Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Combining Anatomical Images with Estimates of Brain Activity
Extracted from Electrographic Data:
Methodology and Applications
Doctorate Thesis
-Patras 2003-
University of Patras School of Medicine National Technical University of Athens Department of Mechanical Engineering National Technical University of Athens Department of Electrical and Computer Engineering
Interdepartmental Program of Postgraduate Studies in Biomedical Engineering
University of Patras School of Medicine Department of Physiology and Department of Medical Physics 26500 Patras Professor George K. Kostopoulos
Alexandra Badea MSc in Medical Physics
Examination Committee:
Professor George K. Kostopoulos
Professor Nicolas Pallikarakis
Professor Ilias Kouvelas
Professor Ioannis Varakis
Associate Professor Panagiotis Dimopoulos
Associate Professor Anastasios Bezerianos
Lecturer Constantinos Papatheodoropoulos
Advisory Committee:
Professor George K. Kostopoulos
Doctor Andreas A. Ioannides
Professor Nicolas Pallikarakis
Εξεταστική Επιτροπή Καθηγητής Γεώργιος Κ. Κωστόπουλος
Καθηγητής Νικόλαος Παλλικαράκης
Καθηγητής Ηλίας Κούβελας
Καθηγητής Ιωάννης Βαράκης
Αναπληρωτής Καθηγητής Παναγιώτης ∆ηµόπουλος
Αναπληρωτλης Καθηγητής Αναστάσιος Μπεζεριάνος
Λέκτορας Κωνσταντίνος Παπαθεωδορόπουλος
Συµβουλευτική Επιτροπή
Καθηγητής Γεώργιος Κ. Κωστόπουλος
∆ρ. Αντρέας Α. Ιωαννίδης
Καθηγητής Νικόλαος Παλλικαράκης
Acknowledgment I am very grateful for the support of the Greek State Scholarships foundation
(IKY). This made possible my graduate studies in Greece and from here all the
experiences described below and many others.
I would like to thank all those who have taught me lessons during these years I have
been a graduate student.
To Professor George Kostopoulos for being an excellent teacher, for teaching among
others what an elegant character and honesty in research are, for being a lot more
to me than a PhD student supervisor.
To Doctor Andreas A. Ioannides for his close supervision, for giving himself example
of what hard work means, for being a never-ending source of ideas.
I am thankful for the chance to work in their laboratories and for their attempt to
create a common workplace, uniting via internet, phone calls and mainly common goals
people on opposite faces of the Earth.
To the members of the Physiology Department for the wonderful atmosphere and
openness to collaborating with people coming with different backgrounds and from
different countries. It has been great to be part of this large family.
To Professor Charles F. Starmer for teaching me about perseverance and marathon
running, about passing through life in a highly interactive, involved way.
To Professors Nicholas Pallikarakis and Basil Proimos for their involvement in the
graduate course in Medical Physics and Biomedical Engineering, an international
experience they host and nurture in Greece for quite a few years.
To my colleague, collaborator and friend Ovidiu Zainea for many discussions, for his
part in our collaborative work, for his care in planning the EEG experiments, and for
his contribution in keeping the programs, the PCs and the local network working.
Thanks to all the people who have used the software described in this thesis and
contributed comments, ideas or criticism related to it (Andreas Ioannides, George
Kostopoulos, Vahe Poghosyan, Marc Schellens, Cristian Badea, Maria Stavrinou,
Milton Ioannides and others).
Thanks to the people who wrote code, made it publicly available, and/or simply
discussed it and taught many other people, including myself, how to write code in
IDL.
Thanks to all the patient subjects who provided data for the studies.
Thanks to Cristian and Andi for teaching me what happiness is!
This thesis is dedicated to my family, without which nothing would have been
possible.
i
Table of Contents
Chapter 1. General Introduction 1
1.1 Motivation and goals 1
1.2 Plan of the thesis 3
Chapter 2. Imaging brain structure and function 5
2.1 Structural brain imaging and the use of MRI 7
2.1.1 History of MRI 7
2.1.2 Physical bases and principles of MRI 7
2.1.3 Clinical and research applications 12
2.2 Functional imaging based on MEG 13
2.2.1 History of MEG 14
2.2.2 Biophysical bases of MEG and EEG generation 14
2.2.3 EEG versus MEG 17
2.2.4 The experimental system 19
2.2.5 Inverse Problem 22
2.2.6 MEG:Clinical and research application 25
2.2.7 Coregistration of MRI and MEG 26
2.3 Conclusion 27
Chapter 3. Brain Segmentation 28
3.1 Introduction 28
3.1.1 Definitions related to image segmentation 28
3.1.2 Applications of Segmentation 29
3.1.3 The problem of brain segmentation 32
3.1.4 Mathematical morphology for image analysis 35
3.1.5 White matter – gray matter separation 38
ii
3.2 A mathematical morphology based method for brain
segmentation 40
3.3 A modified fuzzy c means method for white matter-gray
matter separation 44
3.4 Geometrical, differential properties of the cortex 47
3.5 Results 51
3.1.6 Whole Brain Segmentation 51
3.6.1 Gray matter-white matter separation with bias field compensation 52
3.6 Discussion and Conclusion 53
Chapter 4. Brain Structure Segmentation 56
4.1 Introduction 56
4.1.1 Motivation for segmentation 56
4.1.2 Selected structures of interest 58
4.1.3 Methods for subcortical structure segmentation 63
4.1.4 Background on active contours segmentation 64
4.2 Methods 66
4.2.1 Manual and snake based segmentation 67
4.2.2 Hippocampus Segmentation 69
4.2.3 Amygdala Segmentation 71
4.2.4 Central sulcus segmentation 73
4.2.5 Thalamus segmentation 75
4.2.6 Brain stem segmentation 77
4.3 Results 78
4.4 Discussion and conclusion 80
iii
Chapter 5. Visualization of surface activation 83
5.1 Introduction 83
5.2 Methods 85
5.2.1 Extracting the structural information 86
5.2.2 Computing the activation maps 87
5.2.3 3 D Visualization 88
5.2.4 Slice views 90
5.2.5 The VISIO software features 90
5.3 Results 92
5.3.1 Qualitative Evaluation of Segmentation 93
5.3.2 Surface activation visualization 94
5.4 Discussion 99
Chapter 6. Applications in Neurophysiology 101
6.1. Introduction 101
6.1.1. The somatosensory system 102
6.1.2. Background on the early somatosensory evoked potentials/fields 105
6.2. Methods 107
6.3. Results 108
6.3.1. Electrical stimulation of nerves in the limbs of normal subjects 108
6.3.2. Electrical stimulation of the limbs for a paraplegic subject 115
6.3.3. The primary visual cortex - a combined fMRI and MEG analysis 117
6.3.4. Use of anatomical constraints for EEG dipole localization. Application to
central sulcus 118
iv
6.4. Discussion 121
Chapter 7. A Software Tool for Interactive Determination of
the Plane of Cut through the Rat Brain 123
7.1. Introduction 124
7.2. Methods and materials 124
7.2.1. Reconstructing the rat brain (structures) based on atlas images 124
7.2.2. Search protocol 126
7.3. Results 127
7.4. Practical solution 128
7.5. Discussion 129
7.6. Conclusions 130
Chapter 8. General Discussion 131
8.1 Programs design 132
8.2 Contributions of SAV to understanding brain function,
complementing and integrating the relevant techniques 135
8.2.1 Source localization and extent 135
8.2.2 Source separation 135
8.2.3 Spatial resolution of electrophysiological techniques 136
8.2.4 Comparison of multimodal data 136
8.3 Applications 136
8.4 Outlook 137
Chapter 9. Conclusion 140
Publications i
Abbreviation List iii
References vi
Abstract Advances in hardware and software have made possible the reconstruction of brain
activity from non-invasive MEG and EEG data over a large part of the human brain. The
appreciation of the information content in the data is enhanced when relevant anatomical
detail is available for visualization. Different neuroscientific questions give rise to
different requirements for optimal combination of the information from functional and
anatomical data. Much of the software available today deals with scalar measures of
activity, e.g. changes in hemodynamic demand. The brain activity reconstructed from
MEG and EEG incorporates scalar but also vectorial information, which can be presented
in juxtaposition with relevant anatomical detail from MRI. Furthermore the direction of
the current density vector is expected to be related to the local cortical surface.
To address these problems we introduce an object-oriented software tool dedicated to the
visualization of spatio-temporal brain activity which allows the interplay of geometry and
vector properties of the current density directly in the representations.
The software (SAV) provides modules dedicated to: a) segmentation of the cerebrum
and/or b) subcortical or extra cortical structures and ultimately c) visualization of scalar
and vector fields in the background of the anatomy of the segmented surfaces. The
software succeeds to: a) bring forth the timing of activations and their relationships to the
cortical surface topography; b) allow the user to study the functional data in easy-to-
control view settings and hence; c) navigate through large data sets by focusing on
predefined anatomical structures.
We examine the use of detailed anatomical knowledge in functional studies and derive
quantitative properties of the segmented structures.
Additionally we investigate the applicability of quantitative imaging techniques to
planning an electrophysiological experiment on rat brain slices. We developed software to
visualize selected structures within the rat brain and a procedure to derive an optimal
sectioning plane, which would preserve as much as possible the afferent connections to the
selected structure.
Chapter1. Introduction
1
Chapter1. General Introduction
Chapter1. General Introduction.................................................................................................... 1
1.1 Motivation and goals.................................................................................................... 1
1.2 Plan of the thesis .......................................................................................................... 3
1.1 Motivation and goals
One of the most interesting objects of study for medical researchers and biomedical engineers
alike is the living brain. It is being studied from the cellular level to the system level, from the
experimental to modeling aspects and till the possibility of interfacing with electronic devices.
The goals of these studies are three fold: understanding it, healing or replicating it. Yet much
remains to be done to understand its structure and function.
The general goal of this thesis has been to contribute to solving the problem of linking brain
structure and function. A major landmark was set by Brodmann [1909] who classified brain
regions based on their cytoarchitecture. He looked in fact at the appearance of the cortex under
the light microscope. Since then there has been increasing interest in linking the appearance of
a region and its function.
Sophisticated techniques and devices have been developed for imaging the brain and mapping
its function.
Techniques for imaging the anatomy may use X-rays to produce radiographs on films and also
CT volumetric reconstructions. A newer technique is MRI which adds new information
because of the increased soft tissue contrast.
A number of different techniques have been described for non-invasively measuring human
brain activity. These can be broadly classified into hemodynamic, metabolic or electromagnetic
measurements. Current hemodynamic measurements, particularly functional MRI (fMRI),
provide excellent spatial resolution (millimeters) but are temporally limited by the latency of
the hemodynamic response (seconds) as the physiological time limit and the signal to noise
ratio from the technical point of view. SPECT, PET, EEG, MEG and fMRI are techniques with
different physical bases dedicated to imaging function. The main challenge, besides perfecting
Chapter1. Introduction
2
each of these techniques (increasing the image quality), is to combine in an intelligent manner
the complementary information extracted from each of them.
Currently the most widely used method of analyzing functional brain imaging data consist in
projecting the functional data on anatomical slices [Fischl et al., 1999]. For this purpose the
anatomical and functional data sets must be brought into the same coordinate system.
Comparative studies over a large number of subjects require these data sets to be into a
standardized 3D space. The most common procedure is to report to the Talairach atlas
[Talairach et al., 1967; Talairach and Tournoux, 1988]. While this approach has certain
advantages (ease of use, widespread acceptance, applicability to subcortical structures) it also
has significant drawbacks. The atlas is based on a single subject and contains one hemisphere.
Even recently developed statistical atlases, based on large number of brains only give
probabilistic locations since there is high variability between normal brains. These drawbacks
are partly due to the presence of the extensive convolutions of the cerebral cortex, almost as
particular to an individual as the fingerprints. A central problem is the analysis of regions
buried within the deep and irregular sulci of the cortex. Estimates of the amount of buried
cortex range from 60 to 70% [Van Essen and Drury, 1997; Zilles et al., 1988] implying that
distances measured in 3-D space between two points on the cortical surface will easily
underestimate the geodesic distance (measured on the cortical surface), particularly in cases
where the points lie on different banks of a sulcus.
The location and extend of particular functional entity may differ from one normal individual
to another and even more in pathological cases. The difficulty is not simply in visualization,
because buried regions can be exposed either by cutting slices through the cortex using
noninvasive imaging techniques (MRI, CT, etc) or by using computational techniques to flatten
the cortex with minimal distortions. Rather, the main challenge is to decipher the complex
spatial relationships between regions contained on different slices or lying in different sulci.
Our goal has been to develop methods and implement them into software dedicated to
combining anatomical images with estimates of the human brain activity from electrographic
data, in such ways as to reveal novel information.
The largest part of the thesis is dedicated to extracting detailed anatomical information on the
brain from high resolution MR scans of the head and using this information to enhance the
appreciation of information on magnetic sources extracted from MEG.
Similar techniques have been used for reconstructing the 3D anatomy of the rat brain, this time
based on atlas data [Paxinos, 1997] and studying the geometry of connections between separate
Chapter1. Introduction
3
structures. Specifically we examined the possibility of inferring the position and orientation of
a cut plane which would preserve intact as many as possible connections of the fornix to the
mamillary bodies. This would benefit electrophysiological experiments in rat brain slices for
the study of mammilarry bodies and the same methodology could eventually be applied to
other structures.
1.2 Plan of the thesis
The thesis is organized in nine chapters, the first two being introductory. The next five are
intended to be self contained and in general include their own introductory sections, methods,
results and conclusion. The last two chapters are a general discussion and conclusion to the
thesis.
The first chapter is the present one and aims to give a quick introduction to the problem of
combining anatomical images with estimates of brain activity from electrographic data, in the
general context of brain imaging. The second chapter gives a broader introduction to the
physical bases and techniques used for anatomical imaging with MRI and functional imaging
with MEG and EEG, with an accent on MEG. Also some details on the instrumentation used
for recording the MEG data used in this thesis is given in Chapter 2.
Chapters 3 and 4 introduce the problem of segmenting the cerebrum and its main
compartments: the gray and white matter in chapter three, and selected gray matter structures
in chapter four. The methods used for the two chapters are different; however they can be used
in combination for best results. Some of the segmented brain structures may be relevant for
activation studies, and we present examples in the next chapters, and/or may have significance
since their changed morphology may be an indicator in certain disorders (like Alzeimer’s
disease, schizophrenia, etc). Additionally the computation of local, differential properties of
these structures is described.
Chapter 5 introduces the various visualization options we have developed and used for
representing the segmented anatomy and overlaying it with functional data in the form of
equivalent current dipoles and current density or statistical parametric maps. We introduce into
the representation of activation data not only the scalar but also the vector aspect where
applicable.
Chapter 6 describes applications of the above methods in neurophysiology, specifically the
study of somatosensory evoked fields (the cases of median and tibial nerve stimulation), and
visually evoked fields. Another application makes use of the information on the local geometry
of the anatomy. This information is incorporated into an exhaustive search for the EEG dipoles
Chapter1. Introduction
4
which best models the activity evoked by median nerve stimulation at the wrist. The activity
sources are dipoles located in the contralateral hand area of the primary somatosensory cortex.
Chapter 7 describes one of the possible uses of image processing, visualization and
quantification for planning an electrophysiological experiment on rat brain slices, in particular
when one wants to preserve the maximum possible number of connections between two
structures.
Chapter 8 and 9 consist in the discussion and general conclusion of the thesis.
At the end are the references and the lists of publications and abbreviations.
Chapter2.Imaging brain structure and function
5
Chapter 2. Imaging brain structure and function
Chapter 2. Imaging brain structure and function.......................................................................... 5
2.1 Structural brain imaging and the use of MRI............................................................... 7
2.1.1 History of MRI......................................................................................................... 7
2.1.2 Physical bases and principles of MRI ...................................................................... 7
2.1.3 Clinical and research applications.......................................................................... 12
2.2 Functional imaging based on MEG............................................................................ 13
2.2.1 History of MEG...................................................................................................... 14
2.2.2 Biophysical bases of MEG and EEG generation ................................................... 14
2.2.3 EEG versus MEG................................................................................................... 17
2.2.4 The experimental system........................................................................................ 19
2.2.5 Inverse Problem...................................................................................................... 22
2.2.6 MEG: Clinical and research application ................................................................ 25
2.2.7 Coregistration of MRI and MEG ........................................................................... 25
2.3 Conclusion.................................................................................................................. 26
The work of Broca [1865], who demonstrated that speech deficits result from damage to the
left frontal lobe, can be considered as the start of the modern research on correlating brain
structure and function. Intraoperative surgical mapping (through electrical stimulation)
pioneered by Penfield and others [Penfield and Roberts, 1959] allows the correlation of
structure and function in more detailed manner compared to pathological observation. This
technique is however invasive and therefore limited to pathological cases, which can present
altered relationships between structure and function [Woods, 1996]. Only with noninvasive
brain imaging can the normal population be studied. The same applies to studying function in
patients before and after localizing the area of pathology [Damasio and Frank, 1992]. This
approach allows for prediction of structure based on function and prediction of function based
on structure, a situation suited for framing and testing new hypotheses [Woods, 1996].
Functional brain imaging is a multidisciplinary research field that encompasses techniques
devoted to a better understanding of the human brain through noninvasive imaging of the
electrophysiological, hemodynamic, metabolic, and neurochemical processes that underlie
normal and pathological brain function.
Chapter2.Imaging brain structure and function
6
The brain is a complex dynamical system, with many degrees of freedom, and multichannel
measurements are necessary to gain a detailed understanding of its behavior [Mira and Pesaran,
1999]. Such measurements include multielectrode recordings, optical brain imaging, functional
magnetic resonance imaging (fMRI) [Ogawa et al., 1992; Kwong et al., 1992], and
magnetoencephalography (MEG) [Hämäläinen et al., 1993].
The anatomy can be visualized with techniques like MRI and CT, while the functional aspects
can be studied noninvasively using PET, SPECT, EEG, MEG or fMRI.
The coupling between neuronal firing and blood flow is exploited by imaging techniques like
PET, SPECT and fMRI. fMRI exploits the fact that an increase in neuronal activity results in a
local increase in blood flow, exceeding metabolic oxygen demands. There is more
oxyhemoglobin than deoxyhemoglobin locally, oxyhemoglobin is paramagnetic, hence the
greater activity on the image.
PET and SPECT involve detecting gamma ray photons and reconstructing 3D maps of
radioactive tracer concentration. More active brain regions have a higher rate of blood flow and
receive the tracer earlier than other areas or emit stronger radiation. In contrast with these
methods, which give indirect measure of neuronal activity, EEG and MEG give a direct
measure of the electrical activity of neural cells.
MEG and EEG are techniques that measure, respectively, the magnetic induction outside the
head and the scalp electric potentials produced by electrical activity in neural cell assemblies.
They are the only noninvasive techniques that can analyze the whole brain with a sub
millisecond time resolution. Their space resolution (about 5mm for cortical sources) is too
coarse to be sensitive to the activity of one or even few neurons, but it reflects a cooperative
effect of a large number of neurons spread over at least a few millimeters of cortex.
The spatial resolving power of MEG and EEG is limited by the number of spatial
measurements (a few hundreds recording places) and the inherent ambiguity of the underlying
quasistatic electromagnetic inverse problem. Introducing constraints in the models on the
source generation of MEG and EEG signals can increase resolution substantially, but it can
also bias the reconstructions wrongly if these constraints reflect wrong assumptions.
The experiments designed to study brain function usually involve repeated measures of the
method’s specific parameter, a paradigm possible for noninvasive methods only. The simplest
such functional experiment is the activation study, aiming to identify, locate and eventually
quantify the brain region “responsible” for a certain function. Usually multiple scans are
acquired for two conditions, typically one “at rest” and a second “at activation”. These studies
generate huge amounts of data which are usually studied after statistical processing to
Chapter2.Imaging brain structure and function
7
emphasize significance but also as single trials, in an attempt to preserve much of the original
information. It is the analysis and mainly the visualization of these large multichannel,
multidimensional data which is of interest to the present study.
The next subsections provide an introduction to the imaging methods which provide the basis
for the activation studies. Section 2.1 gives a background on MRI, the technique providing
anatomical data. Section 2.2 gives a background on MEG (and EEG). These are the techniques
providing the functional data used in our study. At the end of this section is a discussion on
combining information from anatomical (MRI) and functional (in particular case MEG)
techniques.
2.1 Structural brain imaging and the use of MRI
2.1.1 History of MRI
Wolfgang Pauli suggested in 1924 that the interior of the atom's nucleus could be manipulated
and caused to move with angular momentum (spin) and become magnetic. Isidor Isaac Rabi
won the Nobel Prize in physics in 1944 for his resonance method for recording the magnetic
properties of atomic nuclei. Bloch and Purcell discovered independently the nuclear magnetic
resonance or NMR and shared the Nobel prize in 1952. In 1971 Raymond Damadian
demonstrated that NMR could be used for a medical purpose showing that the nuclear
magnetic relaxation times of tissues and tumors differ. Damadian drove the development of a
full body NMR machine, later called MRI, which would become available only in 1977. This
NMR device was designed to detect cancer but did not produce high resolution images. In
1973 Paul Lauterbur [Lauterbur, 1973] described rotating the magnetic gradient around an
object to create an image based on backprojection algorithms, similar to those developed for
CT.
In 1975 Richard Ernst proposed using phase and frequency encoding, and the Fourier
transform. This technique is the basis of current MRI techniques, approved for clinical use by
FDA in 1985 [Hornak, 2002].
2.1.2 Physical bases and principles of MRI
Magnetic resonance imaging gives information on the proton density and the media in which
they are found, based on the absorption and emission of energy in the radio frequency range of
the electromagnetic spectrum.
At the basis of nuclear magnetism lie the properties of nuclei that contain an odd number of
protons, neutrons or both in combination, such as 1H, 13C, and 23Na. These properties are spin
angular momentum and the associated magnetic moment along the spin axis.
Chapter2.Imaging brain structure and function
8
The classical approach
The proton can be considered to posses electrical charge e and an angular momentum I. The
electrical charge moving around the spin axis gives rise to a magnetic moment (µp).
In the presence of an external static magnetic field, the magnetic dipole moment is subject to a
torque, µp x B0. This results in precession of the magnetic moment around the magnetic flux
density, B0.
The frequency with which the moment precesses is given by the Larmor equation:
ω0=-γ B0
Where ω0 corresponds to the frequency of precession called the Larmor frequency and γ
(MHz/T) is the gyromagnetic ratio
The net magnetic moment (M) of a sample made of many nuclei is the vector sum of all
nuclear magnetic moments and at equilibrium it is aligned to B0 (which defines the z axis), see
Figure 2.1-1.
A second, radiofrequency field (B1), perpendicular to B0, i.e. in the xy plane is used to tilt M
away from B0 to produce a measurable component in xy plane. In a rotating frame reference
system, moving together with B1 the effective field is:
Beff=(ω0- ω1)/γ=B0- ω1/γ
For a B1 characterized by the Larmor frequency the effective field is zero and the
magnetization will precess around B1. Once displaced from the z axis by B1, the motion of the
magnetization vector will induce a current in the radio frequency (RF) coils. A typical such
signal, called free induction decay (FID) is shown in Figure 2.1-1 c.
Figure 2.1-1 (a) A rotating magnetic flux density B1 is applied perpendicularly to the static field of flux
density B0, M will experience an additional torque which moves it at an angle α from the z axis (x’ and y’
are rotating with the Larmor frequency). (b)The motion of M in the laboratory frame following the
application of a 90o pulse. This motion can be considered as composed of the motion of the rotating frame
relative to the laboratory frame plus the motion within the rotating frame. (c) The resulting FID that can be
measured from the xy component of the magnetization decays because of spin-lattice and spin-spin
interactions (from [Webb, 1998]).
Chapter2.Imaging brain structure and function
9
The quantum approach
A thermal equilibrium the distribution of spins in the two energy states (parallel or antiparallel
to the applied field B0) follows Boltzmann’s law.
∆=
−
+
skTE
downspinn
upspinnexp
)21,_(
)21,_(
k is Boltzmann constant, Ts is the absolute temperature of the spin system, and n is the number
of spins in a given state and E∆ is the energy difference between the two states.
The energy of these states is: E=-µB0=- γm1ћB0.
Where µ is the magnetic moment, ћ is the reduced Planck’s constant (1.0546x10-34 Js) and m1 is
the magnetic cuantum number. For the protons the magnetic quantum number is m1=+/-21 and
therefore: E∆ = γћB0
The lower energy state (spin up or parallel to the external field) has a larger population of spins
compared to the high energy state (spin down or antiparallel to the external field). As a result,
the tissue will exhibit a net magnetization, dependent on the strength of the magnetic field and
the thermal agitation.
Transitions can take place from the lower state to the higher state if energy at the Larmor
frequency is absorbed:
Ћω0= γћB0
Figure 2.1-2. The stationary states of a proton spin in a constant magnetic field H0. The application of a
radio frequency (RF) magnetic field at the Larmor frequency results in energy absorption. After cessation
of the RF field the excited protons go back to the lower energy state and a FID is emitted [from Cho at al.,
Chapter2.Imaging brain structure and function
10
1993].
In the external static magnetic field, B0, nuclei can be shifted from the parallel to antiparallel
alignment by the application of radio frequency energy, E = ћγB. If we consider an RF
magnetic field, B1, applied perpendicular to B0, the system will absorb energy. Upon
termination of the RF pulse, the nuclei return to their original alignment parallel to the applied
static field and energy is emitted in the form of a weak RF signal of specific frequency. This
frequency of the emitted signal depends on the strength of the applied static magnetic field as
well as the type of nuclei producing the signal.
T1 and T2 parameters
The emitted RF signal is called the free induction decay (FID) signal and is picked up by a
receiver coil. It constitutes a central part of MR imaging. The waveform of this signal is an
exponentially damped sine wave. The decay is due to relaxation phenomena spin-lattice (or the
longitudinal magnetization realaxation, Mz) and spin–spin (for the transverse magnetization
realaxation, Mxy). These processes are characterized by the T1 and T2 constants. T1-weighted
images reflect the recovery of the original signal as the flipped atoms return to baseline from
the excited state. They reflect the signal intensity along the z-axis. In addition to the rotation,
the net magnetization starts to dephase because each of the spin packets making it up is
experiencing a slightly different magnetic field and rotates at its own Larmor frequency. The
time constant which describes the return to equilibrium of the transverse magnetization, MXY,
is called the spin-spin relaxation time, T2.
The decay of transverse magnetization is due to: 1) molecular interactions; 2) variations in the
main magnetic field (never completely uniform; 3) added gradients. The combination of these
factors is what actually results in the decay of transverse magnetization. The combined time
constant measures the net remaining signal intensity in the transverse plane.
The Bloch’s equations describe the behavior of the nuclear magnetic moment or a sample of
minimally interacting spins.
BMdt
dM×= 0γ
NMR experiments can be performed in “continuous wave” mode or in “pulsed” mode. In the
continuous mode the radiofrequency is slowly scanned and any induced current is detected. As
the frequency approaches values close to the Larmor frequency of the protons, the effective
field in the rotating reference frame becomes smaller till B1 is larger than Beff and the
magnetization will precess around B1. This will induce a current that is detected. In pulsed
mode a short RF pulse is applied, during this time the magnetization will precess around B1.
Chapter2.Imaging brain structure and function
11
When the pulse is turned off the magnetization will precess around B0 inducing a measurable
current, which decays in time. The pulse duration determines how far the magnetization
rotates, or the flip angle.
The equation of Bloch for the net magnetization M ( kMjMiMM zyx ++= ), taking into
account the simultaneously occurring relaxation mechanisms is given by:
1
0
2
)())()(()()()(
TMtMk
TjtMitM
tBxtMdt
tdM zyx −−
+−= γ
and the solutions follow the curves shown in Figure 2.1-3, namely:
2/0
1/0 )1(
Ttxy
Ttz
eMM
eMM−
−
=
−=
Figure 2.1-3: a) T1-weighted images reflect the recovery of the original signal as the flipped atoms return to
baseline from the excited state. B) Decreases in T2* signal are due to dephasing of the flipped atoms [from
Hornak, 2002].
In order to produce an image, each MR signal must be referenced to a specific region of tissue.
This is accomplished by applying a gradient magnetic field and a narrow band RF field in
which the field strength varies linearly with position. Besides this slice selection gradient, other
two gradients are used for phase and frequency encoding. Using a computer-aided
reconstruction program, similar to that used in computed tomography, the signals attributed to
individual volume elements of tissue can be resolved and reconstructed into an image.
Traditional magnetic resonance imaging (MRI) systems measure spatial distribution of several
tissue-related parameters, as longitudinal and transversal relaxation times and proton density.
From a single anatomical slice can be collected a multitude of numerical data. The data
consists of T1, T2 relaxation constants values and/or proton density pixel intensities at each
spatial location in the MRI. These parameters can be aggregated into multidimensional data
which can used to cluster the tissue in the feature space.
Various pulse sequences are available: spin echo of T1 and T2, inversion recovery, cardiac or
respiratory gated imaging, and from the high speed imaging techniques: single shot imaging
Chapter2.Imaging brain structure and function
12
(echo planar or spiral planar), multiple shot imaging.
The choice of the sequence parameters will have an influence on the amount of signal coming
from the object and hence will affect the SNR. These parameters include the time between the
90 degrees pulse and the echo, called the echo time TE, the inversion time TI, the time between
two successive 90 degrees pulses, called repetition time TR, the flip angle , the shape of the
radio frequent pulses, etc.
An example of the different appearance of the same tissue type in T1, T2 and PD images is
shown in Figure 2.1-4.
Figure 2.1-4: From left to right: T1 weighted, T2 weighted, PD images of the brain [from “The Whole Brain
Atlas” available online at http://www.med.harvard.edu/AANLIB/, authors Johnson KA and Becker JA].
The strength of the MRI signal, i.e., the signal emitted when nuclei return to their equilibrium
state, depends primarily on three parameters: the proton density (the higher the density of
protons, the larger the signal), T1, and T2. The contrast between brain tissues depends upon
how these parameters differ in the tissues.
In T1 weighted images the signal from fat is bright, whereas image intensities from areas of
muscle and fluid are lower. Cartilage, ligaments and tendons appear very dark. Bone marrow is
also bright due to its fat content.
In T2 weighted image fluid and edema appear bright. The fat displays moderate intensity
whereas the muscles are dark. Tendons, ligaments and cartilage still appear dark.
For most "soft" tissues in the body, the proton density is very homogeneous and therefore does
not contribute much to signal differences in the image.
2.1.3 Clinical and research applications
Two methods of imaging the 3D anatomy are CT and MRI and their relative advantages can
make useful a complementary approach using both techniques. In comparison with CT,
dedicated more to examinations of the chest, abdomen, pelvis, and the skull base, MRI is the
procedure of choice for neurologic and orthopedic diagnoses, but can be used also for liver,
Chapter2.Imaging brain structure and function
13
kidney, pancreas, etc. and eventually in association with Gadolinium for venography or
artherography. MRI has a higher capacity for displaying soft tissue contrast.
A notable example is the discrimination between the gray and white matter of the brain. In
addition, MRI is unobstructed by bone, and thus it is especially beneficial in high resolution
imaging of the brain and spinal cord. Image contrast can be tailored to the specific clinical
application so that specific types of pathology are emphasized. Three-dimensional
reconstructions of anatomic structure can be obtained. These characteristics render MRI a very
effective and important tool for soft tissue imaging.
Obtaining a structural brain scan in a psychiatric patient can be useful to exclude trauma, or
other brain disorders (stroke, multiple sclerosis) that might have the psychiatric symptom as
their presenting problem.
While CT is the imaging method of choice for examining skull fractures, the presence of lots of
bone at the base of the skull make CT scans difficult to read for the cerebellum and base of the
skull. MRI on the other hand present advantaged related to soft tissue contrast. MRI scans can
be used to identify an old infarct or tumor. Additionally, MRI can identify problems with white
matter. For example, multiple sclerosis (MS) can often present with psychiatric symptoms,
especially depression. On MRI, the MS plaques can be seen as very bright areas in white
matter where demyelination has occurred.
Areas of increased signal intensity often occur as a normal consequence of aging
(approximately one per decade of life). MRI scans can give exquisite information about brain
size and shape, and this is may be helpful clinically in conditions like Alzheimer’s to assess
atrophy. Suddath and coworkers [Suddath et al., 1990] searched for anatomical correlates of
schizophrenia by performing MRI scans in identical twins, one of whom had schizophrenia and
the other did not. Compared to the twin, the schizophrenia patients clearly had increased
ventricular size. However this finding of is not helpful clinically as it is neither specific nor
sensitive for schizophrenia. MRI scans sometimes find evidence of other brain abnormalities in
neuropsychiatric disorders. For example, midline developmental defects in schizophrenic
patients [George et al., 1989; Scott et al., 1993] or the spreading of brain tumors could be
documented.
2.2 Functional imaging based on MEG
A good description of the MEG principles, instrumentation and applications can be found
elsewhere [Hämäläinen et al., 1993]. Here will be given just a short introduction and the
description of the actual system (section 2.2.3) used for the recording of MEG signals used
throughout this study.
Chapter2.Imaging brain structure and function
14
2.2.1 History of MEG
While the first human EEG was recorded in 1924 and reported by Hans Berger in 1929
[Berger, 1967], the first measurements of magnetic fields produced by the brain’s electrical
activity were recorded in by David Cohen in 1968. He detected the magnetic alpha rhythm
using an induction coil magnetometer in a shielded room [Cohen, 1972]. This was possible
because of the observation of Zimmermann and colleagues that quantum interference could be
observed in a superconducting loop with a single junction if it was excited by a radio-
frequency bias. The SQUID (superconducting quantum interference device) was developed in
between 1964-1969 [Zimmerman et al., 1970] and used for the first time in biomagnetism in a
joint experiment done at MIT together with David Cohen and Edgar Edelsak.
Electromagnetic measurements have progressed from acquisition of a single or a few channels
to hundreds of signals acquired simultaneously. The measurements of magnetic and electrical
activity result in the magnetoencephalogram and encephalogram and are able to follow changes
in neuronal activity on a millisecond time scale, comparable to the dynamics of the neuronal
assemblies. MEG and EEG are produced directly by the electrical activity through which the
brain communicates and therefore suitable for measuring some of the electrical brain activity
with high (ms) time resolution.
2.2.2 Biophysical bases of MEG and EEG generation
The principal building blocks of the brain are neurons and glial cells. The two principal groups
of cortical neurons are the pyramidal and the stellate cells. The former are large cells, their
apical dendrites from above reaching parallel to each other. These cells are mostly present in
the 3’rd layer of the visual cortex or the 5-th layer in the motor cortex, they rely mainly local
information within different cortical layers. They yield information about the “vertical” activity
below the cortical surface, corresponding to cortical columns. Since they tend to be
perpendicular to the cortex the resulting direction of the electric current flowing in the
dendrites is also perpendicular to the cortex. The corresponding electric field is associated with
a magnetic field, perpendicular to it.
Not all cortical cells generate fields measurable through EEG/MEG. For example the stellate
cells generate “closed fields”, because their dendrites which receive the input have
ramifications in all directions. They constitute the so called “silent sources”. The pyramidal
cells, perpendicular to the cortex and clearly asymmetric along a direction perpendicular to the
cortex, are the main generators of the measured electromagnetic field. The essential property of
these neurons responsible for the E/MEG signal generation is that they are regularly arranged
and activated in a (more or less) synchronous way.
Chapter2.Imaging brain structure and function
15
The electric and magnetic fields generated in the brain are caused by two distinct processes: a)
postsynaptic and b) action potential currents.
Post synaptic potentials
At equilibrium the concentration of ions inside and outside the membrane is maintained at
specific levels, the Na concentration being higher outside than inside the membrane, whereas
the K concentration is smaller outside than inside. The Na/K pump is involved in maintaining
these levels or returning to equilibrium, it moves three Na ions out and two K ions into the cell
in one duty cycle [Hämäläinen, 1993].
Figure 2.2-1 Networks of cortical neural cell assemblies are the main generators of MEG/EEG signals. Left:
Excitatory postsynaptic potentials (EPSP) are generated at the apical dendritic tree of a pyramidal cell and
trigger the generation of a current that flows from the non-excited membrane of the soma and basal
dendrites to the apical dendritic tree sustaining the EPSPs. Some current takes the shortest route between
the source and the sink by traveling within the dendritic trunk (primary current in blue), while
conservation of electric charges imposes that the current loop be closed with extracellular currents flowing
even through distant parts of the volume conductor (secondary currents in red). Center: Large cortical
pyramidal nerve cells are organized in macro-assemblies, their dendrites normally oriented to the local
cortical surface. Right: Functional networks made of these cortical cell assemblies and distributed at
possibly mutliple brain locations are the putative main generators of MEG and EEG signals (from [Baillet
et al, 2001]).
When a neurotransmitter molecule binds to a postsynaptic receptor, the membrane permeability
is altered for specific ions, and there is a flux of ions through the membrane which gives rise to
a post-synaptic potential of about 10mV, lasting about 10ms. The moving ions give rise to a
current within the cell. This mechanism can be represented as a dipolar field, decreasing as the
inverse of the squared distance. Currents in the surrounding tissues, called volume currents,
close the circuit and so that there is no accumulation of charge in the medium [Plonsey, 1981;
Chapter2.Imaging brain structure and function
16
Karp, 1981] (Figure 2.2-1).
Both primary and secondary currents contribute to magnetic fields outside the head and to
electric scalp potentials, but spatially structured arrangements of cells are important for the
superposition of neural currents such that they produce measurable fields.
Macrocolumns of tens of thousands of synchronously activated pyramidal cells are believed to
be the main MEG and EEG generators because of the coherent distribution of their large
dendritic trunks locally oriented in parallel, and pointing perpendicularly to the cortical surface
[Nunez et al, 2000]. The activity from 5x5mm patch of cortex, assumed to be 4mm thick, and
made of macrocolumns characterized by a current density 100 nA/mm2 is estimated to be
about 10 nA-m, consistent with empirical evidence [Baillet et al, 2001]. A single synapse
contributions is estimated to be of about 20 fA-m [Baillet et al, 2001], very small to be
detected.
The currents associated with the postsynaptic potentials synchronously generated among the
dendrites of the pyramidal cells are believed to be the main source of the signals detected in
MEG and EEG. One reason is that they typically last longer than the rapidly firing action
potentials [Nunez, 1981]. In addition to postsynaptic potentials, other slow variations of the
membrane potential, such as those associated with depolarizing or hyperpolarizing after
potentials and dendritic events as calcium action potentials may be sources of extracellularly
measured potentials.
Action potentials
The AP propagation along the cell’s axon is modeled by a quadripolar structure generating a
field which decreases with the inverse of the cubed distance [Figure 2.2-2 ].
Figure 2.2-2 Extra cellular sources and sinks in the axons [from Ayache et al., 1999]
Its effect is negligible compared to the post synaptic potential, when measured outside the
cortex, despite the fact that the amplitude of the signal is higher.
The time constant is 1ms compared to 10ms for the postsynaptic potentials, accounting thus for
Chapter2.Imaging brain structure and function
17
the high frequency components which tend to cancel when averaged over time. Because of
their short duration they tend to overlap much less than the postsynaptic potentials.
This contribution is therefore difficult to measure by current technology in EEG/MEG.
It is believed that because the field generated by the dipoles decreases with the inverse of the
square distance, only the currents produced in the cortex generate signals strong enough to be
recorded. However some authors reported scalp recordings having as source the hippocampus
[Tesche and Karhu, 2000] cerebellum [Tesche and Karhu, 1997], and thalamus [Tenke et al,
1993], [Llinas et al, 1999].
2.2.3 EEG versus MEG
Although EEG and MEG measure the same electrical phenomena they do differ in a number of
aspects which will be discussed below.
The sources of interest are classified as primary and secondary sources [Figure 2.2-3]. The
primary sources reflect the impressed current density, due to transformation of energy from
chemical to electrical and microscopic passive cellular currents. They represent areas of
activity associated with a given sensory, motor, or cognitive task. The secondary sources are
the volume currents that result from the macroscopic electric field. Both types of sources
produce an electric and magnetic field.
Figure 2.2-3 MEG and EEG measurements differ in that MEG is more sensitive to the intracellular current
or the primary sources (left) and EEG to the extracellular current or secondary sources (right)
EEG measures a sum of potentials caused by primary and secondary sources. It is more
sensitive to the effects of (secondary) currents, since only these can volume conduct to the
scalp to produce the EEG. Cohen and Cuffin [Cohen and Cuffin, 1983] stated that MEG
measures mainly the intercellular current while not being affected by these volume currents,
since their effects cancel due to symmetry. This may explain a more focal pattern in MEG
relative to the EEG potential distribution.
The average behavior of a population of pyramidal cells can be modeled by a dipole with a
magnetic field normal to the cortex and an electric field tangential to it.
MEG's primary response is to tangential generators, whereas EEG is sensitive to both
tangential and radial generators. Given the arrangement of pyramidal cells as dipole layers,
Chapter2.Imaging brain structure and function
18
perpendicular to the cortical surface, in general the EEG measurement will give information on
the activity of cells located in gyri (parallel to the skull) but also in sulci, while MEG will give
information on the activity of cells located in sulci (normal to the skull) [Figure 2.2-1b]. MEG
gives clear information about tangential dipoles, which might be obscured by radial sources in
EEG.
The map representing the distribution of the magnetic field at the surface of the head caused by
a tangential current dipole is rotated at 90o relative to the corresponding EEG map. Differences
in the localization of dipoles with the two methods may be due to the directions along which
measurements are made.
The magnetic field given by a distribution of impressed current source elements ji located
within a finite, inhomogeneneous conductor (divided into homogeneous regions vj with
conductivity σj,) with spherical symmetry is given by:
Equation 2.2-1
Where r is the distance from an external field point at which is evaluated to an element of
volume dv inside the body, dv is a source element, and is an operator with respect to the
source coordinates.
The first term on the right-hand side of Eq. 2.2.1 represents the contribution of the volume
source, and the second term the contribution of the boundaries between regions of different
conductivity.
The influence of inhomogeneities is larger on EEG than on MEG. The field potentials are
influenced not only by the geometry of the neuronal populations and the electrical properties of
individual neurons but also by the presence of inhomogeneities, regions with different
conductivities. The layers or compartments lying around the brain (i.e. CSF, skull and scalp)
attenuate the original signal. The scalp EEG is most influenced by the conductivities of the
skull and scalp while MEG by the conductivities of the brain and CSF [Haueissen, 1996]. The
effect of anisotropic conductivity of the skull is the smearing out of the distribution of EEG of
the scalp, whereas the normal component of MEG is not affected [da Silva and van Rotterdam,
1999].
To explain the EEG the conductor model assumes the existence of layer with different
conductivities. These cause and attenuation of the signal and introduce equivalent dipole layers
at the boundaries of the compartments. However the brain and surrounding tissue behave like a
medium with constant magnetic permeability. Therefore the magnetic field is not influenced by
these layers, nevertheless it is affected by the induced dipole layers existing at the boundaries
Chapter2.Imaging brain structure and function
19
of the compartments with different conductivity.
While EEG is a relative measurement, no electrodes or reference point is needed for MEG.
Also the transducers need not touch the scalp, because the magnetic field does not disappear
where the conductivity is zero.
There are differences in the MEG and EEG measurements regarding the representation of some
non dipolar sources (sources arranged in the form of lines, longer than 2cm give different
results for example) and these differences may be helpful in identifying the type of such
sources in the brain [da Silva and Van Rotterdam, 1999].
In the presence of noise it appears that MEG and EEG have different capabilities of retrieving
the position and orientation parameters: EEG gives better estimates of the orientation, while
MEG on the position. The inverse procedure fails more often in cases of bad SNR for EEG.
The use of MEG and EEG may complement each other and lead to a better model and a more
precise localization of the source.
2.2.4 The experimental system
The MEG system measures the very weak magnetic fields produced by the cooperative activity
of neuronal assemblies. Neuromagnetic fields are typically 50-500fT in strength, while the
earth's magnetic field is about 5x1-5T and the electromagnets used to pick up cars generate a
field comparable to that of MRI clinical machines (1.5-2.0T). The tiny fields generated by
neuronal assemblies require a very sensitive detector and eventually shielding from the
surrounding magnetic fields.
The only device with sufficient sensitivity for high quality biomagnetic measurements of
magnetic flux generated by the brain’s activity is the SQUID magnetometer, which operates at
low temperatures produced with liquid helium.
The SQUID is based on the Josephson effect. It consists of superconducting ring interrupted by
one or two Josephson junctions. The recording dewar contains magnetic detection coils,
continuously bathed in liquid helium to superconducting temperatures of 4K (-269oC). The flux
sensitivity of a SQUID is about 10-15 Tm2. To improve the sensitivity a flux transformer is
used. A flux transformer is a simple superconducting coil which collects ambient flux. This
initial 'pickup' (input) coil is connected in series to a second input coil which is tightly
inductively coupled to the SQUID itself. This simple circuit has the effect of funneling the flux
trapped in the pickup coil into the SQUID loop, thus providing enhanced flux sensitivity
(Figure 2.2-4 a).
Chapter2.Imaging brain structure and function
20
The gradiometer consists of an inductively coupled flux transformer with two coils connected
in series and oppositely wound. Because of the opposite winding a uniform magnetic field will
not induce flux into the SQUID. If a magnetic field with a spatial gradient is applied in the z
direction there is a difference in fluxes in each coil and therefore a net current in the
gradiometer - a flux is then generated in the SQUID loop (Figure 2.2-4 b).
Figure 2.2-4. Flux transformers used to transport flux to the SQUID loop. (a) shows a simple one coil flux
transformer, (b) a first order gradiometer [from Singh, 1995].
The detection of such small magnetic fields in the presence of high background magnetic noise
requires shielding and compensation of the existing fields and therefore the recordings are
usually done within magnetically shielded rooms.
The main problems with shielded rooms are that they are very expensive and often do not
provide adequate signal to noise ratios. They fail to provide defense against noise sources from
within the patient such as the heart. This problem was overcome by considering distant noise to
be spatially uniform at the detector and the use of gradiometers which makes the system
sensitive only to spatial gradients of the field, rejecting sources of noise. The magnetic fields
generated by the brain decay rapidly with the distance.
Because the lower coil is much closer to the brain's activity than the upper coil, the output is
roughly equal to magnetic field. The gradiometer is less sensitive to distant sources of
interference and slight vibrations of the instrument than a magnetometer. Forming a virtual
second or third order gradiometer by using a reference system further extends this principle.
The formation of third order gradiometer technique actually allows the system to be operated
without any magnetic shielding, or in cooperation with a shield to achieve even better
performance.
a b
Chapter2.Imaging brain structure and function
21
Magnetometer based systems optimize signal strength but gradiometer based systems optimize
signal to noise ratio, which is the more important value when examining MEG data.
Whole head MEG systems are available with a different number of channels, including may
have between 64, 149, 151 and 275 MEG channels to simultaneously measure activity across
the entire cortex.
The MEG of the BSI in RIKEN is a whole-head Omega biomagnetometer (CTF Systems, Inc.,
Vancouver, BC, Canada). The biomagnetometer consistes of Niobium detection coils emerged
in liquid helium contained within a Dewar. The helmet-shaped, liquid helium-filled dewar is
placed inside a shielded room (NKK, Japan). The helmet houses 151 primary channels
uniformly distributed in the inner lower surface of the dewar and 28 reference sensors. Each
primary channel is a first-order axial gradiometer with the two 1-cm-radius coils separated by 5
cm. The proximal sensor coils are on average a little over 3 cm from the (outside) wall of the
helmet. The distance between a proximal sensor coil and the brain surface varies from 5 to 8
cm depending on the shape and actual position of the subject’s head inside the dewar. The
reference channels are a collection of magnetometers and gradiometers placed about 30cm
above the uppermost gradiometer inside the dewar away from the subject’s head. The signal
from a subset of the reference channels is used to construct in the software a synthetic higher
order signal which behaves like a gradiometer with a long baseline and is therefore effective in
eliminating noise from distant sources. In addition to the MEG channels the EOG and ECG can
be acquired at the same time for offline noise elimination
Figure 2.2-5. The MEG system at RIKEN BSI is an OMEGA 151 (151 channels) (from
www.vsmmetdtech.com).
The equipment used to measures the very weak magnetic fields produced by electric activity in
the brain is a Whole-Cortex MEG System (Omega151) produced by CTF System Co. The
Chapter2.Imaging brain structure and function
22
recording of the magnetic field around the head is characterized by good spatial sampling
(average distance between coil sensors: 3.1cm). The signal from all 151 sensors is sampled
simultaneously with (routine) sampling rate 1250 Hz.
The sensors include: axial gradiometer with 5cm baseline and 2cm coil diameter. Additionally
29 reference sensors are placed on the far side of the MEG array to allow software third
gradient mode for effective elimination of environmental noise. It is possible to record
simultaneously with the MEG electrical channels to monitor heart and eye blinks and 64
channels of EEG (electroencephalogram) Measurements are possible with the subject in sitting
or supine positions Figure 2.2-5.
The EEG system consists in two Synapms Amplifiers (Neuroscan, LTD) each having 32
channels; 28 monopolar and 4 bipolar channels, connected to the cap electrode array. The
Neuroscan Scan 4.1 (EEG analysis software) is used for recordings. The Polhemus Stylus
Fastrack digitizer is used to localize the electrode positions and to acquire the head shape.
2.2.5 Inverse Problem
The localization of neuronal activity based on measurements of electric potentials (EEG)
and/or magnetic field (MEG) involves solving two distinct but related problems. The “forward
problem” consists in calculating the electric potential and magnetic field outside the head,
given current sources and conductive properties of the tissues. The “inverse problem” consists
in determining the properties of the current sources within the brain based on the external
electric and/or magnetic measurements.
For a problem to be well-posed in the Hadamard sense, it must meet the following criteria: 1)
for each set of data, there exists a solution; 2) the solution is unique; 3) the solution depends
continuously on the data. While the bioelectric forward problem is well defined and is
governed by the quasi-static limit of Maxwell equations the inverse problem based solely on
EEG and/or MEG measurements lacks two of the three criteria for being well-posed: there is
not a unique solution and the solution doesn't depend continuously on the data. The lack of
uniqueness means that for any distribution of EEG and MEG signals outside the head there are
infinitely many possible configurations sources within the brain that are consistent with these
recordings, a principle originally pointed out by Helmoltz [1853]. The second property means
that small errors in measurements may cause large variations in the solution. This indicates that
there is a serious mathematical problem but in practice and especially for superficial sources
this does not lead to serious practical limitations.
Additional information or assumptions are added in order to solve the inverse problem and
these include [Malmivuo and Plonsey, 1995]: using simplified models for the source and the
Chapter2.Imaging brain structure and function
23
volume conductor; imposing physiological constraints; recognizing the typical signal pattern
associated with certain source configurations (empirical approach); examining the lead field
pattern from which the sensitivity distribution of the lead and therefore the statistically most
probable source configuration can be estimated.
The model for the volume conductor can be a simple, spherical model or realistic, based on the
MRI of the head [Wieringa and Peters, 1993] which allows the extraction of the 3D geometry
of the brain and surrounding tissues. Numerical methods like the boundary element or volume
element methods are used in the later case. The necessity of using realistic models, particularly
those accommodating inhomogeneous volumes, is stronger for EEG than for MEG [da Silva
and Van Rotterdam, 1999]
The simplified signal source models traditionally used are equivalent current dipoles ((ECDs)
which should have good correspondence with the physiology and anatomy associated with the
actual source distribution. In the case of simplified source models at least as many independent
measurements are made as the model has independent variables in order to have as many
equations as unknowns and to evaluate the variables of the model [Malmivuo and Plonsey,
1995]. To reduce the sensitivity to noise the number of measurements greatly exceeds the
number of variables in the source model. The over specified system of equations is solved
using least squares approximation and possibly other constraints to improve the stability.
For the single moving dipole the position and moment are determined for each time moment by
a separate fit of the field. The position is allowed to move as time progresses. For the fixed or
rotating dipole the time course of the dipole’s strength is also specified, in addition to the
position, and in the case of the rotating dipole also the orientation. Time-varying multidipole
models were used with complex global optimization and fitting methods.
In the time-varying dipole model introduced by Scherg and von Cramon [Scherg and Cramon,
1985] an epoch of data is modeled with a set of spatially fixed dipoles whose amplitudes vary
with time. De Munck [deMunck, 1990] proposed the following approach based on the
stationary dipole model, proposed by Scherg and Von Cramon, and separating the parameters
into linear and nonlinear. First the functions that describes the change of the source as a
function of time must be estimated for a given position and orientation of the ECD; second the
orientation must be found given the time functions and position. These steps are performed
iteratively till the best time functions and orientation are found for a given position. Then the
nonlinear position parameters are updated and the process repeated until a best fit is obtained.
The localization accuracy reported for dipolar inverse solutions is smaller than 13 millimeters
[Leahy et al., 1999]. One drawback of this approach is that ECDs do not indicate the activated
Chapter2.Imaging brain structure and function
24
volume but only its gravity center. Also the use of dipole models is meaningful if the activity
has a focal character and the number of possible sources can be anticipated with reasonable
accuracy [da Silva and Van Rotterdam, 1999].
The continuous current source mode is more biologically plausible than the discrete current
dipole model and readily allows the incorporation of external information, e.g fMRI [Liu et al.,
1998]. The number of solutions in this case is in general much larger than the number of
sensors and the problem is underdetermined. The minimum norm algorithm introduced by
Hämäläinen and Ilmoniemi in 1984 was the first effort in this direction. An algorithm using
direct Fourier based inversion was proposed by Dallas and colleagues [Dallas 1985; Kullmann
and Dallas, 1987] and reconstructions confined to an anatomically defined surface [Wang et
al., 1992]. More general linear frameworks for 3D reconstructions were introduced by Sarvas
[1987] and a probabilistic approach by Clarke and Janday [1989]. Magnetic Field Tomography
(MFT) introduced a two step procedure for the estimation of the general parameters, a
probability weight which could be used for the incorporation of additional information and an
iterative procedure to solve a non-linear algorithm for the estimation of the current density
(Ioannides et al. 1990). The MFT algorithm adopted the theoretical framework of Clarke and
Janday but used problem-defined expansion functions (the lead fields used by the original
minimum norm method) which together with the non-linear estimation algorithm were capable
of tomographic reconstructions of focal or distributed generators, without prior assumptions
about their number or nature [Ioannides et al 1989; 1990]. MFT allows the 3D reconstruction
of the dynamic brain activity with a spatial resolution of 2-5 mm at the cortical level [Ribary et
al., 1991; Moradi et. al., 2003], which deteriorates with depth. This procedure produces more
focal images than the traditional minimum-norm solutions. Another possibility is to use a
MUSIC-type probability weighting [Mosher, 1992] combined with cortical constraints to focus
the image [Dale and Sereno, 1993].
The techniques for estimating the distributed sources may include Bayesian approaches [Baillet
and Garnero, 1997; Phillips et al., 1997; Schmidt et al., 1999] or constrain the current location
and orientation perpendicular to the cortical surface [Wang et al., 1992; Dale and Sereno, 1993;
George et al. 1995; Baillet and Garnero, 1997]. The use of constraints has been introduced as a
restriction of the activated regions to foci identified by similar fMRI experiments [Di Russo et
al., 2002]. Adaptive beamformer techniques were used for MEG [Van Veen et al., 1997; Gross
and Ioannides, 1999; Sekihara et al., 2001] but these work best if the sources are uncorrelated.
Methods for tomographic reconstructions are also available today for EEG [Michel et al., 2001;
Stern and Silbersweig, 2001] but these tend to have less resolution than the corresponding
Chapter2.Imaging brain structure and function
25
reconstruction for MEG data, at least when simple models for the conductivity profile are used
for the forward problem.
2.2.6 MEG: Clinical and research application
Due to EEG's long tradition and its role as a gold-standard in certain clinical applications it is
anticipated that simultaneous MEG and EEG will be used as complementary tools.
MEG has been used in normal subjects to give information on brain functions including touch,
speech, vision, hearing and motor skills but also in the assessment of pathological functional
deficits, neuropharmacological investigations, trauma assessment, epileptic assessment and a
growing list of research investigations in neuroscience and psychiatry.
MEG can contribute to clinical care in: (1) the pre-surgical MEG mapping of neurosurgical
candidates who have brain tumors or vascular malformations in the region of the sensori-motor
strip; (2) the assessment of pathological functional deficits or evaluation of residual
neurological function after severe head injury; (3) the assessment of pathological functional
deficits in epilepsy, the most common clinical application of MEG being in the planning of
epilepsy surgery; (4) assessment of traumatic brain injury. In the case of mild head trauma
patients may have normal MRI and CT and normal clinical EEG, but abnormal
neuropsychological function. MEG provides evidence of traumatic brain injury (for ex.
anomalous delta and theta bursting) in this population with a higher sensitivity compared to
MRI and EEG.
Also the role of MEG has been studied for in: (1) fetal brain activity monitoring; (2)
schizophrenia [Rojas et al., 2002] or; (3) migraine [Bowyer et al. 2001], (4) neurological
evaluation of neuropharmaceuticals.
The comparison of pathological (in the presence of tumor, infarct, injury, etc.) versus normal
subject responses following auditory, visual or somatosesnsory stimulation may reveal
abnormal sources location and response patterns. Somatosensory evoked magnetic fields have
been employed for example for assessment of traumatic brain injury.
MEG shows potential as a diagnostic method which could replace invasive EEG monitoring in
some patients, guide intracranial electrode placement, and reduce the length of EEG evaluation
prior to surgery.
2.2.7 Coregistration of MRI and MEG
The location of the activated area is best examined in the background of the anatomy, using the
MRI. In our case MRI images were acquired for each subject with either of a: 1. 1.5-T Siemens
Chapter2.Imaging brain structure and function
26
Magneton Symphony or 2. Varian Unity Inova 4T system. The co-registration of MEG and
MRI data is performed using 3 or more landmarks on the head of the subject. These points are
digitized in the MEG coordinate system using a position sensing device, a Polhemus Isotrack
device in our case. This locates the dewar and gradiometers with respect to a head-centred
coordinate system.
Before MRI scanning, capsules containing vitamin E oil or another contrast media can be
attached to the fixed marks on the head. These show up as bright white marks on the MRI
scans and allow a coordinate transform to be calculated which places the MRI slices in the
head-centered coordinate system. Alternatively more complex registration procedures can rely
on the head anatomy.
Figure 2.2-6 Landmark sites can be used for corregistration. Capsules of vitamin E are placed at the nasion
and preauricular points, and in the same places are placed the MEG localization coils. Images from
http://www.ctf.com/Pages/page27.html.
Before the MEG procedure (three) probe coils are attached to the scalp, one close to the nasion
and two close to the preauricular points (these are the same landmark sites as the Vitamin E
capsules used during the MR scan). The coil positions and subject’s head shape (a few
thousand points) are digitized with a three-dimensional digitizer (Polhemus, 3DSpace/Fastrak,
U.S.A.). The coils can be detected by the MEG system itself, form a coordinate system that can
be used to register the MEG and MRI together. The digitized points are then matched to the
surface of the scalp using dedicated software developed at RIKEN for this purpose. Before and
after each run the probe coils are activated and their positions identified by the CTF software.
This defines the exact position of the sensors relative to the coils and hence with respect to the
head of the subject and its MRI representation. The procedure yields an accurate
superimposition of the estimated sources on the corresponding MRI slice. This allows solutions
to be compared on the same anatomical background from runs of the same experiment or even
across different experiments on the same subject.
2.3 Conclusion
A problem in functional brain mapping is how to best combine the advantages of
Chapter2.Imaging brain structure and function
27
complementary techniques as those dedicated for anatomical imaging (MRI) and those for
functional imaging, like MEG and EEG or PET, SPECT. Unique among the functional
techniques because of a high temporal resolution, complemented with noninvasive nature,
MEG and EEG can provide new insights for understanding brain function. The measurements
are non-invasive, and make possible the repetition of measurements with the same human
subject.
The information EEG and MEG provide is considered to be complementary, as MEG is best in
detecting sources within the sulci of the cortex while EEG detects both tangential and radial
sources but has a lower spatial resolving power and is hindered by the high resitivity of the
skull. For MEG this layer appears as transparent and one more advantage is that it does not
require that EEG electrodes be pasted onto the scalp. However MEG is an expensive technique
yet because of the high costs involved in creating the conditions for a high sensitivity magnet
and the suppression of magnetic background.
Chapter3. Brain segmentation
28
Chapter 3. Brain Segmentation
Chapter 3. Brain Segmentation ........................................................................................................28
3.1 Introduction ...................................................................................................................... 28
3.1.1 Definitions related to image segmentation...............................................................28
3.1.2 Applications of segmentation...................................................................................29
3.1.3 The problem of brain segmentation .........................................................................32
3.1.4 Mathematical morphology for image analysis .........................................................35
3.1.5 White matter – gray matter separation .....................................................................38
3.2 A mathematical morphology based method for brain segmentation................................ 40
3.3 A modified fuzzy c means method for white matter-gray matter separation................... 44
3.4 Geometrical, differential properties of the cortex ............................................................ 47
3.5 Results .............................................................................................................................. 51
3.1.6 Whole brain segmentation........................................................................................51
3.6.1 Gray matter- white matter separation with bias field compensation........................52
3.6 Discussion and Conclusion .............................................................................................. 53
3.1 Introduction
3.1.1 Definitions related to image segmentation
Image segmentation is defined as partitioning of an image into non-overlapping, constituent
regions, homogeneous with respect to image characteristics like intensity or texture [Pham et al.,
2000].
The segmentation problem is to determine the sets Sk included in the image domain I, whose union
is the entire image I. The sets that make up the segmentation result must satisfy:
UK
kkSI
1=
= , where =jk SS I Ø ; for k≠j and each Sk is connected.
When the connectedness constraint is removed the problem is called pixel classification and the
sets are called classes.
Chapter3. Brain segmentation
29
Labeling is the process of assigning a meaningful designation to each region or class and can be
performed separately from segmentation. Essentially the numerical index k of a set Sk is mapped
to an anatomical designation. Often in this case the value of K is assumed to be known based on a
priori knowledge of the anatomy. An example of labeling is the assignment of labels to healthy
versus tumoral tissue type.
To differentiate between classification and segmentation one shall think of segmentation to be a
top-down parceling of an image into anatomically meaningful continuous groups of voxels; while
classification shall be viewed as the bottom-up (data driven) labeling of voxels with a tissue class
label without demanding spatial contiguity for a class. The image data represent only one measure
(or a few measures in the case of multi-spectral data) concerning the underlying anatomy, and by
itself is sufficient only for classification. Anatomically distinct regions of the brain are
differentiated on the basis of histology, cytoarchitecture, connectivity, cytochemistry or function.
As such, data from external sources are required to constrain and guide the segmentation process.
3.1.2 Applications of Segmentation
Segmentation has application in robot vision, object recognition and medical imaging.
Segmentation is an important issue in biomedical engineering with applications at different levels:
from large anatomical structures like: brain, heart, knee, jaw, spine, pelvis, liver, prostate, and
blood vessels till the microscopic, cellular or nuclear level.
Frequently used to improve visualization of medical imagery and allow quantitative measurements
of image structures, segmentations are also valuable in building anatomical atlases, researching
shapes of anatomical structures, tracking anatomical changes over time, and cell counting.
The traditional method of medical image analysis, used to be the inspection of two-dimensional
grayscale images produced on films placed on a light box. New imaging techniques produce
digital images, carrying a lot more information stored in a format (DICOM for example) suitable
for sharing and communicating data. One main advantage of these techniques relates to their
suitability for detailed quantitative analysis of the location, appearance, size, or shape of patient
anatomy and the segmentation is one important step in this analysis. Starting from the
segmentation the analysis may include three-dimensional visualization, volumetric measurement,
shape analysis, image-guided surgery, and detection of anatomical changes over time.
Chapter3. Brain segmentation
30
3.1.1.1 Visualization
Segmentation of medical imagery allows the creation of 3D surface models which can be then
inspected from all angles and also sliced to offer the traditional representation of 2D grayscale
images, this time at any orientation and position by means of interpolation algorithms.
3.1.1.2 Volumetric Measurement
Measurement of the volumes of anatomical structures is necessary in medical studies, both of
normal anatomy and of various pathological conditions or disorders.
One example comes from the study of schizophrenia, where volume measurements of the lateral
ventricles, structures in the temporal lobe such as the hippocampus, amygdala, and
parahippocampal gyrus, the planum temporale, and the corpus callosum are used to study the
variation in neural anatomy between schizophrenic and control patients.
3.1.1.3 Image registration
Improvements in segmentation aid the techniques for image registration, such as surface matching
methods, landmark matching or atlas-based registration.
3.1.1.4 Shape Representation and Analysis
Various quantitative representations of shape are studied in order to mathematically describe
salient anatomical characteristics.
One example of a shape representation is a skeleton, a construct which is similar to the centerline
(or medial axis) of a segmented structure. One way to imagine a skeleton is the “brush fire”
approach: one thinks of simultaneously lighting fires at all points on the boundary of the structure.
The fires burn inward, traveling perpendicular to the boundary where they started, and then
extinguish when they hit another fire. The connected “ash” lines left where the fires extinguish is
the skeleton of the structure.
A richer shape representation is the distance transform, a function that measures the distance from
each point in a structure to the nearest point on that structure’s boundary. The distance transform
can also be imagined with the pyrotechnic approach: it is the time that the fire first reaches each
point in the structure. Consequently it is considered richer than the skeleton, since it contains more
information.
Chapter3. Brain segmentation
31
Shape representations are important for quantitative anatomical comparisons. Distance transforms
shape representations have been applied to the classification of anatomical structures in a study
that aims to differentiate between the hippocampus-amygdala complexes of schizophrenics and
normals [Golland et al., 2000]. An example of grayscale MR image data and the shape
representation derived from it for this study can be seen in Figure 3.1-1.
Figure 3.1-1. Shape representation example. A segmentation of the hippocampus-amygdala complex (left), a 3D
surface model of the hippocampus-amygdala complex (center), and a distance map used to represent the shape
of the hippocampus-amygdala complex (right) [from Golland et al., 2000]
Shape representations can also be used to aid the segmentation process itself by providing
anatomical knowledge. A generative shape model, trained from a population of shape
representations, can be used to visualize new shapes according to the learned modes of variance in
the shape population (allowing visualization of “average” anatomy and of the main anatomical
variations that may occur). Then, at each step of the segmentation of new data, fitting the model to
the current most likely segmentation can provide anatomical information to the algorithm.
3.1.1.5 Image-Guided Surgery
In order to remove brain tumors or to perform difficult biopsies, surgeons must follow complex
trajectories to avoid anatomical hazards such as blood vessels or brain areas which are of
functional importance to the patient (speech, use of right hand, etc). Before surgery, path planning
and visualization is done using preoperative MR and/or CT scans along with three-dimensional
surface models of the patient’s anatomy
During the procedure, the results of the preoperative segmentation may still be used: the surgeon
has access to the pre-operative planning information, as three-dimensional models and grayscale
data are displayed in the operating room. In addition, “on-the-fly” segmentation of real time
imagery generated during surgery has been used for quantitative monitoring of the progression of
surgery in tumor resection and cryotherapy. The tissue may be deformed during the process.
Chapter3. Brain segmentation
32
3.1.1.6 Change Detection
For longitudinal studies, spanning over a lengthy time interval, segmenting regions of interest is
crucial for quantitative comparisons. Multiple sclerosis (MS) is a disorder that progresses over
time, accurate temporal measurements of neural changes may lead to a better understanding of the
disease. The goals of this kind of projects are analysis of lesion morphology and distribution,
quantitative evaluation of clinical drug trials, and monitoring of disease progression in individuals.
To this end, automatic segmentation is used to identify MS lesions. The volume of such lesions, as
measured from segmented data, has been shown to correlate with clinical changes in ability and
cognition.
3.1.3 The problem of brain segmentation
Cortical segmentation using MRI has been one of the most needed items on the list of engineering
in medical research. It is one of the most pressing needs for neuro-anatomical analysis because it
helps in: (1) the quantification of cortical thickness and cortical folding curvatures [Zeng et al.,
1999] and (2) in determining the spatial inter-relationships of the neuro-anatomical structures. The
quantification of specific regions is required for long-term monitoring of a disease progression or
remission. Since the manual cortical segmentation methods are subject to errors both in accuracy
and reproducibility and are time-consuming, fast, accurate and robust semi-automatic or
completely automatic techniques are needed.
The problem of brain segmentation is challenging due to the complexity of the images and of the
structure of interest. Manual segmentation is extremely tedious for large structures, especially for
the gray or white mater boundaries. The complexity of the cortical sheet folding makes slow and
impractical even the segmentation for a single subject. Moreover, the manual segmentation may
result in large inter and intra-segmenter variability. Automatic methods are desirable since studies
which compare different aspects of brain morphology across subjects typically involve vast
amounts of data and therefore manual segmentation is very time consuming. There are still
challenges to be met for the “automatic” methods. These mainly relate to the complex shape and
appearance of structures, i.e. anatomically distinct tissues can have the same gray-level appearance
while the same tissue can have different gray-level appearance in different regions of the image
due to distortions caused by the imaging equipment.
Traditional image segmentation methods rely on intensity values and thresholding. The
thresholding process is complicated by the overlapping intensity distributions of brain and non-
Chapter3. Brain segmentation
33
brain structures (skin- gray matter, fat, bone, and optic nerves –white matter), image acquisition
limitations (intensity inhomogeneity, partial volume effects, etc), and motion artifacts. One effect
of the intensity inhomogeneity may be that the same tissues do not have a unique range of values
and the histogram analysis for different image regions shows that the grey value ranges may vary
from slice to slice for the same type of tissue, and even within the slice.
One solution to this problem is to identify the largest connected component with gray values
corresponding to the brain. The method proposed by Höhne and Hanson relies on the use of
mathematical morphology [Höhne and Hanson, 1992]. The next step would be to ensure that the
topology is equivalent to a sphere. If this is not the case the topological defects are removed using
editing tools [Teo et al., 1997] or automatically [Kriegeskorte and Goebel, 2001]. Later
developments use this surface to initialize a deformable surface which aims to refine the boundary
[Kapur et al., 1996]. The model be placed outside the brain but these methods have difficulty
progressing within the narrow sulci the sulci or from within, from the WM/GM interface which
will contain larger sulci. An approach relying on the use of double deformable surfaces has been
proposed by MacDonald [MacDonald et al., 2000] for identifying both the internal and external
surfaces of the gray matter.
In general these methods need an initial model, based on previous segmentations, or a statistical
atlas. Spherical models circumvent this need and have been applied for structure’s segmentations
[Tek and Kimia, 1997]. But these models still need to be placed in the proximity of the structures
boundaries for good results and sometimes need to be manually initialized. These methods may
require also the transformation of the current brain and may therefore introduce uncertainties.
Modern techniques for segmentation relying on deformable surfaces require
Some of the main difficulties in achieving a robust and fast cortical segmentation are:
(1) Partial volume averaging (PVA): This is caused by the finite extent of an imaging system's
point spread function (PSF). If there is more than one tissue within the extent of a partial density
function or one voxel, then PVA is prominent;
(2) Tissue inhomogeneity and non-uniformity are intrinsic properties that add to the boundary
fuzziness regardless of the imaging system's quality. Due to the spatial inhomogeneities in the
radio-frequency (RF) gain in the RF coil, the intensities associated with these two tissues overlap;
(3) Shading artifacts: There are three kinds of artifacts: (i) hardware, (ii) those due to MR physics
and (iii) patient-related. The hardware related artifacts include: zipper artifact, corduroy artifact,
dot artifact, data clipping artifact, spurious echo artifact, coherent ghosting artifact, and calibration
Chapter3. Brain segmentation
34
artifact. The MR physics related artifacts are: magnetic susceptibility artifact, chemical shift
artifact, truncation artifact and criss-cross artifact. The patient-related artifacts are a result of the
motion artifacts due to voluntary or involuntary patient movements.
(4) Random noise associated with the MR imaging system;
(5) Convolutedness and variability of the brain structure. This is due to the complex topology
causing bends and twists. Besides that, this morphological shape differs from subject to subject.
(6) Variability in tissue types: This is due to the number of tissue types and connectivity (classes)
present in the tissue volume such as optic nerve and blood vessels.
(7) Size and types of brain tumor. Current techniques are suitable mainly for medium to large-
sized tumors.
(8) Operator variability: This is due to the variability in tracing (intra-and inter-observer
variability) of the cortical boundary/regions for image segmentation algorithms;
(9) Error susceptibility. This refers to failures for fully supervised methods;
(10) Imaging variability: Variability in imaging parameters such as inter-scan interval, voxel
dimension, signal-to-noise ratio, position and orientation of the subject in the scanner cause
complications in the segmentation process
(11) Availability of shape models: This accounts due to the absence of explicit shape models that
capture the deformations in human brain anatomy and topology. Work has been done in this
direction, especially for gray matter structures and the parietal and occipital lobes 20 normal MR
brain data sets and their manual segmentations are provided by the Center for Morphometric
Analysis at Massachusetts General Hospital and are available at http://neuro-
www.mgh.harvard.edu/cma/ibsr.
When choosing among the many segmentation approaches the following factors shall be taken into
consideration: (1) the application used, (2) the kind of supervision needed/not needed, (3) the
accuracy, (4) the robustness desired, (5) the speed issues.
The segmentation techniques may be classified into three core classes. Techniques which use: (i)
the regional-based approach, (ii) the boundary/surface-based approach and (iii) the fusion of
boundary/surface with region-based approach.
Region-based techniques: (1) atlas-based and threshold; (2) mathematical morphology; (3)
probability; (4) clustering; (5) texture-based; (6) knowledge-based; (7) neural-network; (8) region-
linking with hyperstack and (9) fusion of the above. Probability-based techniques are further
classified into: (i) Bayesian-based, (ii) expectation-minimization-based and (iii) Markov Random
Chapter3. Brain segmentation
35
Field-based. Clustering-based techniques are classified into: (i) FCM and (ii) AFCM types. A class
of techniques exists which consists of the fusion of: (i) texture with pixel-classification-based, (ii)
knowledge-base with pixel-classification, (iii) neural-network based with pixel-classification and
(iv) pixel-classification with edge-based.
Boundary/surface-based techniques are classified into four major types: (i) edge; (ii)
reconstruction; (iii) parametric and (iv) geometric-based. Parametric-based techniques are further
classified into two types: (1) 2-D-based and (2) 3-D-based. The 3-D based parametric-based
techniques are further classified into: (1) ribbon; (2) topological-surface; (3) constrained
parametric; and (4) parametric-based with threshold/connected components.
Boundary/Surface fused with Region techniques are classified into two types: (i) parametric and
(ii) geometric-based. Parametric-based is classified into two types: (1) parametric fused with EM-
based and (2) parametric fused with clustering. Geometric-based are classified into four types: (1)
level sets with Bayesian-classification; (2) coupled level sets with classification; (3) level sets with
clustering, and (4) level sets with shape modeling.
One categorization can be made in model based methods or methods which do not rely on a model
to initialize the segmentation. Our method belongs to the first class.
Within the methods relying on intrinsic parameters found in the image there are multispectral or
unispectral methods, ours uses T1 weighted images only. At the chore of the method is the
application of mathematical morphology. A short introduction in mathematical morphology is
given in the next section 3.1.4.
3.1.4 Mathematical morphology for image analysis
Mathematical morphology (MM) was introduced in 1964 by Matheron and Serra, [Serra, 1982]
from a need to analyze texture in petrography. Later on MM methods came to be used not only for
studying the arrangement of spectrographic phases, milling of rocks but also for brain histology,
dynamics of cloud movements, automatic reconstruction of relief maps, character recognition,
segmentation of cancer cells, etc. These methods use the notion of geometrical structure, or
texture, quantified by introducing the concept of “structuring element”. The structuring elements
interact with the object of study, modifying the shape and reducing it to a form more expressive
than the original. The main objective of morphology is to reveal the structure of the objects by
transforming the sets which model them.
Four principles must be satisfied by every morphological transformation and its measure, namely:
Chapter3. Brain segmentation
36
(i) translation invariance; (ii) scale invariance; (iii) locality (local knowledge); and (iv) semi-
continuity.
Mathematic ally these properties can be expressed:
(i) hh xX )]([)( Ψ=Ψ ;
(ii) )()( XX Ψ=Ψ λλ
(iii) :ZboundedZbounded ∃′∀ ZXZZX ′∩Ψ=′∩∩Ψ )()]([
The transformation Ψ satisfies the local knowledge principle if, for any bounded set Z’ in which
we want to know Ψ x , we can find a bounded set Z in which the knowledge of X is sufficient to
locally perform (within Z’) the transformation.
(iv) For every decreasing sequence of closed sets tending towards a limit∆ , and every increasing
transformation1 Ψ , there must correspond a sequence of transformed sets tending towards the
transform of ∆ . In other words, the boundary of the transform is equal to the transform of the
boundaries.
A central place in mathematical morphology is occupied by the hit or miss transform, a point by
point transform of a set.
If X is a set and B a structuring element and suppose B is centered at the point x. with components
B1 and B2. Suppose B is ceneterd in x and we denote this by Bx1, Bx
2. A point x belongs to the hit
or miss transform BX ⊗ of X, if and only if Bx1 in X and Bx
2 is included in the complement Xc of
X.
{ }cxx XBXBxBX ⊂⊂=⊗ 21 ;:
It was thought that characterizing quantitatively a structure mean assigning it a few representative
parameters (for the plane: area, number of intercepts, number of particle). Behind each parameter
lies one elementary transformation. The area can be interpreted as the number of times a test point
(belonging to the structuring element) hits the image under study when translated on the grid,
vertex to vertex. The number of intercepts has behind a transformation involving a pair of
consecutive points. The connectivity number can be calculated in a similar manner. The smallest
structuring element of 0 dimension (the point), of 1 dimension (the doublet) and two dimensions
(the triplet) open the way to the area, length, and number measurement on the set X. If the 1 An increasing transformation satisfies the following: if A included in B then the transform of A
is included in the transform of B.
Chapter3. Brain segmentation
37
dimension of the structuring element is added to the dimension of the parameter measured we
obtain a constant, which represents the number of dimensions of the object space.
Two basic operators are dilation and erosion.
Erosion is a particular case of the hit and miss transform, where B2 is the empty set.
The eroded set X B of X is defined as the locus of centers x of Bx included in the set X:
X B }|{ BBx x ⊆= =IBb
bX∈
The dilate BX ˆ⊕ is the locus of the centers of the Bx which hit X. The definition of dilation is
given relative to the transposed (reflected) set of B, so as to make an analogy with erosion.
UXx
xBBX∈
=⊕ ˆˆ ≠∩= ]|{ BBx x Ø}
The algebraic properties of the dilation and erosion dual transforms include: a) translation
invariance; distibutivity; iterativity; increasing and inclusion.
After having eroded X by B it is not possible in general to recover the initial set by dilating the
eroded set by the same structuring element B. This dilate reconstitutes only a part of X, simpler
and with less details , but may be considered as that part which is the most essential.
The successive application of erosion and dilation results in opening of X with respect to B:
XX B (= BB ⊕)ˆ
The effect of opening is that it smoothes the contours of X, cuts the narrow isthmuses, suppresses
small islands and the sharp capes of X.
The properties of opening include: antiextensive XX B ⊂ , increasing BB XXXX '' ⊂⇒⊂ ,
idempotent BBB XX =)(
The closing of X with respect to B is the result of applying successively dilation followed by
erosion:
XX B (= )B̂⊕ B
By dual interpretation with opening, the effect of closing is that it blocks up narrow channels, the
small lakes and the thin gulfs of X.
The properties of closing include - extensive: XX B ⊃ , increasing BB XXXX ⊂⇒⊂ '' ,
idempotent BBB XX =)(
Analyzing the same object X by dissimilar structuring elements (like disk or a set of points) may
result in very different pieces of information on its geometric structure. The structuring elements
Chapter3. Brain segmentation
38
can be chosen from a wide number of possible shapes but they must be geometrically simple,
bounded. From multiple shapes with similar properties the most extreme being: isotropic (disk)
versus anisotropic (segments) for studies of orientation, convex (disk or segment like) for size
distribution versus nonconvex (circle, boundary of the segment or clusters of points) for set
covariance, infinitesimal iterations.
3.1.5 White matter – gray matter separation
The segmentation in different classes has applications in image-guided surgery and neuroscience.
In image-guided surgery, 3D visualization are created by rendering 3D models of the brain with
overlays of tumors for visualization in pre- and intraoperative surgical planning.
The analysis of MRI of the human brain often involves white matter (WM) segmentation.
In the functional brain mapping the surface of interest may be the WM/CSF interface, the
WM/GM interface or run midway through the white matter (WM/GM). One of the reasons for
preferring the GM/WM surface display is that the folding of the cortical sheet obstructs the view
of activations buried within sulci if the GM-CSF interface is used for visualization. The
visualization of the WM-GM interface allows a better view of the activations in the sulci. A
second reason relates to the higher accuracy of the white matter-gray matter boundary surface
segmentation, as compared to the gray matter- CSF boundary [Dale et al., 1999].
The quantification of cerebral volume is important for assessing brain development and/or changes
in the normal brain or in pathological states [Iwasaki, 1997]. The segmented volumes of white and
gray matter (GM) in different parts of the brain are measured and compared in order to develop
hypotheses about various disorders.
For example in studies of ageing was observed a decline of global gray matter with age, involving
cortical and deep gray matter structures and cerebellum diffusively, but not significantly for white
matter [Good et al., 2001]. Separating gray matter from white matter is particularly important if
one considers that gray matter reduction is relevant in schizophrenia [Lawrie and Abukmeil, 1998;
McCarley et al. 1999] and Alzheimer’s disease [Jack et al., 1997], while myelination is an
indicator of brain maturation [Giedd et al 1999; Pauss et al., 1999]. Myelination can also serve as a
marker showing if brain abnormalities are related to the neuronal body (nonmyleinated: gray
matter) or to the axons (often myelinated: white matter).
Chapter3. Brain segmentation
39
Significant correlations between brain volume and other putative MR neuronal markers indicate
that atrophy reflects axonal loss in multiple sclerosis. There has been considerable interest in
measuring tissue loss (atrophy) as a global marker of multiple sclerosis outcome.
Brain white matter bulk consists predominantly of axons (46%) followed by myelin (24%), and
progressive atrophy implies loss of these structures, especially axons, although variable effects on
tissue volumes may also arise from glial cell proliferation or loss, gliosis, inflammation and edema
[Miller et al., 2002].
Intensity-based segmentation of the GM, WM compartments is challenged by the complex
geometry of the gray matter-white matter interface and by the fact that a substantial part of the
voxels at the interface sample both white and gray matter (the partial volume effect).
The direct use of global thresholding is hardly suitable because of the presence of intensity
inhomogeneities. These may be caused by non uniformities in the RF field during acquisition as
well as other factors. For a detailed explanation see [Webb, 1988; Bushberg et al., 1994]. The
result is a shading effect or a bias field where the pixel or voxel intensities of the same tissue class
vary slowly over the image domain.
Several methods have been described to separate and estimate the white matter and gray matter
volumes. Some of these methods employ histogram analysis and assume that the histogram is
composed of a gray matter and white matter distribution. One method to separate the two
distributions is to find the best separation threshold in terms of variances within and between the
two distributions [Momenan et al., 1997; Otsu, 1979]. Other methods [Kennedy et al., 1979;
Wieringa, 1993] use as threshold the position of the local minima between the two distributions as
threshold. Similarly the Utrecht method [Schnack et al., 2001] uses shape information for the two
histogram peaks to find one single threshold for separating the white matter from gray matter. The
intensity threshold is found from the crossing of the tangential lines to the steepest slope points of
the polynomials fitted to the wm and the gm peaks, multiplied by a scaling factor obtained from
calibration. The calibration factor is used to compensate for noise and artifact influences. A
prerequisite of their technique is the intensity nonuniformity correction. Schnack’s method as well
as the nadir algorithm [Filipek et al., 1994, Kennedy et al. 1989] and Otsu’s algortihm [Otsu,
1979] start from an initial total brain segmentation. Otsu’s algorithm is a variance optimization
method, comparable to c-means which allows the separation into more than two classes.
Although most brain segmentation methods do not take into account the scanner nonuniformity
various methods have been developed to correct for the introduced bias [Wells et al., 1996;
Chapter3. Brain segmentation
40
VanLeemput et al., 1999] and some combine the segmentation with the bias field correction
[Kapur et al., 1996; Zhang et al., 2001].
It has been shown that intensity inhomogeneities are well modeled by the product of the original
image and a smooth slowly varying multiplier field. Pham and Prince [Pham and Prince, 1999]
have analyzed the performance of a classification algorithm based on fuzzy c means (FCM) in the
case where the bias field is scalar and vector field and concluded that the assumption of a scalar
field gives results similar to the vector field assumption while involving simpler and faster
computation. Another method for labeling the white matter and gray matter relying on fuzzy c
means and compensating meanwhile for the bias field is the one reported by Ahmed and others
[Ahmed et al., 2002]. They modify the objective function to be minimized in the FCM method in
such a way as to include the effect of the neighbors’s labeling. This approach should thus be
suitable for cases where the images are corrupted by noise.
3.2 A mathematical morphology based method for brain segmentation
We use a semi-interactive, region based approach, based on mathematical morphology for
extracting the entire brain. Our approach is similar to the ones described in [Höhne and Hanson,
1992; Kapur at al., 1996], and is shown schematically in Figure 3.2-1.
The method relies on the fact that the intensities at each spatial location of the MRI data can serve
for identifying tissue clusters and selected anatomical structures. The intensity information
however does not suffice in many cases because different tissues may have similar MR
characteristics and because of image acquisition artifacts like noise, RF inhomogeneity, partial
volume effects, etc.
The procedure starts by pre-processing the MRI i.e. by applying a filter for a noise reduction
purpose. We use median filtering or anisotropic diffusion filtering. For cases where the
inhomogeneity field is introducing major problems to the segmentation process the anisotropic
diffusion filter can be applied but this leads to a significant increase in the total time required for
segmentation. Practically the segmentation starts with interactive thresholding, using values
suggested from the smoothed histogram. All these factors introduce errors in the initial
binarization into brain and non-brain voxels. To overcome this problem, we use mathematical
morphology to isolate the largest connected component classified initially as brain, by cutting its
connection to non-brain structures.
Chapter3. Brain segmentation
41
Figure 3.2-1.Flowchart of the brain segmentation process done by the CORTSEG module
The core of the method consists in a succession of morphological operation starting from the
thresholded image, as shown in Figure 3.2-2.
Thin connections between regions with similar intensity values (bridging veins from the cerebral
cortex to dura, from dura to the skull, the optic nerves, etc.) are cut using erosion with a ball like
structuring element of radius 1. Interactive selection of the number of erosions is allowed. The
next step of the segmentation process is a region growing operation. The region growing starts
from a seed selected by the user, by mouse clicking on a voxel within the eroded mask. The
algorithm adds to the initial region all voxels connected to it and whose values are between the
selected thresholds. Dilation with a ball like structuring element is applied to restore the
boundaries of the brain, distorted in the erosion step. The dilation is conditioned on the initially
assigned brain labels, through thresholding, aiming that no boundaries are actually expanded. The
number of dilations is equal to the number of erosions. Closing is applied in the end to fill in small
holes, i.e. voxels that are not part of the brain but are surrounded by voxels who are.
Preprocessing
Erosion
Seed planting & region growing
Thresholding
Closing
Conditional Dilation
Surface rendering
MRI
Y
N OK?
Chapter3. Brain segmentation
42
a) original b) thresholded c) eroded
d) region grown e) dilated f) closed
Figure 3.2-2. The stages of the morphology based segmentation: a) original MR slice, b) the voxels in between
the selected thresholds are shown in pink, overlaid on the original MR slice (in gray) , c) in the eroded image
(shown in pink) the brain (B) and non brain compartments (nB) are disconnected through severing thin
connections (C), d) the region grown from a seed is in pink retrieves the brain compartment, e) the dilated
image restores the original size, f) the closed image fills in little holes (h) (voxels which are classified as brain
surrounding voxels classified as background)
The results at any stage of the segmentation process can be visualized and evaluated qualitatively
by varying the degree of transparency of the segmented slice overlaid on the original MRI slice
[Figure 3.2-3]. The segmented volume can be shown in varying color scales while the original
MRI is always shown in gray scale.
The output is either a binary volume with pixels classified as brain or non-brain or a volume
containing the gray value information for the brain volume.
B
nB
C
h
Chapter3. Brain segmentation
43
The methodology is implemented into the Cortical Segmentation (CORTSEG) software module,
whose interface is shown in Figure 3.2-3.
Figure 3.2-3. Cortseg interface
The segmentation result is visualized using the VISIO module, presented in Chapter 5.
Figure 3.2-4.Rendering of segmented brain
If unsatisfactory, the segmented brain can be edited and refined with STRUCTSEG, presented in
Chapter 4.
Chapter3. Brain segmentation
44
3.3 A modified fuzzy c means method for white matter-gray matter
separation
A new step in the segmentation procedure is necessary for separating the segmented brain into
classes: white matter, gray matter and CSF starting from an eventually noisy, inhomogenously
“illuminated” MRI scan.
Inhomogeneity artifacts in the MRI, often disregarded by segmentation protocols, can spoil the
quality of the segmentation. To avoid this problem, we use a modified fuzzy c-means method
incorporating neighborhood information for labeling a voxel and simultaneously correcting for the
bias field. The first order neighbors are considered only.
The fuzzy c means clusters data by interactively computing a mean intensity for each class and
segmenting the image by classifying each voxel in the class with the closest mean. The fuzzy
partition obtained is in the end defuzzified using the maximum membership conversion.
The MRI data are first segmented into brain and background interactively, based on thresholding
and applying a succession of mathematical morphology operators: erosion, region growing,
dilation and closing.
We assume that the gray matter is a sheet of approximately constant thickness and we proceed to
iteratively erode the initially segmented brain. From the statistical properties of the gray values of
the eroded and remaining voxels, we infer the optimal number of erosions and the initial mean
value for the WM and the GM clusters [Figure 3.3-1].
Figure 3.3-1: The intensity values of the eroded voxels (considered to be gray matter GM) increases with the
number of erosions, reaching a constant value, meanswhen no more gray matter has been left to be eroded.:.
This values is assigned to the white matter (WM) mean.
Chapter3. Brain segmentation
45
After a number of erosions the mean value of the voxels does not change too much, indicating that
only WM is left.
Assuming that the MRI recorded signal is the product of the true signal, the one generated by the
tissue, and the spatially varying gain field:
kkk GXY = , k=1, …N
Yk is the observed and Xk the true signal intensity and Gk the gain field, measured at voxel k. N is
the total number of voxels.
Applying the log transform allows the artifact to be modeled as an additive field [Wells et al.,
1996.]
kkk xy β+= Equation 1
where yk and xk are the true and observed log transformed intensities at the k-th vozel. βk is the
bias field at voxel k.
The standard FCM objective function for partitioning a set a set { }Nkkx 1= into c clusters is:
2
1 1ik
c
iik
pN
kvxuJ −=∑∑
= =
Where: c is the number of classes; N is the number of voxels; { }ciiv 1= are the prototypes of the
clusters and the partition matrix U=[uik] ∈ U must satisfy:
∀<<∀=∈ ∑∑==
N
kik
c
iikik iNuandkuu
11,0:,,1]1,0[ and p is a weighting factor describing the
amount of fuzziness of the resulting classification.
The modified objective function accounts for the neighborhood effect (the labeling of a voxel is
influenced by the labels in its immediate neighborhood) as well as for the bias field [Ahmed et al.,
2002].
−+−= ∑∑∑∑∑
Ν⊂= == = krxir
c
iik
pN
kRik
c
iik
pN
km vxu
NvxuJ 2
1 1
2
1 1
α
Nk is the set of neighbors in a window around xk and NR is the number of neighbors.
α is a parameter which controls the effect of the neighbors, a higher value being used for low SNR
images.
Substituting the kkk xy β+= Equation 1
Chapter3. Brain segmentation
46
−−+−= ∑∑∑∑∑
Ν⊂= == = kryirr
c
iik
pN
kRik
c
iik
pN
km vyu
NvxuJ 2
1 1
2
1 1βα
The optimization problem can be formulated as:
{ } { } mviU
JNkk
ci 11 ,,
min== β
, subject to U∈U
The zero gradient condition involved derivation of Jm with respect to uik, vi, and βk and yields three
necessary conditions for Jm to be a local minimum.
A Lagrange multiplier is used for solving the constrained optimization in equation. 1
−+
+= ∑∑∑
== =
c
iik
c
i
N
kiik
p
Rikik
pm uu
NDuF
11 11λγα
Where 2ikkik vyD −−= β and ∑
Ν⊂
−−=kry
irri vy 2βγ
We put the conditions that the derivatives of Fm with respect to uik, vi, and βk are null:
;0=∂∂
ik
m
uF ;0=
∂∂
i
m
vF ;0=
∂∂
k
mFβ
and knowing that the sum of all partitions at a point is 1: kuc
jjk∑
=
∀=1
,1 we obtain the values for
the partition matrix, the cluster centers and the bias field:
11
12
2
* 1
−
=
⊂
⊂∑∑
∑
−−+
−−+
=p
c
j
Nxjrr
Rjk
Nxirr
Rik
ik
kr
i
kr
vxN
D
vxN
D
u
βα
βα Equation 2.
( ) ( )
∑
∑∑
=
⊂=
+
−+−
= N
kik
p
Nxrr
Rkkik
pN
ki
u
yN
yuv kr
1
1*
)1( α
βαβEquation 3.
∑
∑
=
=−= c
iik
p
c
iiik
p
kk
u
vuy
1
1*β Equation 4.
Chapter3. Brain segmentation
47
The algorithm for segmenting the images into classes while correcting for the bias field can be
summarized as follows:
1. Select the initial class prototypes { }ciiv 1= and initialize the bias field { }N
kk 1=β with small and equal
values, for example 0.01.
2. Update the partition matrix using 1
1
12
2
* 1
−
=
⊂
⊂∑∑
∑
−−+
−−+
=p
c
j
Nxjrr
Rjk
Nxirr
Rik
ik
kr
i
kr
vxN
D
vxN
D
u
βα
βα Equation 2
3. Calculate the new clusters centers using ( ) ( )
∑
∑∑
=
⊂=
+
−+−
= N
kik
p
Nxrr
Rkkik
pN
ki
u
yN
yuv kr
1
1*
)1( α
βαβ
Equation
3.
4. Estimate the bias field using∑
∑
=
=−= c
iik
p
c
iiik
p
kk
u
vuy
1
1*β
Equation
4.
Steps 2) to 4) are repeated till ε<− oldnew VV , where V is the vector of cluster centers and ε is a
small number, for example 0.01.
We use α between 0.7 and 0.85, p =2 and Nr =9 for 2D and 27 for the 3D case.
The segmentation accuracy can be measured as: a) 100* Number of correctly classified
voxels/number of voxels in the class or b) using the Dice coefficient, defined as (2*Volume of
voxels assigned to class k by both the ground truth and the algorithm)/(volume of voxels assigned
o k by algorithm + volume of voxels assigned to k by ground truth) - this coefficient approaches 1
if both segmentations are in agreement and 0 if there is no overlap.
3.4 Geometrical, differential properties of the cortex
The cortical surface is an important feature of the brain, but its precise folding, geometry and
variability are little understood. Once the cortex has been extracted it can be represented as a set of
surface patches. We aim to find a parametric representation and infer the local geometrical
properties and differential properties. A precise mathematical representation of a typical cortical
surface can be done by means of a family of local quadratic [Joshi et al., 1995]. The local
Chapter3. Brain segmentation
48
structural information relevant to shape is embodied in the principal curvature, normal and
principal direction fields over the surface.
Differential geometry provides a natural mathematical framework to study the cortical surface
geometry. Among many geometric quantities defined for the surface, curvature information is
valuable since it quantifies the structure of the sulci and gyri, and provides the basis for
comparative studies.
This representation allows to make statements about the intrinsic geometry of the surface as well
as to map one surface geometry onto another or onto an atlas based surface and thus to study the
variability of the cortical surface in an ensemble of brain data. These features are thus relevant for
spatial transformation and registration of brain images using for example elastically deformable
models [Davatzikos, 1997].
The digital and mathematical representation allows the precise tracing of sulci or/and gyri onto the
cortical surface, based on the sign of the mean curvature. This is important for parcellating the
cortex in distinct entities and helps understanding the structure and functionality of these areas.
The segmented 3D object can be represented as a set of voxels, or points, and the isosurface
delimiting the boundary with the background as a set of surface patches. We are interested in
finding surfaces defined in the neighborhood of one of their points (surfaces in the small).
We define surfaces to be manifolds supported by local coordinate systems with at least two
derivatives allowing for the generation of curvature maps. We fit quadratic functions to the
elementary patches obtained from triangulation. Practically we want to estimate the parameters
that minimize the mean square distance from the data points to the surface defined by those
parameters.
Consider a surface M. This can be defined as the set of zeros of a smooth function of 3 variables x,
y, and z : f(x, y, z)=0 . It can also be defined parametrically as:
S={(x, y, z): x=d(u,v), y=e(u, v), z=f(u,v), u, v belong to D in R2 }
Or in a less general form, equivalent to the graph surface (or Monge patch):
S={(x, y, z): x=u, y=v, z=f(u,v), u, v belong to D in R2 }
The points on the graph that are directly connected to a vertex i. form the neighborhood N(i). For
each point i the surface(M) is expressed as the graph of the function z(x, y) of f (x, y).
We determine using the triangulated 3D volume, the mesh vertices and polygons connectivity, and
compute the mesh normals for the set of polygons.
Chapter3. Brain segmentation
49
We determine the neighborhood of each vertex xi and the tangent plane at the vertex point i. and
the tangent plane. Each point in the tangent plane, x satisfies: ( ) 0=−• ii xxn rrr
We can define an orthonormal coordinate system with xi as origin and choose the basis vectors in
the tangent plane:
Where: ( )zyx nnnn ,,=r
We have to express coordinates of the vertex i and of its neighbors in the new system:
coordinates transformed= (coordinates –fiducial)*(transforormation matrix)T
We store the data relative to the patches: fiducial, neighbors, vertices (including fiducial and
neighbors), normal, transformation matrix (the new vector basis).
In a neighborhood of the fiducial point, the surface can be given in the form:
S=(u, v, f(u, v))
The coordinate vectors: ),1,0(),0,1(vfXv
ufXu
∂∂
=∂∂
= are linearly independent, thus the
representation z=f(u, v) yields a regular surface as long as f(u, v) has a few continuous derivatives.
We fit the surface with a quadric: 2)2()1(
2)0(),( vguvgugz vu ++=
The tangent plane is determined by Xu and Xv .
The normal vector X3: 2)(
3XvXuXvXuX
×
×=
In the fiducial point (0, 0, 0): )0,0()1()2()0,0()1()0( )2,1,0(;)2,0,1( ugvgXvvgugXu +=+=
Let s be the length of the regular curve X(t)=X(u(t), v(t))
The first fundamental form (given by the inner product of the tangent vectors) determines the arc
length of a curve on a surface and satisfies: 22
2
++
=
dtdvG
dtdv
dtduF
dtduE
dtds
121 bnbandaab i ×=⟩⟨
=
≠+−
≠+−
≠+−
=
0,)),(,,(1
0,)),(,(1
0,),),((1
zT
zxzzz
yT
yzxyy
xT
xxzyx
nifnnnnn
nifnnnnn
nifnnnnn
a
Chapter3. Brain segmentation
50
Where: vvvvvuuvuuuu XXXGXXXFXXXE ======
The second fundamental form is given by: 22 2 NvMuvLuII ++=
Where: 333 XXNXXMXXL vvuvuu ===
The normal curvature in a direction vp=axu+bxv is given by 22
22
22)(
GbFabEaNbMabLavpk
++++
= .
The maximum and minimum values of the normal curvature at a point on a regular surface are
called the principal curvatures (k1, and k2). The principal curvatures and principal directions are
obtained from the solutions of:
The Gaussian (K) and Mean (H) curvatures are obtained after expanding the determinant above:
It results that the principal curvatures are given by KHHk −±= 22,1
Figure 3.4-1Brain rendered based on the mean curvature: note that the sulci have low values (black in the
figure) of the mean (Subject A.I.)
0det =−−−−
kGNkFMkFMkEL
21
)21(212
21
2
2
2
kkFEGMLNK
kkFEG
GLFMENH
=−−
=
+=−
+−=
a) b)
Chapter3. Brain segmentation
51
3.5 Results
3.1.6 Whole Brain Segmentation
A software module, CORTESG, has been created and has been used to segment the cerebrum out
of MRI scans from various machines and, for verification purposes, a computational phantom,
with various degrees of noise and inhomogeneity. We use the computational phantom from the
McConnell Brain Imaging Center at the Montreal Neurological Institute, McGill University,
Canada [Collins et al., 1998; www.bic.mni.mcgill.ca/brainweb. The phantom is available with
varying pixel size, levels of noise and bias field for T1, T2 or PD images.
We measured the segmentation accuracy for the T1 phantom with isotropic voxels of 1 mm size.
The noise levels were of 3, 5, and 9% and the inhomogeneity field of 20% or 40% and we
obtained values above 97% in all cases.
Our actual MRI-s are used for visualization purposes in the context of studying the spatio-temporal
pattern of activations for evoked fields (MEG) or potentials (EEG). In these cases the evaluation of
the segmentation result is done qualitatively, by visual inspection of the segmented surface.
The segmentation can be evaluated qualitatively by varying the transparency of the segmented
slice, seen in juxtaposition with the original MR slice. For a quantitative check we have used the
simulated brain phantom from the McConnell Brain Imaging Center with T1 data. The
segmentation accuracy, measured as 100 by the ratio between the correctly identified voxels and
the total number of brain voxels, is 98.6% for 3% noise and 20% inhomogeneity, and 98.0% for
9% noise and 40% inhomogeneity.
The speed performance of the software modules was evaluated mainly on a PC, 1700 MHz AMD
Athlon, 512 MB Ram, under Windows 2000 operating system. The morphology based brain
segmentation from T1 weighted MRIs of 256x256x256 pixels takes approximately 20 minutes for
a trained user. An exception is when anisotropic diffusion filtering is used, since this adds
considerable more time, increasing the time to about 1 hour.
An example of whole brain segmentation is given in Figure 3.5-1A, which shows a slice of the
segmented volume superimposed on the MR scan. The 3D rendering of the segmented brain is
shown in Figure 3.5-1B and C. The transparent rendering of the brain surface allows to asses the
location and shape of selected brain structures or surface parts of the cortex and an example is
shown in Figure 3.5-1B for the hippocampus-amygdala complex. It is sometimes useful to
visualize in the background of the cortex the location of landmark features as the central sulcus
Chapter3. Brain segmentation
52
outline in Figure 3.5-1C. The SAV program allows the delineation of important landmarks on the
3D brain representation, by drawing in 3D (in VISIO as was the case for the central sulcus outline)
or by fast marking the landmarks in subsequent slices (in combination with STRUCTSEG).
Figure 3.5-1. Segmentation examples: a) whole cortex segmentation, visual evaluation. b) a combination of
brain segmentation and subcortical structures reveals information abut the shape and location of the structure,
here the hippocampus; d) the central sulcus as an outline over the cortex. Subject J.R.
SAV has been designed to allow the removal of segmented areas so that parts of the brain which
are normally covered by other structures can be visualized. For example, the cerebellum was
removed for activation studies which required the ventral part of the brain, or the adjacent part of
the brain stem to be exposed.
SAV allows volumetric calculations to be performed. Evaluation of the brain volume gave results
ranging from 815 to 1200, 1500 cm3 (the ventricles are not taken into account).
3.6.1 Gray matter- white matter separation with bias field compensation
We apply the procedure based on FCM for separating the white and gray matter compartments of
the brain to real and phantom MRI data. Figure 3.5-2 shows examples of white-gray matter
segmentation derived from T1 weighted anatomical MRI scans for one of our MRI data sets and
for the Montreal computational phantom.
d c
b a
C
Chapter3. Brain segmentation
53
Figure 3.5-2. Examples of segmented slices for a phantom (a, c) and a real MRI data set (b, d) (subject J.R.)
Examples of 3D segmentations for a real MRI and a set of functional data corresponding to left
arm stimulation of the median nerve at the wrist are shown if Figure 3.5-3. The hand area of SI is
activated. The latency is 31 ms and the threshold used is 25%. For ease of orientation an outline of
the central sulcus is overlaid on the segmented MRI.
Figure 3.5-3. 3D rendering of the gray matter surface and (a) the white matter surface (b). Activation maps
elicited by left arm stimulation of the median nerve at the wrist, 31 ms after stimulation. In pink the outline of
the central sulcus (subject J.R.).
The WM and GM volumes are calculated counting the number of voxels belonging to each class
and multiplying these values by the voxel size, since no gap exists between the slices. The
segmentation is evaluated by comparing the results of the present method on the Montreal brain
phantom. The segmentation accuracy has been evaluated against the computational phantom data
as the ratio between the number of correctly classified voxels and the number of voxels in the
class, multiplied by 100, and we found the accuracy to be in the range of 70%.
3.6 Discussion and Conclusion
Surface based analysis has demonstrated potential for enhancing progress in understanding cortical
structure, function and development in normal and pathological conditions. Accurate
Chapter3. Brain segmentation
54
representations and mapping algorithms are therefore needed for the analysis of high resolution
data.
Depending on the application one or more parameters are prevailing in the evaluation (or choice)
of the segmentation method. Often the evaluation is based on accuracy, efficiency and
repeatability. Accuracy is most important for quantitative measurements of brain structures while
efficiency is most important for visualization purposes. While the efficiency can be easily and
objectively measured the evaluation of accuracy and repeatability are more a subjective matter
since there is a lack of “golden truth”/ or ground truth. Alternatively the manual segmentation can
be used as ground truth or a simulation by means of brain phantoms. The manual segmentation
approach is extremely time consuming due to the large size of the data set and the complexity of
the white matter -gray matter surface, requires a long training period and may have limited
reproducibility. However it is the manual contouring which is often used as the ground truth for
such segmentation. Alternatively one can employ computational phantoms but these phantoms
offer a limited number of conditions and although very important for validation purposes may not
span the whole range of problems one can meet in reality with a particular MR scan.
We have adopted a relatively fast, interactive segmentation approach, which enables to
successfully segment the brain surface from MRI, using mathematical morphology.
The first attempts at segmentation were done in 2D and sometimes required the use of multiple
seeds for region growing. The 3D approach is superior in that it requires only one seed for the
whole volume and therefore less interaction from the user and the connectivity analysis is done in
3D.
Figure 3.6-1. Effect of using multiple seeds for the region growing in the 2D segmentation approaches. The left
figure shows the eroded image in yellow and the thresholded image in blue. The center image is the result of
using one seed for starting the region growing. The use of a second seed was necessary, as shown at the right.
Chapter3. Brain segmentation
55
The same method worked reasonably well for the segmentation of the lateral ventricle since it
constitutes a region with distinct gray values, CSF being much darker than the surrounded brain
tissue. For other structures however a special procedure needed to be developed, and will be
detailed in Chapter 4.
While this method appears suitable for the whole brain segmentation it may not suffice in cases
where the separation of white matter and gray matter is needed.
The GM sheet is a thin (~2mm) and convoluted sheet and therefore its segmentation is likely to be
affected by partial volume effects. Noise and inhomogeneity artifacts add to the complexity of the
MR segmentation problem. This may not be so much import when segmenting the whole
cerebrum and the results from a morphology based segmentation techniques proved to be
satisfactory in general. Further refinements could be done in difficult cases using the editing
possibilities of STRUCTSEG. However for separating the gray matter and white matter
compartments the partial volume effects and the inhomogeneities of the field become important.
We attempt to overcome these using a modified FCM algorithm which incorporates neighborhood
analysis and estimates the bias field. The neighborhood effect achieves regularization, useful for
segmenting noise corrupted images.
The initial choice of the parameters for FCM affects the convergence and the accuracy of the
algorithm. We propose to select these parameters from the analysis of voxels left after eroding the
brain segmented with mathematical morphology. These features add to the functionality of our
surface activation visualization software.
The integration of segmentation and visualization modules within the same software environment,
offers new opportunities, in addition to the direct estimation of the cortical thickness. For example,
the results of tomographic analysis can be either constrained to the identified gray matter domain,
or the bands of maxima in such tomographic solutions can be contrasted to the gray matter
topology.
Chapter 4. Brain Structure Segmentation
56
Chapter 4. Segmentation of Individual Brain Structures
Chapter 4. Segmentation of Individual Brain Structures ........................................................... 56
4.1 Introduction ................................................................................................................ 56
4.1.1 Motivation for segmentation .................................................................................. 56
4.1.2 Selected structures of interest................................................................................. 58
4.1.3 Methods for subcortical structure segmentation .................................................... 63
4.1.4 Background on active contours segmentation........................................................ 64
4.2 Methods...................................................................................................................... 66
4.2.1 Manual and snake based segmentation .................................................................. 67
4.2.2 Hippocampus Segmentation................................................................................... 69
4.2.3 Amygdala Segmentation ........................................................................................ 71
4.2.4 Central sulcus segmentation................................................................................... 73
4.2.5 Thalamus segmentation.......................................................................................... 75
4.2.6 Brain stem segmentation ........................................................................................ 77
4.3 Results ........................................................................................................................ 78
4.4 Discussion and conclusion ......................................................................................... 80
4.1 Introduction
4.1.1 Motivation for segmentation
Brain structure segmentation is essential for quantitative MRI analysis, i.e. studies on the
anatomy of specific brain areas or nuclei and provides information on their normal appearance
as well as on the onset and progress of certain neurological diseases (dementia, schizophrenia,
etc.). Traditional measures refer to cross section, area, length, or volume of a structure. For
bilateral structures an asymmetry index can be computed and this proved useful for the study
of hippocampus in epilepsy [Wang at al., 2001]. Flattening is a computational technique [Engel
et al., 1997 and http://white.stanford.edu/~brian/mri/segmentUnfold.htm, Fischl et al., 1999;
Hurdal et al., 1999; Van Essen and Drury, 1997] which can be used for comparing structures.
Small volumetric changes or regionally specific changes in a tissue compartment can be
assessed by a voxel by voxel comparison of gray matter density [Wright et al., 1995], image
intensity [Andreasen et al., 1994], or probability of a segmented structure [Paus et al., 1996;
Penhune et al., 1996]. Modern approaches address the shape information and statistical
Chapter 4. Brain Structure Segmentation
57
measures. These can be obtained through high dimensional transformations from one brain to
another or to a brain atlas, and may incorporate information on the normal brain variability.
Deformation based morphometry aims to characterize spatial modes of anatomical variability
globally or to identify regionally specific differences in a deformation field which warps one
brain onto another.
The quantitative analysis of MRI data is important in: a) functional brain mapping [Dhawan et
al., 1992], b) assessing structural brain abnormalities [Shenton et al., 1992], c) computer-
assisted neurosurgery [Kikins, 1996] but also d) investigating the relationship between brain
structures and memory processes, emotion and personality [Sullivan et al., 1995; Mori et al.,
1999; Mori et al., 1997] and their evolution with aging.
In functional neuroimaging, exact localization is crucial for the correct interpretation of focal
activations with respect to specific functions [Pruesner et al., 2002]. High precision in
measurement is required due to subtle volumetric differences between brain regions in patients
and in normal control subjects. MRI volumetry is an established research tool in the
investigation of the relations among brain structures to the onset and time course of pathologies
like carcinoma, multiple sclerosis, schizophrenia, dementia, epilepsy, depression or
Alzheimer’s disease. Some of these brain disorders involve volume changes in several brain
regions simultaneously and an ideal study should be able to address all those changes in the
same patient population. Besides volume or area measurements shape analysis is emerging as a
powerful morphometric tool, as exemplified by the work of Shenton and colleagues [Shenton
et al., 2002].
A few applications of quantitative MRI analyses are in:
a) schizophrenia: Some of the robust MR findings regarding abnormal structures include:
enlarged lateral ventricles, reduction of the medial temporal lobe volume (amygdala–
hippocampal complex and/or parahippocampal gyrus), and reduction of gray matter volume of
the superior temporal gyrus [Shenton et al., 2002]. The findings reported in a recent paper
[Lawrie et al., 2002] suggest that even persons with high risk for schizophrenia may exhibit
reductions in temporal lobe volumes.
Shape analysis studies indicate abnormalities of the amygdalo–hippocampal complex in
schizophrenia and an increased left/right asymmetry (R>L) in volume and shape [Wang et al.,
2001]. Shenton and colleagues [Shenton et al., 2002] suggest that some of the structural
abnormalities are neurodevelopmental in origin. They localize brain regions responsible for the
left/right asymmetry differences in patients in the tail of the hippocampus and in portions of the
amygdala.
Chapter 4. Brain Structure Segmentation
58
b) temporal lobe epilepsy (TLE): The temporal lobe structures have been the main subject of
morphological studies in TLE. The mophometric measurements address the hippocampus
[Webb et al., 1999], amygdala or the amygdala–hippocampal complex [Watson et al., 1997]
but also the thalamus [Kim et al., 2002]. Intensity abnormalities, correlated with gliosis, have
been noticed in the MR scans in these areas. The lesions are noticeable because of a subtle
decrease in T1 intensity and an increase in T2 intensities (correlated with increased water
content). The glucose metabolism appears decreased in these areas in TLE [Kim et al., 2002].
c) Alzheimer’s disease Amygdala atrophy appears more prominent in advanced stages of
Alzheimer’s disease [Tsuchiya and Kosaka, 1990]. More intense neuronal loss occurs in the
corticomedial nuclear group than in the basolateral group. Volumetric measurements of the
amygdala and the amygdalohippocampal complex appear more accurate than those of the
hippocampal formation alone in distinguishing patients with Alzheimer disease in the early-
stages [Lehericy et al., 1994; Watson et al., 1997; Lawrie and Abukmeil, 1998].
4.1.2 Selected structures of interest
Of importance in the morphometric studies are the structures of the temporal lobe. In particular
the hippocampus plays an important role in the pathogenesis of temporal lobe seizures [see
Webb et al., 2000 and references therein]. We have concentrated on a number of brain
structures including the hippocampus, amygdala and the hippocampal-amygdalar complex, the
thalamus and the ventricles (the lateral ventricles). We have been also interested in studying
functional aspects where the central sulcus, the postcentral gyrus, the primary visual cortex, the
brain stem, or cerebellum play significant roles.
4.1.1.1 The hippocampus
The hippocampus (HH) is a horseshoe shaped region of the subcortical brain, located in the
medial temporal lobe, inferior to the choroidal fissure and the temporal horn of the lateral
ventricle. It is part of the limbic system, which includes: the cyngulate gyrus, hippocampal
formation, septal nuclei, and amygdala; and also the mammilary nuclei and the anterior
thalamic nucleus, according to Brodal [Brodal, 1992].
Deviations from normal morphology of the hippocampus may be related to temporal lobe
epilepsy, medial temporal sclerosis, schizophrenia, alcoholism or psychosis. For example MR
findings in studies of medial temporal sclerosis (MTS) indicate hippocampal atrophy and
increased signal intensity in T2 weighted images (increased tissue free water, gliosis).
Comparative hippocampal volumetry provides means to quantify these changes. Zeineh and
colleagues [Zeineh et al., 2000], segment and unfold the hippocampus, allowing the
Chapter 4. Brain Structure Segmentation
59
demarcation of the fusiform, parahippocampal, perirhinal, entorhinal, subicular, and CA fields
to be viewed and compared across subjects.
The visualization of electrophysiological data localizing deep bioelectromagnetic sources can
give insight into the role of the hippocampal structure, believed to be involved in the formation
of spatial memory and other kinds of memory [Squire, 1992].
In the sagittal plane the hippocampus is divided into three parts: head, body, and tail. The more
anterior head, called the pes hippocampus, is marked by digitations [Figure 4.1-1a]. The body
is more cylindrical in shape, and the tail tapers posteriorly.
The gray matter of the hippocampus is an extension of the subiculum of the parahippocampal
gyrus. In coronal plane the hippocampus and parahippocampal gyrus form an S-shaped
configuration. The hippocampus itself consists of two interlocking C-shaped structures: the
cornu ammonis and the dentate gyrus. Histologically, the cornu ammonis is further divided into
four sections: CA1 to CA4 [Figure 4.1-1].
The alveus and fimbria, white matter tracts along the superior surface of the hippocampus,
continue posteriorly as the fornix and serve as major efferent pathways to the rest of the brain.
Figure 4.1-1 A diagram of the hippocampus and a coronal cross section (adapted from [Hui et al., 1997]). A.
Medial view of the hippocampus showing the head, body, and tail sitting on the parahippocampal gyrus. 1:
intralimbic gyrus; 2: band of Giacomini; 3: uncus; 4: fimbria (a band of white matter sweeping posteriorly
to form the fornix); 5: margo denticulatus (a row of small bumps in the hippocampal body which
represents the visible portion of the dentate gyrus. Anteriorly, the Margo denticulatus becomes the band of
Giacomini, which then merges with the uncus. Posteriorly the Margo denticulatis becomes the fasciola
cinerea, and the visible continuation of CA3 is known as the fasciolar gyrus); 6: fasciola cinerea; 7:
fasciolar gyrus; 8: gyri of Andreas Retzius (represent the bulge of CA1 at the tail); 9: parahippocampal
gyrus; 10: isthmus; 11: fornix. B. Coronal section of hippocampus. 1: CA1; 2: CA2; 3: CA3; 4: CA4; 5:
dentate gyrus; 6: alveus; 7: choroid plexus; 8: subiculum; 9: fimbria; 10: hippocampal sulcus. Coronal
sections of the hippocampal body show the interlocking relationship between the cornu ammonis and the
dentate gyrus. The cornu ammonis is divided into four segments, CA1 through CA4. The dentate gyrus
A B
Anterior Posterior Medial Lateral
Chapter 4. Brain Structure Segmentation
60
forms the medial U in this configuration. CA4 is located at the hilus of the dentate, whereas CA1 is
continuous with the subiculum. The hippocampal sulcus separates the dentate gyrus from the subiculum.
The alveus envelops the cornu Ammonis and forms the fimbria at the superomedial surface.
4.1.1.2 The amygdala
The amygdala (AG) is an olive shaped mass of gray matter, made of several distinct nuclei,
located in the supero-medial part of the temporal lobe, immediately anterior to the inferior horn
of the lateral ventricle. It is partly anterior and partly superior to the hippocampus and to other
gray matter structures like claustrum and the tail of the caudate.
The amygdala is composed of several distinct nuclei, including three main groups: a) the
basolateral amygdala nuclei (the lateral nucleus, the basal nucleus and the accessory basal
nucleus), b) the centromedial nuclei (the central nucleus and the medial nucleus), and c) the
central amygdaloid nucleus. The last two groups are sometimes referred to as the cortiomedial
nucleus. The amygdala receives input from the neocortex, cingulate cortex and hippocampus.
Major efferent pathways include the ventral amygdalofugal pathway and the stria terminalis,
both innervating the hypothalamus.
Figure 4.1-2 The amygdala is located in the temporal lobe. A: shows the locally zoomed area of interest, the
right temporal lobe. From B to D: the coronal slices advance from posterior-to the most anterior. B: a slice
through the posterior amygdala and its relation to hippocampuss (below) and tail of the caudate (laterally).
C: a slice through the middle part of amygdala. D: shows the anterior-most aspect of the amygdala, at the
level where the optic tract emerges from the cerebri to form the optic chiasm. Numbers in the figure refer
to: 1: Amygdala; 2: Hippocampus; 3: Temporal horn of the lateral ventricle; 4: Posterior aspect of optic
chiasm; 5: Mammillary bodies; 6: Claustrum; 7: Tail of the caudate; 8: Optic tract; 9: Entorhinal cortex;
10: Sulcus semiannularis.(from Convit et al., 1999).
Chapter 4. Brain Structure Segmentation
61
Figure 4.1-2 shows the amygdala positioned directly anterior to the pes hippocampus and
above the tip of the temporal horn.
4.1.1.3 The central sulcus and its neighboring gyri
A major landmark among cortical feature, the central sulcus has been studied from the
anatomical and functional point of view, perhaps as much as the visual cortex.
The central sulcus, or the Rolandic sulcus (abbreviated CS or rol), delimits the boundary
between the sensory and motor cortices, as well as the boundary between the frontal and
parietal lobes. One way to locate it from a surface view is to notice that it is placed between
two parallel sulci, the precentral and the postcentral sulci, it descends from the top medial
surface of the brain laterally to the towards the sylvian sulcus. However the lower part of the
central sulcus never actually intersects the sylvian fissure, but terminates just above it [Figure
4.1-3a].
Figure 4.1-3The major sulci have been used for parcellating the brain. The central sulcus (Rolandic Sulcus)
in red. f1=the superior frontal sulcus, f2=the inferior frontal sulcus, ips= intraparietal sulcus, t1= superior
temporal, cm=callosomarginal sulcus (from [N. Tzourio et al., 1997]).
The precentral gyrus is limited posteriorly by the central sulcus and anteriorly by the precentral
sulcus. It is occupied by the motor cortex of the pyramidal tract. The precentral sulcus is
frequently composed of two parts, having an inferior segment lying more anteriorly, often
intersecting the superior frontal sulcus.
The postcentral gyrus is limited in front by the central sulcus and posteriorly by the
interparietal sulcus. It often intersects the intraparietal sulcus [Figure 4.1-3 a]. It is occupied by
the end station of the sensory tract, the somatosensory cortex.
The identification of the central sulcus often necessitates a combination of all possible cardinal
views and a side view of the brain.
Chapter 4. Brain Structure Segmentation
62
4.1.1.4 Thalamus
The thalamus is considered a central relay station for information throughout the brain. It
processes information from all sensory modalities except olfaction and receives major afferent
connections from the cerebellum [Jones, 1985; Steriade et al., 1997). In addition, the thalamus
has major reciprocal connections with the frontal lobe [Jones, 1985; Fuster, 1997; Steriade et
al., 1997). The thalamus also relays cortical information back to the body to allow for complex
processes such as movement and speech [Jones, 1985; Crosson and Hughes, 1987; O’Leary et
al., 1994; Fuster, 1997; Steriade et al., 1997]. This connectivity has led to the suggestion that
the thalamus may be the site of “gating” or filtering sensory stimuli [Jones, 1985]. More
recently this role has been extended to include global brain integration and pacemaker roles of
the thalamus [Llinas, 2001]
Anatomically the thalamus is a paired gray matter structure, located paramedially and
consisting of a cluster of nuclei, making up about 80% of the diencephalon. Its two large lateral
portions are in general (in about 70% of cases) connected across the midline by the
intermediate mass.
It is placed medial to the internal capsule. Its medial aspect forms part of the lateral wall of the
third ventricle. It is composed of several anatomical and functional groups of nerve cells: some
groups are essential part of central sensory pathways. Some others are part of the
extrapyramidal motor system. Some groups are part of the limbic system or relay for the
association cortex.
Figure 4.1-4 Thalamus outlines in a coronal slice at the middle of the thalamus (from [Portas et al., 1998a])
4.1.1.5 Brain Stem
Located at the junction between the cerebrum and the spinal cord, the brainstem relays
information between the peripheral nerves and spinal cord to the brain and it contains most of
the cranial nerves (all but two of the twelve pairs). It consists of the midbrain, medulla
oblongata, and the pons. The brain stem has a role in: alertness, arousal, breathing, blood
pressure control, digestion, heart rate and other autonomic functions.
Chapter 4. Brain Structure Segmentation
63
The brain stem is made up of three distinct parts, all of them containing ascending and
descending nerve tracts: a) medulla oblongata, a center for several important reflexes (heart
rate, breathing, swallowing, vomiting); b) the pons, a relay between cerebrum and cerebellum;
and host to reflex centers c) the midbrain, a visual reflex center; and part of the auditory
pathway.
The medulla oblongata is approximately 3 cm long, located at the most inferior portion of the
brain stem and is continuous inferiorly with the spinal cord. On its anterior side are two
prominent enlargements, called pyramids because they are broader near the pons and taper
towards the spinal cord, extending the length of the medulla. The pyramids consist of
descending nerve tracts involved in the conscious control of skeletal muscles. An important
feature is that near their inferior ends the descending nerve tracts cross to the opposite side, or
decussate.
Two rounded, oval structures called olives, protrude from the anterior surface of the medulla
oblongata just lateral to the superior margins of the pyramids. The olives consist of nuclei
involved in functions such as balance, coordination and modulation of sound impulses from the
inner ear. The following nuclei of cranial nerves: ix (glossopharyngeal), x (vagus), xi
(accessory), and xii (hypoglossal) are also located in the medulla.
Functionally the medulla oblongata acts as a conduction pathway for both ascending and
descending nerve tracts. Various medullary nuclei also function as centers for several reflexes
(regulation of heart rate, blood vessel diameter, breathing, swallowing, vomiting, coughing,
and sneezing).
4.1.3 Methods for subcortical structure segmentation
Three classes of approaches have been proposed to segment brain structures: manual, fully
automatic or semiautomatic.
While most clinical applications use manual methods, semiautomatic as well as automatic
segmentation methods have been reported in the literature [Duchesne et al., 2002; Joshi et al.,
2002; Dawant et al., 1999; Kelemen et al., 1999; Wang et al., 2002]. In the manual method the
rater identifies and labels the structure of interest in subsequent slices in order to reconstruct it
in 3D. The 3D reconstruction is achieved by stacking contours traced in subsequent slices
[Filippi et al., 1998].
The T1 images are generally used as basis for brain structure segmentation [Pucci et al., 1998],
but also the potential of T2 and PD images is investigated. A recent study [Spinks et al., 2002]
makes use of multispectral images and report that the performance of (thalamus) segmentation
is enhanced through the combined analysis of the three types of images.
Chapter 4. Brain Structure Segmentation
64
The model based methods may use surface tessellation and map the surface model to the given
brain structure surface. The models or templates can be constructed from atlases [Haller et al.,
1997] or from a single subject manual segmentation [Shen et al., 2002] or the mean in a
population. The models are positioned close to the structure’s position in the actual MRI and
allowed to deform, to find the configuration which will minimize an energy functional.
The models incorporate information on the brain structure’s geometry and eventually statistical
shape variation. A particular approach has been introduced by Barra and Boire, [Barra and
Boire, 2001] which uses information fusion and fuzzy maps for segmenting selected structures.
Methods for volumetric measurements include: a) the use manual tracing of boundaries in
subsequent slices and calculating the volume within the structure by pixel counting; b) the use
of tessellation [Arndt et al., 1994]; c) high dimensional brain mapping [Cserrnanasky et al.,
1994; Wang et al., 2001].
To enable the comparison across subjects Laaks and colleagues [Laaks et al., 2000] report the
ratio of the between the volume and the intracranial area, multiplied by 1000; where the
intracranial area is measured in a coronal section, at the level of the anterior commissure.
The results can also be normalized to the intracranial area or volume: for example the mean
hippocampal values can be expressed as a proportion of cerebral hemisphere volume
4.1.4 Background on active contours segmentation
Kass, Witkin and Terzopoulos [Kass et al., 1987] developed the active contour models, also
called snakes. The snake is defined by an energy functional for which the global minima
should be found. The snake is a contour of controlled continuity and can be acted upon by:
internal contour forces, image forces, and external forces which are either supplied by an
interactive user or another, higher process. The applications of active contour models include
contour extraction, motion tracking, image interpretation and shape analysis.
The initial contour is drawn manually and placed near the edges under consideration, and then
image forces draw the contours to the edge in the image. As the algorithm iterates, the energy
terms can be adjusted to obtain a local minimum.
The contour is represented by a vector v(s) = (x(s), y(s)) having the arc length as parameter.
The total energy of the active contour is the integration of local energies along its normalized
contour, which is written as:
dssvEsvEsvEdxsvEE conimagesnakesnake ))(())(())(())((1
0int
1
0
* ++== ∫∫ (Equation 4.1-1)
Chapter 4. Brain Structure Segmentation
65
The energy term being minimized has: an internal energy (due to bending or discontinuities),
an image energy part (given by lines, edges and terminations), and a third energy term, which
incorporates external constraints.
Eint, the internal energy of the contour is given by:
2/))()()()(( 22int svssvsE sss βα += (Equation 4.1-2)
where vs and vss are the first and second derivative of v by the arc length s.
The first order term has large values where there is a gap in the curve (vs) and it makes the
contour behave like a membrane. If α is 0 a discontinuity may occur. The second order term
has large values where the curve bends rapidly (vss) and makes the contour act like a thin plate.
If β is 0 a corner may develop. The values of α and β at a point determine how much the
contour is allowed to stretch or bend at that point. For example a large β determines the
minimum energy contour to occur when the curve is smoother.
Several modifications/enhancements of this original model have been published, including the
incorporation of global shape constraints and its variability, through the use of point
distribution models in their active shapes models [Cootes et al., 1994].
Amini and coworkers [Amini et al., 1988] proposed a dynamic programming approach using
hard constraints, as well as first order and second order continuity constraints (soft constraints).
Their method is however computationally expensive. Williams and Shah [Williams and Shah,
1992] introduce a fast, greedy algorithm which retains the use of hard constraints but it is one
order of magnitude faster than the original one.
The minimum energy contour may be found using techniques like: a) variational calculus
[Kass et al. 1987]; b) dynamic programming: using hard constraints, as well as first order and
second order continuity constraints (soft constraints) [Amini et al., 1988]; c) a greedy
algorithm [Williams and Shah, 1992], allowing a contour with controlled first and second order
continuity to converge on an area of high image energy, the edges.
The performance of the snake algorithm depends on the initial states and given weighting
parameters. If the initial contour is not placed around an object, it tends to settle into a local
minimum near its initial location in state space. Even if placed near the true object boundary,
the snake can be trapped by weak ridges or isolated edge points. The original internal energy
measure of the contour makes the snake to shrink to a point or a line if no external force is
around to attract it. Due to this problem, blurred edge images are used to attract snake points
from distance. However, the true object boundary cannot be found if too much blurred edge
image is used. Hence, Cohen [Cohen and Cohen, 1993] introduced the balloon models to drive
the snake automatically to a good position. An external force is added, that makes the snake to
Chapter 4. Brain Structure Segmentation
66
behave like an expanding balloon. The snake can pass over weak local minima due to its
inflating force. If the weighting parameter for the inflating force is chosen appropriately, the
snake can evolve until it settles on the object boundary. Gunn and Nixon [Gunn and Nixon,
1994] introduced a dual active contour. They used two inter-linked snakes to solve problems in
original snake method. One snake expands from inside the object, the other contracts from the
outside. The two snakes evolve until they meet at the same equilibrium state. If a snake stops at
local minima, the algorithm adds a driving force to the snake to move toward the other snake.
It is possible to segment each slice that contains the object of interest individually and
reconstruct the 3D object from consecutive contours [Cohen and Cohen, 1990; Chung and Ho,
2000]. However, these methods do not guarantee results of the same quality of results as the
3D based approaches.
Deformable surfaces are 3D models directly applicable to volumetric data [Staib and Duncan,
1996; Ghanei at al., 1998]. The initial model for the deformable templates may be generated
from a stack of initial contours, drawn by the user on cross sections of the volumetric data
[Ghanei at al., 1998] or can be derived from statistical atlases [Webb et al., 1999].
The hippocampus can be segmented via a high-dimensional transformation of a segmented
atlas based brain, to the individual brain [Haller et al., 1995], comparing the values of the
voxels that correspond to the hippocampus in the labeled image (initially hand segmented).
Linear [Webb et al., 1999] or nonlinear registration and normalization [Haller et al., 1995;
Collins and Evans, 1997] methods between a particular MRI and a labeled volume derived
from a large number of subjects have been used, as well as registration of two brains, one of
which was previously labeled.
Deformable models seem more appropriate than other segmentation and edge detection
methods due to their ability to treat a structure as a unit object, producing a closed contour.
However the result of using models is dependent on the initial placement of the model and the
parameters of the model, which may need tuning from a trained user.
4.2 Methods
We introduce a software tool, STRUCTSEG for segmenting deep cortical structures or cortical
patches based on stacking contours traced within subsequent slices. Theses contours are either
manually traced or refined from those coarsely traced, by means of an active contours method.
The algorithm we use is based on the fast algorithm proposed by Williams and Shah, [1992].
This method was chosen: a) for its simplicity; b) for its speed relative to dynamic programming
approach; c) for its stability and flexibility regarding the inclusion of hard constraints relative
to the variational calculus approach.
Chapter 4. Brain Structure Segmentation
67
The data consist of 3D digital MRI scans of the head from normal subjects, T1 weighted, 256
x256x256 voxels, of about 1mm size. The software is implemented in IDL language.
The image can be presented in any of the three standard orientations. In general is presented a
zoomed image, three times greater than the original. The zoom factor can be interactively
changed.
The image brightness and contrast can be adjusted according to the parameters of the original
scan images.
The current point in the window where the segmentation is done is also visible in the two
complementary orientations, in separate windows.
4.2.1 Manual and snake based segmentation
In the manual mode the outline of the selected structure is drawn onto an MRI slice (either
coronal, sagittal or horizontal).
The contour from one slice is exported to the next slice on the anterior-posterior direction,
where it serves as initial contour. The procedure described for the previous slice is repeated.
The 2D contours obtained in this way are stacked and used for reconstructing the 3D structure
of the hippocampus. The structure’s volume is calculated after filling in the polygons, counting
the voxels and multiplying by the voxel dimensions.
The brightness and contrast can be adjusted so that the user can better understand the
information in the image. The user can see how the current slice fits in the 3D context of the
rest of the data by examining the "cardinal" views: coronal, sagittal, and axial.
The snake based segmentation is done after applying an edge operator (Sobel for example) and
starts from a manually traced contour. This contour undergoes deformation in order to find the
object/structure of interest. The energy functional to be minimized is:
Equation 4.2-1
The first and the second term are the first and second order continuity constraints and
correspond to the internal energy in the original model.
The last term measures image forces, such as edge strength or intensity.
The algorithm is iterative. During each iteration, a neighborhood of each point is examined and
the point in the neighborhood that has the smallest energy value is chosen as the new location.
Only closed contours are considered.
The continuity term is calculated as the difference between the average distance between points
and the distance between the two points under consideration:
Equation 4.2-2
Chapter 4. Brain Structure Segmentation
68
The points having distances near the average will have the minimum value; the value of this
term is normalized to the largest value in the neighborhood to which the point may move,
giving a value between [0, 1].
The curvature term can be estimated in the case of evenly spaced points using:
Equation 4.2-3
The value is normalized by dividing it by the largest value in the neighborhood, thus ranges
from [0,1].
The image energy is given by the image gradient magnitude and normalized to the maximum
range of the gradient value in the neighborhood: (min-magnitude)/(max-min). The values are
negative, so that points with large gradient will have small values. In order to prevent the
occurrence of large values of this term in areas where the gradient magnitude is nearly uniform,
if (max-min) <5 then min is given the value (max-5).
At the end of each iteration the curvature at each point along the new contour is calculated. If
the value is a curvature maximum, then the β parameter at this location is set to 0. This step
functions as primitive high level process giving feedback to the energy minimization step. The
curvature at these points is computed as:
Equation 4.2-4
Curvature maxima points having values of the curvature above a threshold are considered as
corner points for the next iteration. A further condition for designating a point as corner is that
the gradient magnitude must be above some minimum value, preventing thus the formation of
contours unless the contour is near an edge.
β is set to 0 at the points satisfying the conditions: a) curvature maxima above a curvature
threshold; b) gradient above a threshold.
In summary, the active contours algorithm is applied on MRI slices where an edge detector was
applied. The energy functional is computed for the current location Vi and its neighboring
points. The new position of the point is the position having the smallest energy value. It is
assumed that checking local minima will lead to a global minimum.
Vi-1 has been moved already to its new position, and its location is used, together with that of
Vi to calculate the first order continuity term. Vi+1 is used, together with Vi-1, to compute the
second order constraints. V0 is processed twice, as the first and as the last point in the list
(close contour and obtain a behavior close to that of other points).
The parameters (the size of the neighborhood, point spacing, edge strength, curve strength,
maximum number of iterations, the contour change limit etc.) can be changed for every one
Chapter 4. Brain Structure Segmentation
69
slice. If the result is not satisfactory, for example the contour got trapped by other edges, the
manual segmentation can replace the automatic one.
Finally the contours are stacked one on top of the other and the structure is reconstructed in 3D.
The results of segmentations done on different slice orientations can be combined. If further
editing is necessary, the segmented structure can be examined in the background of the original
scan and the original segmentation can be corrected.
A first examination of the segmentation result can be done using the visualization module,
VISIO.
4.2.2 Hippocampus Segmentation
Most methods for segmenting the hippocampus use coronal images, because both hippocampi
can be assessed simultaneously and the errors in estimating the volume of an anisotropic
structure are minimal if the slices that build up the volume are oriented perpendicularly to the
long axis of the object.
The “coronal slices” in which the hippocampus is studied are defined by various authors as: a)
perpendicular to the intercomissural line (the most usual definition); b) perpendicular to the
axis connecting the genu of corpus callosum to the clava (gracile tuberculus); c) perpendicular
to an axis tilted at 20o negative to the canthomeatal line (unites the corner of the eye to the
external auditory meatus); this direction corresponding to the longitudinal axis of the
hippocampus.
In our volumetric study of the hippocampal formation we include: Cornu Ammonis, dentate
gyrus, and subiculum.
Figure 4.2-1 Hippocampus tracing at the level of the hippocampal: a) head; b) the body and c) the tail.
A:alveus, th: temporal horn of the lateral ventricle, lv: lateral ventricle, the body; cc: the crus cerebri; ac:
ambient cistern; s: splenium of corpus callosum (subject J.R.).
th
a th
lv
cc
ac
s
Chapter 4. Brain Structure Segmentation
70
We use coronal slices for segmenting the hippocampus and proceed along the antero-posterior
direction. We also examine axial and sagittal slices to obtain additional information on the
hippocampal anatomy and its relations to neighboring structures.
The following landmarks are used for segmenting the various parts of the hippocampus,
illustrated in Figure 4.2-1:
1. The hippocampus head: At the level of the hippocampal head, coronal scans show the
amygdala as a close superomedial relation. The two structures are separated by the anterior tip
of the temporal horn, although they are often not easily distinguished. The tip of the lateral
ventricle horn lies just lateral to the hippocampal head. Slightly posterior to this, the temporal
horn may extend medially above the hippocampal head separating it from the amygdala lying
above and medially to it.
The first slice where we segment the hippocampus is where the hippocampus first appears
below amygdala, where is a clear separation between the two structures [Abrahams et al.,
1999]. The hippocampus can be differentiated from the amygdala by visualization of the alveus
and typically, by a region of CSF superior to the alveus [Webb et al., 1999]. The alveus is
visible in images of good quality as a narrow band of white matter separating the rest of the
hippocampal head from the amygdala.
2. The hippocampus body: More posterior coronal sections reveal the transition between the
head and body of the hippocampus, at about the level of the red nucleus. The intralimbic gyrus
serves as a landmark demarcating the hippocampal body from the head. The body is more oval-
shaped and sits squarely on the parahippocampal gyrus. The temporal stem and inferior horn of
the lateral ventricle are used as lateral borders. The alveus and fimbria are used as dorsal
border, but should be excluded. The CSF of the ambient cistern and the crus cerebri are used
for medial border. The white matter of the temporal lobe helps define the inferior border.
3. The hippocampus tail: More posterior, differentiation between cornu Ammonis and dentate
gyrus becomes more difficult as the body flattens further to become the tail. The fimbria loops
upward and anteriorly to form the fornix, which is seen as a band of white matter adjacent to
the inferomedial border of the lateral ventricle and inferolateral to the corpus callosum. The
hippocampal tail loops similarly to form the indusium griseum, a thin strand of gray matter that
lies at the upper surface of the corpus callosum at the midline. Therefore if visible in the MRI
the lower border of the splenium of the corpus callosum can serve as the dorsal border. In more
anterior slices the pulvinar of the thalamus serves as dorsal border.
Chapter 4. Brain Structure Segmentation
71
The CSF of the quadrigeminal cistern helps define the medial border. The lateral border is
defined by the ascending crura of the fornices and the CSF of the atrium of the latereal
ventricle.
The white matter of the temporal lobe helps define the ventral border.
The most posterior position where the hippocampus is to be segmented can be identified as the
slice where the crura of the fornices depart from the lateral wall of the lateral ventricle [Laaks
et al. 2000], or where the lateral ventricles split into the frontal and temporal horns [Webb et
al., 1999].
An example of segmented slice is shown in Figure 4.2-2, comparing the manual and snake
based segmentation.
Figure 4.2-2. Example of hippocampal segmentation: a) manual, b) snake based (subject J.R.).
A combination of coronal with sagittal images for identifying the most anterior slice, where it
connects to the amygdala and coronal slices proved to be the most helpful method.
For the hippocampus, the volumetric studies include: the dentate gyrus, hippocampus proper,
the subicular complex [Laaks et al., 2000] and eventually fimbria.
4.2.3 Amygdala Segmentation
We used coronal images for segmenting the amygdala in combination with sagittal images for
identifying the boundary between the amygdala and hippocampus or with the hippocampal-
amygdala transitional area. This method is particularly helpful for defining the amygdalar
boundaries, especially its posterior end [Figure 4.2-3].
A major problem is the separation of the hippocampal-amygdala transitional area, a thin strip
of gray matter which forms a distinct anatomical structure with distinct connections and
cellular composition. It is located at the posterior end of the amygdala, where the rounded
cortical nucleus of the amygdala transitions to strip of gray matter which connects to the
A B
Chapter 4. Brain Structure Segmentation
72
subiculum of the hippocampus. In the coronal plane it appears at the level of the mammillary
bodies. To isolate it horizontal sections are most suitable.
Figure 4.2-3. The sagittal images are used to identify the anterior end of the hippocampus and the posterior
end of amygdala. The points in saggital slices are visible as stars in the corresponding coronal slices. They
indicate the coronal slice position and help identify the boundary amygdala–hippocampus (subject J.R.).
Several landmarks can be used for defining the boundaries of the amygdala:
a) the most anterior slice can be defined using as landmarks: i) the level of closure of the lateral
sulcus, identified in horizontal sections [Puessner, 2000], ii) the level the temporal stem (a
white matter tract linking the temporal lobe with the rest of the brain) [Shenton 1992], iii) the
section posterior to the one where the optic chiasm appears as a continuous structure.
b) the most posterior slice is defined in the coronal orientation using as landmarks: i) the level
where gray matter first starts to appear superior to alveus and lateral to the hippocampal head
[Puessner, 2000] (the alveus may not be visible because of partial volume effects (PVA) and in
this case the inferior horn of the lateral ventricle is used as border); ii) the mammillary bodies
for delimiting the amygdala from the anterior hippocampus region [Shenton, 1992].
c) the superior border can be defined using the thin layer of white matter dividing the amygdala
from adjacent structure, sometimes though this layer is not visible and the amygdalar gray
matter mixes with that of the putamen. In this case can be used an imaginary horizontal line
between the superior-lateral part of the optic tract and the fundus of the inferior part of the
circular sulcus of the insula.
d) the inferior boundary : the tendorial indentation can serve for separating the amygdala from
the enthorinal cortex, excluding gray matter inferior to the indentation. Sometimes in coronal
slices the hippocampal–amygdalar boundary is subjectively traced as a horizontal line, the
boundary being affected by PVA. In other slices the alveus can serve as a demarcation
landmark.
Chapter 4. Brain Structure Segmentation
73
Horizontal sections can be used for increasing precision in the definition of the medial and
lateral borders, helping to exclude the enthorinal cortex which may have been considered as
part of the amygdala in the coronal sections.
e) the alveus , the ambient cistern or the uncal recess of the inferior horn of the lateral ventricle
can be used as the medial landmark
f) the inferior horn of the lateral ventricle is used as a lateral landmark
Figure 4.2-4 Segmentation of the amygdala, contoured in green. The temporal horn of the lateral ventricle
and the alveus serve as landmarks (subject A.I.).
4.2.4 Central sulcus segmentation
We followed the approach described in [Tzourio et al., 1997] to identify the central sulcus. The
approach is based on combining information from slices of different orientations.
a) Starting from (para) medial saggital slices the central sulcus is identified as the first sulcus
anterior to the callosomarginal sulcus (CM) [Figure 4.1-3 d, Figure 4.2-5b]. On paramedial
saggital slices, at the top of the hemisphere , the CS forms a typical notch , just in front of the
end of the ascending part of CM [Figure 4.2-5 c]. This is the rolandic genus and corresponds
to the hand area.
b) Next the CS is identified in lower horizontal slices as lying between the precentral and
postcentral sulci, the group of these three sulci making a typical pattern at the level of the
vertex. On upper axial slices the CS is characterized by a specific curvature, the rolandic genu,
and never intersects any of the surrounding sulci that run in a different direction, such as the
(superior) frontal sulcus or the intraparietal sulcus. [Figure 4.2-5-c, bottom]
On lateral saggital slices the CS is the third sulcus encountered when starting from the
ascending branch of the sylvian sulcus and moving backward [Figure 4.2-5-c, up].
Once the CS is identified the precentral and postcentral sulci are identified as the sulci which
lie anteriorly ands posteriorly respectively, and run parallel to it.
Chapter 4. Brain Structure Segmentation
74
Figure 4.2-5 Identification of central (rolandic) sulcus (CS or rol) on (a) upper axial slices. (b) a paramedial
sagittal slice; (C) on external parasaggital slices (up) for the inferior part of CS and a horizontal slice at the
level of the hand area (down), CS (Rol) in red, the precentral sulcus in yellow, the postcentral sulcus in
blue, intraparietal in light blue, the superior frontal in green, the calosomarginal sulcus (CM) in orange.
f1=the superior frontal sulcus, f2=the inferior frontal sulcus, ips= intraparietal sulcus, t1= superior
temporal, cm=callosomarginal sulcus. (from N. Tzourio et al., 1997).
Since we were interested mainly in the hand area, the first two methods of identifying the CS
were usually for outlining the CS in MRI sections, the vertices stored being used for creating a
3D volumetric and surface model. A few typical slices are shown in [Figure 4.2-6].
Figure 4.2-6. The CS segmentation starts from medial slices, proceeds laterally and uses horizontal slices
for the lower parts of CS. However sagittal slices may be very useful for the most lateral parts. Most of the
times the combination of the two is used (subject R.B.).
Chapter 4. Brain Structure Segmentation
75
4.2.5 Thalamus segmentation
To identify the thalamus on MRI we used the region of interest (ROI) approach, since it is
highly reliable and respects individual variability [Andreasen et al., 1990; Jernigan et al., 1991;
Gur et al., 1998; Portas al., 1998a,b; and Lawrie et al., 1999].
For segmenting the thalamus we use coronal and sagittal slices and take advantage of the fact
that the thalamus can be identified on the basis of its relationship to the ventricles.
Figure 4.2-7 The tracing of the thalamus in coronal sections is helped by the following landmarks: 3v- the
third ventricle, lv-the lateral ventricle, ic-internal capsule, PU: pulvinar, bs: brain stem, cf-crura of the
fornix. The images shall be counted from left to right and top to bottom. The number of the right most
image in a row is indicated next to it (image modified from Portas et al., 1998a]. For details see text.
The thalamus is first visualized on MRI slightly posterior to the anterior commissure [Spinks,
R. et al., 2002], extending between the foramen of Monro and the posterior commissure,
Chapter 4. Brain Structure Segmentation
76
Visualization terminates just beyond the level of the corpora quadrigemina (the two superior
and two inferior colliculi).
The landmarks which we used are listed below, based on [Portas et al., 1998a]:
1. The most anterior slice is defined using as landmark the mammillary bodies of the
hypothalamus or after the slice containing the clearest view of the anterior commisure [Spinks
et al., 2002]. The internal capsule also will appear thicker than in the previous section.
2. The most posterior slice is defined as the slice where the thalamus merges under the crura of
the fornices.
3. The most inferior slices is where thalamus merges with the brain stem. The zona incerta and
its junction with the internal capsule serve as the inferior border of the thalamus, thus
excluding the subthalamic nucleus, the substantia nigra and the nucleus rubor.
4. The superior margins are defined using the lateral ventricle as a landmark. As one moves
from anterior to posterior, the thalamus loses its magnitude. The coronal height of the thalamus
decreases, thus allowing the fornix to serve as the superior boundary in more posterior slices.
Medially the boundary is defined by the third ventricle.
The two separated volumes, one obtained from coronal and the other from sagittal slices are
combined in the end.
An illustration of the shape variability form the most anterior to the most posterior coronal
slices in a normal subject us shown in Figure 4.2-7, [based on Portas et al., 1998a]. Several
features and landmarks useful for segmenting the thalamus are described below.
1. The thalamic reticular nucleus is noticeable first. For the most anterior slices the boundaries
are: the internal capsule (ic), (laterally), the main body of the lateral ventricle (lv) (dorsally),
the third ventricle (3v) (medially).
2. The VA (ventralis anterior) nucleus is just dorsal to the hypothalamus (h).
3. The mammillary bodies (mb) of the hypothalamus are visible.
4. The lateral edges of the thalamus are affected by partial volume effects
5. The mammilothalamic tracts can be viewed, and as marked in 6 the interpeduncular fossa.
7. Globus pallidus (gp) is marked for a reference
8. Note the extension of the interpedunculare fossa in slice 8 and next slices
9. The medial lemnisucs (ml), the putamen (p), zona incerta (zi) as the inferior border, the
subthalamic nucleus, and the crus cerebri are visible.
10. The brain stem (bs) is evident, substantia nigra (sn)
Chapter 4. Brain Structure Segmentation
77
11. The intherhalamic adhesion (the intermediary mass) of the thalamus is visible. The
dorsomedial nucleus appears. White matter can be seen separating the main body of the
thalamus from the lateral geniculate nucleus (lgn).
13. To delineate the brain stem from thalamus (when the habenula is not visible) it is necessary
to draw a line from the hypothalamis sulci of the third ventricle to the deepest indentation of
CSF, ventrally and medially to the medial geniculate nuclei (mgn).
14. Partial volume effects make the delineation of the lateral thalamic boundary difficult.
15. The lateral limit is marked by the internal capsule. To delineate the brain stem from
thalamus is necessary to draw a line dorsally from habenula to the deepest indentation of CSF,
ventrally and medially to mgn.
16. The cerebral aqueduct (ca) is evident and the lgn and mgn appear as ventral bumps on the
pulvinar nucleus (pu).
17. The brain stem is clearly delineated from the pu. CSF serves to identify the central border.
18. The pulvinar is ball shaped. The fourth ventricle is visible and perhaps the pineal gland.
20. The boundaries are: the lateral ventricle (dorsally), the cistern of the great cerebral vein
(laterally), temporal stem (medially) and CSF. Note also the superior colliculi (sc)
21. A small portion of the pulvinar appears between the crus fornices (cf) and the CSF of the
great cerebral vein (superior cistern).
4.2.6 Brain stem segmentation
The brain stem is relatively easy to identify in horizontal slices but some user interaction is
necessary to cut the connections to the cerebellum through the cerebellar peduncles and to
establish the superior limit with the diencephalon. For the definition of these boundaries see the
section on thalamus segmentation.
Figure 4.2-8. Tracing the brainstem at the level of: a) medulla; b) pons; c) middbrain. Cp: cerebellar
peduncles (subject J.R.).
iv ventricle cp
cp
Chapter 4. Brain Structure Segmentation
78
4.3 Results
Methods for segmentation, visualization and volumetric analysis of brain structures from MRI
scans are proposed in this chapter, and implemented into the software called STRUCTSEG.
We have applied it to various subcortical structures and also to the brain stem, cerebellum,
cortical areas like the pre and postcentral gyri, the central sulcus, and the primary visual cortex.
A number of structures were segmented and used primarily for activation visualization
purposes (brain stem, central sulcus, pre and post central gyri, amygdala and thalamus).
Preliminary volumetric studies were conduced for the hippocampus and amygdala and our
results are comparable with results reported in the literature.
The segmentation accuracy evaluation was done by comparing the volumes of the
corresponding structures with those reported in other studies and qualitatively by inspecting the
segmented brain structures in the background of the original MRI.
The volumetric measurements use voxel counting, and the number of voxels is multiplied by
the voxel size in case there is no gap between the slices. Examples of volumes of segmented
structures and the percentage they represent out of the brain volume, are shown in Table 4.3.1
for one subject (J.R.).
Table 4.3-1Segmented brain structures volumes and the percentage they represent out of the brain volume
(subject J.R.)
Examples of segmented structures are shown in Figure 4.3.1, the right most column shows the
individual structures in relation with the cerebrum.
Structure (both left and right
hemisphere)
Volume (cc) Area
(cm2)
Volume/Brain
Volume (%)
Brain 1423.00 1345.11 100.00
Cerebellum 157. 88 350.94 11.09
Brain Stem 36.25 86.50 2.55
Hippocampus 7.99 39.30 0.56
Amygdala 4.41 20.49 0.31
Hipopcampus-Amygdala Complex 13.71 51.87 0.96
Thalamus 15.78 56.59 1.11
Lateral Ventricle 28.73 92.34 2.02
Central Sulcus 10.08 79.96 0.71
Pre and Post central gyri (GM) 70.87 21.36 4.98
Chapter 4. Brain Structure Segmentation
79
Hippocampal-amygdala
Amygdala
Figure 4.3-1.Selected segmented structures presented independently and in relation to the cortex (right
most column). (Subjects J.R., T (brain stem), V.P. (calcarine sulcus walls), R.B. (lateral ventricle). The right
column shows the individual structures in the background of the whole cortex.
Cerebellum Brain, no cerebellum
brain stem (BS) BS, medial lemniscus
Hippocampus Thalamus
Pre and postcentral gyri Central sulcus
Lateral ventricle Calcarine walls
Chapter 4. Brain Structure Segmentation
80
We segmented the amygdalas for 3 subjects and the results are shown in Table 4.3-2.
Table 4.3-2. Amygdalar volumes
The hippocampus volumes for subject A.I. are 3.3cm3 (LH) and 3.5cm3 (RH). For subject JR
the hippocampus volumes are 4.3 cm3 (LH) and 3.6cm3 (RH).
The main use of our segmentation results, pertaining to the: central sulci, gyri, thalamus and
brain stem, is for mapping functional data from MEG and EEG. Another use of the
STRUCTSEG program is for editing preexisting segmentations of the brain where undesired
connections like those to the eyes were still present.
The speed performance of the STRUCTSEG software module was evaluated mainly on a PC,
1700 MHz AMD Athlon, 512 MB Ram, under Windows 2000 operating system. The duration
of the segmentation process for brain structures (i.e. hippocampus, amygdala etc.) depends on
the complexity of the structure and its extent. The larger the number of slices where the
structure is present, the longer the process. Typically, the segmentation process does not take
more than 1 hour. An exception is the central sulcus, which is thin and sometimes interrupted
and may require editing and the combination of segmentations done on different slice
orientations. The snake based segmentation adds a few seconds for each slice (approximately
2-5s).
4.4 Discussion and conclusion
The subcortical structures may have relatively low contrast and multiple as well as
discontinuous edges in MRI. This characteristics make challenging an automatic segmentation.
The application of active contours to the problem of segmenting selected brain structures and
nuclei is investigated. In some slices the snake “bleeds” and finds other edges. In these cases
the manual segmentation is preferred. Because the MRI scans have different image quality, the
interactive tools for changing the zoom factor, contrast and brightness have been found helpful.
In the case of hippocampus segmentation a problematic area is the hippocampal–amygdala
transitional area. The amygdala, located at the anterior border of the hippocampus, presents
challenges because of its small dimensions and the nondistinct edges (especially in the
hipocampal-amygdala transitional areas) but also because parts of the basal ganglia from
Subject Right Amygdala (cm3) Left Amygdala (cm3)
TR 2.08 2.71
TK 1.93 2.45
JR 2.81 2.92
Chapter 4. Brain Structure Segmentation
81
superior and enthorinal cortex from inferior blend in [Pruessnes, 2000]. Similar problems
appear for the thalamus, where is it difficult to find “hard” boundaries in axial or sagittal
sections, but the combination of the two orientations can be useful. The brain stem and
cerebellum appear easier to segment, however user intervention is required to separate these
structures at the cerebellar peduncles. The good definition of the lateral ventricle’s boundaries
makes it a good candidate for semiautomatic segmentation, using the same procedure as for the
whole brain.
The volumetric measurements are derived from segmentation of specific brain structures based
on one or at most three subjects for the same structure. Although insufficient for statistics our
measurements compare well in general with other data reported in the literature. We measured
the hippocampus volume and this ranges from 3.3-4.3 cm3 but we cannot yet make a statement
about differences between the right and left hippocampus due to the reduced number of
segmentations. The ratios between the hippocampal and the brain volume are close to 0.5, as
reported elsewhere [Webb et al., 1999].
Most authors report absolute measurements for the volume of the hippocampi, which range
from 2.5-3.5 cm3, with lower limits for the normal hippocampus 1.7cm3 (LH), 2.0cm3 (RH).
Pruessnerr and colleagues [Pruesnerr et al., 2000] manually segment the hippocampus and
amygdala and report larger volumes of the right hippocampus, 3.324 versus 3.208 cm3, but no
interhemispehric differences for the amygdala (1.154 versus 1.160 cm3). The measured
amygdala volumes usually overestimate the true size according to postmortem studies [Convit
et al., 1999], the reported values range from 1.3cm3 [Rossi et al., 1994] to 3.4 cm3 [Watson et
al., 1997]. Convit and colleagues report values of 1.05 cm3 for the right, 1.14cm3 for the left
amygdala. Our result are in the range 1.9-2.9cm3.
Spinks and colleagues [Spinks et al., 2002] used manual tracing in multispectral images and an
artificial neural network and estimate the volume for the left (7.05+/-1.12 cm3) and right
thalamus (6.55+/-0.94 cm3). The bilateral thalamic volume for one subject is 16.71 cm3,
unnormalized.
While the structure’s shape can be retrieved from 2D contours, the use of a predefined model
may speed up the procedure and be beneficial for cases where a large number of subjects are to
be analyzed. The segmentations we provide can serve as initial models but may need to be
corrected for the individual brains.
In general a certain amount of knowledge is needed to separate a structure from its
environment. Our manual method is helped by the simultaneous visualization of the three
cardinal planes and the use of various image operators, like contrasting and zooming. While
Chapter 4. Brain Structure Segmentation
82
tedious the manual segmentation is considered the gold standard against which automated
segmentation methods are to be evaluated.
Our semiautomatic method combines the advantage of human expertise with the power of
active contours in detecting and tracking a structure’s edges, aiming to reduce the time
necessary to accurately and reproducibly segment the selected structures.
A more advanced step would be to incorporate the knowledge of an expert anatomist into
models used to segment the structures from individual MR scans, eventually registered into the
same (Talairach) space.
There are still discrepancies across laboratories which perform segmentation of the same
structures. These are due to the use of different: a) definitions of boundaries b) segmentation
protocols and data parameters (slice thickness, orientation) but also c) methods for volume
quantification. The most widely used seems to be the manual tracing of boundaries in
subsequent slices and calculating the volume within the structure. Arndt [Arndt et al., 1994]
investigated the use of surface tessellation for estimating the volume and compared it to the
simple pixel counting technique which proved to be more robust. Others [Cserrnanasky et al.,
1994; Wang et al., 2001] used high dimensional brain mapping.
A natural follow up of this study would be the extension of the method to 3D. We plan to use a
segmented structure as an initial model fed to the active contours algorithm. The model would
undergo deformation to find the true boundary of the structure in a second brain, after
registration and normalization of the two brains.
The normal brain is subject to a large variability and the shape characteristics of its structures
may be quantified and included into a statistical atlas as reference. The methods currently used
for analyzing those differences involve laborious manual tracing of the contours of anatomical
structures derived from MRI scans. Repeatability, inter rater variability and lengthy times
required are problems characteristic to the manual segmentation. Automated procedures, like
the one reported by Spinks and colleagues [Spinks et al., 2002], are very promising and would
increase both the number of regions as well as the number of individuals who could be
investigated in any one study. The normal variability of structures, the multiple edges, the fact
that neuroanatomy experts make decisions based on external landmarks and the need for a
close initialization of deformable models are challenges for the automatic methods. The
possible errors involved in the automatic segmentation of subcortical structures are important
for the small dimensions of these structures. A semi-supervised approach like the one we
adopted is simple and yet accurate since it allows the control of the results of segmentation.
Chapter 5 Visualization of surface activation
83
Chapter 5. Visualization of surface activation
Chapter 5. Visualization of surface activation .................................................................................83
5.1 Introduction.........................................................................................................................83
5.2 Methods ..............................................................................................................................85
5.2.1 Extracting the structural information ...........................................................................86
5.2.2 Computing the activation maps....................................................................................87
5.2.3 3 D Visualization..........................................................................................................88
5.2.4 Slice views....................................................................................................................90
5.2.5 The VISIO software features........................................................................................90
5.3 Results.................................................................................................................................92
5.3.1 Qualitative Evaluation of Segmentation ......................................................................93
5.3.2 Surface activation visualization ...................................................................................94
5.4 Discussion...........................................................................................................................99
5.1 Introduction
Visualization has played an important role in mathematics. While some people can see in their
mind’s eye the beauty and structure in mathematical or statistical relationships, most require a
visual representation to appreciate these dependencies. It is no wonder then that it is said: “the
most profound use of computers in mapping is in the field of visualization” [Toga, 1996].
Imaging the electrical activity of the brain is becoming one of the major challenges in functional
neuroimaging. The scope for visualizing the brain activation is to extract the meaningful
information from complex data sets, display, summarize and interact with these data. In the case of
MEG and EEG this information refers to changes in the brain activity, eventually linked to a
particular task.
MEG and EEG are the unique noninvasive modalities which provide a measure of brain activity at
the time scale of neuronal currents. The goal is to represent the functional information in the
background of the anatomy so as the reveal the spatiotemporal evolution of the activations and the
interplay between distinct activated areas. The efficient visualization of large MEG data sets in
Chapter 5 Visualization of surface activation
84
combination with the anatomy is a challenging task partly because of the computational cost of
processing or simply manipulating the data, and the apparently low signal to noise ratio.
The multidimensional nature of brain data lends itself to a variety of visualization techniques
concerned with modeling, manipulation and display.
The visualization can be done using 2D (slices) or 3D (volume) representations. The classical
representation for visualization uses slices extracted from MR scans but may fail to capture the
true 3D nature of cortical paths and the relationship to the orientation of the local anatomy. The 3D
based approach for visualizing activation in the background of the anatomy presents the advantage
that it completes the picture with information missing from the slice representation. This
representation allows to better examine, in a single view, the timing of distinct sources and the
relationship between these sources to the topography of the anatomical structures.
Most of the existing software for brain activity visualization are 3D based (BrainVoyager,
MRI3dx, Freesurfer, Brainnstorm, etc). Furthermore these 3D surfaces can be viewed in a single
plane after flattening [Fishl et al, 1999 a and b; Hurdal 1999; Van Essen et al, 1998; Wandell et al,
2000]. The fully or partially flattened representations have the advantage that they expose
activated areas buried within sulci. However, the flattening methods introduce metric distortions in
the surface representation [Fischl et al, 1999b] while they hide anatomical relationships of
potential importance when studying the interactions of brain areas.
The structure being visualized can be the whole cortex, the white matter only, selected parts of the
cortex, or specific subcortical structures. New imaging technologies allow the visualization of
white matter tracts. The structural information provided in this way is combined with functional
data.
The functional data support different types of representation but they rely in our case on the
original raw data, as recorded from the EEG or MEG sensors. However the EEG or MEG raw
signals cannot be directly attributed to underlying cortical regions [Gross et al 2001; Dimitrov
1998]. Either the statistically processed data [Dimitrov, 1998] or the coherence measures [Gross et
al, 2001] are usually mapped on the brain. The explanation lies in the complex relationship
between a signal source and the recorded signal and this is obtained solving the forward
electromagnetic problem. Especially for EEG the potentials are smeared because of the
inhomogenous conductivity of the head. The activity from even small cortical patches is therefore
likely to be recorded by several sensors. The sources of signals can be modeled as single or
multiple dipoles, multipoles [Mosher et al. 1999] or continuous [Malmivuo, 1995; Hämäleinen,
Chapter 5 Visualization of surface activation
85
1993]. These sources can be remapped on the cortex to the vertices, elementary patches or
collections of patches [Mosher et al 1999; Kincses et al, 1999; Kincses et al 2000] on the
tessellated brain structure’s surface. The intensity or magnitude of a response is conveyed by the
value of a voxel or patch, associated to the physiological measurement and pseudocolored to
enhance differences. The original structural data are thus texture mapped with values
corresponding to the activity.
A combination of these data representations can be used. For example dipoles, or maximum
current density vectors for a region of interest [Ioannides et al, 2002 a and b] can be represented in
the background of the activity maps.
The volume representation provides the most data about the brain and the resolution can be
manipulated to use fewer voxels for less demanding applications but in general are more
computationally expensive than surface representation. The surface representation provides no
information on the inside of the brain while explicitly defining the geometry of the structure
exterior.
From the digital representation of a brain structure one can infer morphometric measurements such
as area, volume and shape related information. These calculations are more efficient for a surface
than for the volume representation of the same brain structure, because of the smaller number of
voxels involved. Furthermore statistical analyses can be done to compare and correlate selected
brain aspects from different modalities or from different subjects. The results can be visualized as
individual segmented structures or in relationship to the whole brain anatomy, which can be
rendered as a transparent surface.
In this thesis we choose as visualization method, a 3D surface representation of the anatomy,
pseudocolored in accordance with the activity values. We developed a software program for
visualization of activation data (VISIO), imported from source analysis programs like MFT or
BESA. This helps reveal the relationship between surface topography and function and helps the
recognition of possible interacting cortical areas.
5.2 Methods
We introduce VISIO, the third module in the software suite called SAV (Surface Activation
Visualization). Like the first two modules VISIO is implemented in IDL and follows an object
oriented approach in the software design. While the first two modules are concerned with the
processing of anatomical information in order to make it suitable for activation studies, this last
Chapter 5 Visualization of surface activation
86
module is concerned with mapping the activation data on the brain structure, and the efficient
visualization of these combined data.
5.2.1 Extracting the structural information
The MRI data are suitable for our visualization only after performing several image processing
operations. In our case, the MRI data consist in T1 weighted scans, 256x256x256 almost isotropic
voxels. The scans are transformed to 8 bit/pixel. The background noise is reduced through the use
of low pass filtering. In the end only the selected segmented structure is retained from the MR
scan. The segmentation process is described in chapters 3 for the whole cortex and 4 for selected
brain structures.
In the segmentation modules the visualization pipeline starts from the tomographic volume which
is preprocessed with various operators in order to enhance the observation of meaningful
information (low pass filters, contrast, brightness, and threshold. The resulting 2D contours are
stacked in STRUCTSEG, while in CORTSEG a grey level volume is reconstructed directly. This
volume may contain other attributes, in addition to the grey level value, like the probability to
belong to a tissue type after classification into the grey matter, white matter or background. The
volume visualization pipeline ends with the surface based rendering.
The extracted volume is retained and used in slice representation and the boundary with
background voxels is extracted as an isosurface, tessellated and characterized in terms of vertices,
polygons and local normals.
If the resulting surface is noisy, a laplacian smoothing can be applied to each vertex on the
triangulated mesh.
)( )(00
)()()1( ni
M
jnjnini xx
Mxx −+= ∑
=+
λ
xi(n) is vertex i for iteration n
λ is the smoothing factor
M is the number of vertices that share a common edge with xi(n).
Another option is to examine a less complex surface, by decimating the mesh while reserving as
much of the original anatomical information.
These data are the basis for visualization of functional data, as described in the following section.
Chapter 5 Visualization of surface activation
87
5.2.2 Computing the activation maps
The functional data are read from files which contain the result of MFT analysis [Ioannides, 1990]
on MEG data. The MFT solutions are computed within the source space covering the whole brain,
brainstem and cerebellum in the nodes of a lattice. The activity values to be displayed are
interpolated from the MFT solutions, at locations on the surface of the structure of interest, such as
the cortical mantle, brain stem, thalamus, etc. The surface data are of interest for us and therefore
only the grid points in the neighborhood of the extracted surface are considered. The neighborhood
size can be adjusted by the SAV user. In this way, the cortical current images are created,
representing the current density on the cortical surface at a single time point.
The user has the possibility to choose the neighborhood size, related to the depth measured from
surface (the cortex for ex.) where from the solutions can still be projected onto the surface. Once
the neighbors of a point on the cortex are known, the activation value at that point is calculated
using Sheppard’s method for the given neighborhood, as the weighted sum:
)()()(1
i
N
ii xfxwxs ∑
=
= ; where p
N
j
pji
ii
xx
xxw
−
=
−
∑ −
−=
1
Where N is the number of source points in a vicinity defined by the maximum distance to the point
x on the cortex. The weights are nonnegative and sum to 1 and the exponent (the p factor) used is
either 1 or 2. Additionally the condition that the interpolated values are equal to the “observed”
values in the nodes of the lattice must be satisfied.
The vector as well as the scalar properties of the activation data can be simultaneously displayed
by showing the contour plot of the power or current density modulus and the current density vector
map.
For the actual display, a part of the color scale is reserved for grey scale representation of the
anatomy. Information on the activation modulus is shown using color coding, for example in the
red-yellow color scale, while information on the current flow direction relative to the local
normals is conveyed by associating outgoing currents with colors in the red-yellow range and
incoming currents with colors in the blue-green range.
There are several possibilities to visualize with VISIO the distributed functional data.
The display may consist in contour plots of activation maps which represent:
a) the moduli of the cortical current distributions, usually represented in a color scale from yellow
to red, the red color being assigned to the highest values;
Chapter 5 Visualization of surface activation
88
b) the signed values of the moduli of the cortical current distributions, usually represented in a
color scale from blue to red. The red color is assigned to the highest positive values and the blue
colors to the highest negative values. The small absolute values are in this case yellow if positive
or green if negative;
b) the statistical parameter maps derived after t-tests or Kolmogorov Smirnoff tests. In fact the
modified p values derived from these test are represented and they include the signed value from
the test, these being multiplied by (1-p). In this way the most significant values (for small p
values) are shown in red or blue.
Another way to present information on the current flow direction is to represent the current density
vectors or dipoles as a vector map on the structure surface. The otherwise dense mesh of vectors
(resolution 1mm) can be trimmed so that the user can make sense of the general behavior. The
area of maximal activity can be indicated by one vector only. The outgoing vectors are assigned
one color, while the incoming vectors are assigned a different color. The display of equivalent
current dipoles uses the same conventions and spheres replace the traditional arrow symbols.
In SAV, two visualization methods are possible: surface and slice representation. On each of these
representations are shown dipoles, vector maps or surface activation maps. Each of these
approaches is described in detail in the following subsections.
5.2.3 3 D Visualization
Our main goal is to visualize the activations on the surface of specific segmented structures,
ranging from the entire brain to selected subcortical structures or patches of cortex. The
segmentation program allows the user to define interfaces as (thin) surfaces or anatomically well-
defined parts, gyri or sulci, of the cortex like the somatosensory or visual cortex and display them
independently or in the background of the rest of the brain, or together with other structures.
SAV uses a surface representation of the anatomy data, the extracted isosurface based on the
segmentation performed with either CORTSEG or STRUCTSEG modules. One or more surface
objects can be visualized simultaneously by mapping a semitransparent texture image on one
surface.
The activation data are rendered onto the anatomy, which is so viewed as a pseudo colored
surface. Different types of rendering and shading are possible: from a (colored) mesh object to a
surface which can use flat/constant shading or interpolated/Gouraud shading.
Chapter 5 Visualization of surface activation
89
The choice of the color scale used to represent the activation values is left at the latitude of the
user who can simply choose from a predefined list. We preferred to use a color scale with “warm”
colors (from yellow to red) for outgoing currents or and for the statistical parameters indicating
hyperactivation and a “cold” color scale for “incoming” currents or for the statistical parameters
indicating hypoactivation. The color coding mentioned above is our usual choice but one may
choose other color scales with a smaller number of colors for example, which may be able to
reflect the changes in activity values from one region the neighboring one in a more striking
manner.
The color bar which accompanies the display indicates the range of activity values (min, max
values) for each time slice, in accordance to the time /spatial normalization options.
SAV allows the user to customize the viewing parameters interactively. Thus, the user can set the
object orientation, zoom factor, and position within the draw area. Three sliders are controlling the
rotation of the 3D model around each of the X, Y and Z axes, corresponding to left-right (LR),
anterior-posterior (AP), superior-inferior (SI) directions. Alternatively the object can be
interactively rotated by clicking and dragging with the left mouse button on the drawing area.
Outlines of sulci or structures of interest can be overlaid on the image to help a more precise
orientation relative to the specific anatomical landmarks (Figure 5.3-1, D). The outlines can be
traced in slices using the STRUCTSEG module or on the rendered surface using the VISIO
module.
All objects on the display area (anatomy, axes, wire frame box, color bar, etc) can be shown or
hidden. The displayed objects can be translated and rotated in order to better expose the areas of
interest. Also they can be made transparent or opaque to allow visualization of structures lying
inside or behind them as well as dipoles locations.
It is possible to select zoom factors and predefined viewing transforms: top view, left side view,
right side view, etc. User defined viewing transformations can be created and saved for later use,
to enable the observation of the activation for a different condition or latency from exactly the
same perspective and angle. Labels can be introduced in the images in positions defined by the
user to ad extra information on the experiment or anatomy.
The latency is displayed for each time slice analyzed to track time in successive data and
animations.
Chapter 5 Visualization of surface activation
90
5.2.4 Slice views
Brain slices can be visualized in one of the three orientations: coronal, sagittal or horizontal,
allowing quick browsing through the data for one time slice and inspection of activity buried deep
within the sulci. The data may be presented as clusters of 16 slices, each of which can be
individually zoomed. Alternatively one may prefer quick interactive browsing through the data by
moving the cut planes to the desired location [Figure5.3-4].
5.2.5 The VISIO software features
Surface representation and slice views
Current density maps are the main way to examine the time course of activation. The activation
values are interpolated in the nodes of the mesh which represent the selected brain structure. This
surface is pseuocolored according to the activation values. The activation data can also be
examined in slice cuts at any of the three stereotaxic planes through the brain volume. The
parameters of the analysis and the file name containing the activation data are indicated on the
display and can be saved for “offline” analysis.
Statistical data display
The same kind of contour maps used to illustrate the current density distributions are used to
illustrate differences between activations at two distinct latencies or between conditions. The
results of statistical analysis using the t –test and the Kolmogorov-Smirnov (KS) test can be
visualized in this way. The t or KS maps are pseudo-colored in accordance to the p factor in order
to identify areas where the activity is statistically different from the base line activity of the same
area or from the activity of the area during a specific condition. The color map uses red or blue for
the most significant changes. The same threshold buttons used for the classical contour maps or for
the vector maps are used for the p value displays
Animations
SAV makes possible the dynamic analysis of activation. Succesisve images of the type described
in the above paragraph can be assembled into animations, enabling the eye to peruse in its usual
function the temporal dimension of the data. The start and the end time for the animation analysis,
the normalization type and other display parameters are defined by the user and can be changed for
each analysis. Alternatively the parameters can be provided in a text file and the program may
Chapter 5 Visualization of surface activation
91
analyze automatically multiple slices with the same parameters and assemble the resulting images
in an html page. An example can be seen at the website
http://nucleus.med.upatras.gr/~alex/sav_results.html.
Reducing surface complexity
The large data sets describing the anatomy, combined with even larger data sets describing the
activations time course slow down the analysis and therefore any reduction in computing time is
important. For this reason one may choose at times to work on a simple mesh, obtained from
decimating the original one. The original mesh can be decimated and/or smoothed, in an attempt to
reduce its complexity while preserving as much detail as possible where this is important.
Thresholds, normalization, and regions of interest
The subtle differences in magnitude values may not be well noticeable when observing the whole
brain. We make use of thresholds to emphasize regions of high activity; and we use separate
thresholds for the modulus (displayed by the contour maps) of incoming and outgoing vectors. The
same thresholds are used for the vector maps. Additionally the vector maps are “trimmed” to a
user-defined percentage of the total.
The analysis can be done one time slice at a time to emphasize local maxima but it is most useful
to compare the time course of activations. For the analysis of succesive time slices the initial time
slice and the end time slice have to be specified by the user. The results will be normalized to the
minimum and maximum values from a given time moment, time range or for the whole time range
in the hyper file. The global normalization can be used for analyzing a specific time range, but the
normalization is done to the global minimum and maximum. The normalization can be done also
in space, to the range of values pertaining to a region of interest (an anatomical structure), or to the
entire source space. This is particularly useful since some subcortical activations may sustain
electrical fields which are orders of magnitude smaller than the cortical ones.
Activation curves
Activation curves can be produced to analyze the time course of activity for a selected voxel on
the brain surface. The voxel of interest is selected by mouse clicking on the 3D representation of
the segmented cortical surface. The activity then is computed using nearest neighbor interpolation.
Chapter 5 Visualization of surface activation
92
Output
Several types of output files can be generated from the VISIO program and these are listed below:
a. Static
The snapshot of the display window can be saved for further analysis as JPEG or GIF files. The
names of these files are automatically generated and comprise information on the functional data
like: time, view (anterior, posterior, etc), latency, threshold, normalization and weighting options,
or zoom factor.
b. 3D scene content -VRML
To add the third dimension of the data, the content of the scene can be saved as VRLM and viewed
later from within a web browser.
c. Dynamic
The dynamic analysis of activation data is done by creating animations from images generated for
successive time slices. The movie is presented online, while the activations are calculated. These
activation movies can be saved for “offline” analysis in MPEG format and played using any
MPEG player. For an easy interpretation of the images the time latency in ms is written onto each
frame of the movie.
d. Suitable for web
The results of automated analysis with a predefined set of parameters is presented to the user as an
HTML page, each image being accompanied by its self explanatory title, describing the
parameters used for analysis and the time moment or condition.
e. Exporting data
Data relative to the anatomy: vertices and local normals can be exported as text files for use
outside the SAV program.
5.3 Results
The results of this chapter are the different visualization options which are available within SAV
and in particular the VISIO module.
The speed performance was evaluated mainly on a PC, 1700 MHz AMD Athlon, 512 MB Ram,
under Windows 2000 operating system. The computing times are related to the number of vertices
Chapter 5 Visualization of surface activation
93
and triangles characterizing the anatomical surface on one hand and to the resolution of the grid
used by MFT and the number of time slices on the other hand.
The mesh which represents the surface of the segmented structure can be smoothed or decimated
to reduce its complexity. The duration of the decimation process depends on the geometry and size
of the structure of interest, and on how much the surface is to be decimated. In general, this is a
computationally expensive step for complex structures like the whole cortex. However for the
hippocampal- amygdala complex made of 10144 vertices, the decimation to 20% of the number of
vertices takes about 1.2s, and the resulting surface is made of 2011 vertices only.
In SAV, specifically in the VISIO module, the anatomical and functional data as well as various
operations applied to these constitute an object. We will refer to this as an active object. The
creation of an active object starts from brain anatomy and involves the extraction of vertices, local
normals and the polygons describing the connectivity. These steps take about 7s. The reading of
functional data and association of source points with vertices on the cortex found in a
neighborhood is the longest step and takes about 70 s for 234000 vertices.
The activation mapping involves interpolation, and processing like normalization in space and
time. For a subcortical segmented structure this processing is considerably faster than the analysis
for whole brain, due to the smaller number of voxels involved.
The uses of visualization and the applications of the methods introduced in SAV are: a) for
qualitative evaluation of the segmentation results; b) visualization of activation data; c)
quantitative measurements pertaining to the structure being visualized. Other applications are also
possible and will be discussed in the section 5.4.
Apart from visual information regarding the shape and location of selected brain structures
volumetric measurements can be obtained using SAV. The activation data are visualized and
assigned colors in various color ranges and the maximum and minimum activity values are
displayed on the screen, together with information regarding the activation file location and other
parameters used during analysis, including user defined labels.
5.3.1 Qualitative Evaluation of Segmentation
We have used visualization for the qualitative evaluation of segmentation results. The transparent
view of the segmented slice in the background of the original MR slice is used to determine if
different parameters need to be employed for the semiautomatic cortical extraction procedure (in
CORTSEG) or if manual editing is necessary (using STRUCTSEG).
Chapter 5 Visualization of surface activation
94
Figure 5.3-1(A) Segmented brain (pink) superimposed on original MRI image; (B) and (C) 3D
rendering of transparent segmented brain along with segmented amygdalo-hippocampal complex (B)
and thalamus (C); (D) Outline of central sulcus on a 3D brain; (E) Segmented central sulci and (F)
segmented pre and post central gyri. In inserts the brain is displayed at the same orientations as the
segmented structures (Subject JR).
Figure 5.3-1 A shows how the qualitative assessment of the segmentation accuracy is done in
CORTSEG by continuously seeing the segmented slice in the background of the original slice, at
any one step of the segmentation procedure.
The appreciation of the structure shape and location relative to the cortex is done using the
representation of multiple volumetric objects in the same view [Figure 5.3-1, B, C and D]. The
outline of the central sulcus was traced onto the surface representation of the cortex, using VISIO.
Visualization of segmented cortical structures is necessary to compare its shape against other
segmentations or atlas images [Figure 5.3-1B -F].
5.3.2 Surface activation visualization
The main display of SAV consists of contour maps. The surface representing the selected
anatomical structure is peusdocolored in accordance to the activation values. We analyzed in this
way the cerebrum and distinct cortical and subcortical structures in an attempt to visualize sources
Chapter 5 Visualization of surface activation
95
of activity reconstructed from MEG with MFT. The cerebellum and brain stem have been
analyzed in this manner for the study of eye movements’ electrophysiology.
The structure’s surface is textured with activity values and represented either a smooth surface
with Gouraud shading or as a colored mesh, as shown in Figure 5.3-2
Figure 5.3-2. The smooth surface (a) and mesh representation (b) of the anatomy. The example shows the
amygdala activated during REM, 181 ms after eye movement to the left. The axes are: X for anterior-posterior,
Y for superior-inferior, Z for left right. Subject JR.
Smoothing, mesh decimating or various qualities of rendering can be used to increase the speed of
analysis.
Selected cortical patches like the pre and postcentral gyri were analyzed in studies of evoked fields
through stimulation of the median nerve (at the wrist). The study revealed activation on Brodmann
area (BA) 3b, hard to visualize if looking from outside the cortex [Figure 7.4.2 d]. The central
sulcus was selected for exposing activation in an area close to the bottom of the sulcus, BA3a
[Figure 7.4.2 g]. The areas around the calcarine sulcus (V1, V2) were used for visual stimulation
studies.
We have focused on the analysis of activation following arm stimulation and used current density
maps and statistical parametric
maps [Figure 5.3-3], calculated
with software developed in the
laboratory of Human Brain
Dynamics, RIKEN, BSI, led by
Dr. A. Ioannides.
Figure 5.3-3.The main display of SAV.
A) Current density map evoked by left
arm stimulation of the median nerve
R L4.3x10-
-4.3x10-
a) b)
Chapter 5 Visualization of surface activation
96
at the wrist for subject J.R. at 22ms after stimulus onset, threshold at 25%/ The outline of the central sulcus is
shown in pink. B), C) and D) show statistical parametric maps obtained for right arm stimulation of subject
R.B. at latencies: 27 ms (B) and 87 ms in (C) and (D). The hyperactivated areas are shown in a warm color scale
and the hypoactivated in a cold colorscale.
If the surface representation cannot reveal enough information, i.e. a source may be buried in the
folds of the cortex then the slice representation is used, as illustrated in Figure5.3-4. The whole
volume of the cerebrum is presented and not only the boundary GM/CSF, used for calculations.
Figure5.3-4. Slice
representations reveal
sources of activations for the
hand area buried within the
sulci (subject JR). Two
coronal cuts are presented.
The center image (A2) is at
the level of SI and the right
image (A3) at the level of SII.
The brain activation can be represented using SAV in various ways, as exemplified in Figure 5.3.5
for the left arm stimulation, at 31ms.
Chapter 5 Visualization of surface activation
97
Figure 5.3-5. The brain activation following left arm stimulation , at 31 ms can be represented using SAV as: a)
contour maps and vector maps on the cerebrum (threshold 11%; vector density 2/1000); b) the density of the
vector map can be selected by the user (threshold 11%; vector density 8/1000); c) the maximum activity vector
(points to the hand area in SI) in the background of the anatomy, which is pseudocolored in accordance to
activation values. Multiple views are necessary to infer the vector orientation (d) contour maps and vector
maps on the segmented region of interest, i.e. postcentral gyri (e) or the central sulcus (f). The dipoles solutions
derived from MEG and EEG data model the activity in BA1 and BA 3b; they are shown relative to the central
sulcus, in a top view. The pink ball indicates the dipole sense, the mauve indicates the origin. The same colorbar
as in (a) applies to all the next figurines.
In (a) contour maps texture the brain surface and vector maps convey the direction of current
density vectors. The dense mesh of activation vectors can be “trimmed” at values selected by the
user. In (a) a value of 2/1000 has been used. In (b) this value has been increased to 8/1000. The
maximum activity vector is shown in (c) in the background of the anatomical surface,
e
d
ba
c
f
Chapter 5 Visualization of surface activation
98
pseudocolored in accordance to activation values. Sometimes multiple views are necessary to infer
the true orientation of the maximum activation vector, as shown in (d). The activation data
following left arm stimulation can be shown in the background of the segmented cerebrum or
analyzed relative to the anatomy of the postcentral gyri (e) or the central sulcus, seen in (f). The
choice of using the segmented postcentral gyri is justified not only by the fast computation time
but because the source is relatively deep and difficult to visualize on the extracted cortex The
dipoles solutions extracted from MEG and EEG analysis are shown in (f) relative to the anatomy
of the central sulcus.
Visualization can be applied to identify how pathology is correlated with morphological changes,
as it is the case shown in Figure 5.3-6.
Figure 5.3-6. The arrows point to the medial ending of the central sulcus in the left column and to the
upper/medial part of the central sulcus in the center. This is where is expected to be the representation of the
lower part of the body, and in the paraplegics can be seen areas of marked atrophy. In the right column are
shown the gyral walls anterior and posterior to the central sulcus. The arrows indicate an area where the
central sulcus appears enlarged in the paraplegic (subject EP).
The region where the lower part of the body is represented in the normal and paraplegic subject
appears atrophied in the paraplegic subject compared to the normal one, age matched.
Chapter 5 Visualization of surface activation
99
5.4 Discussion
Visualization is an essential part of the segmentation process and of the functional data analysis. In
the functional imaging context visualization makes possible a unified data representation: function
in a structural context. In this way it enhances data readability and helps interpreting it by making
available structural information and spatial relationships.
The surface constrained current density analysis that we use, helps to relate the location and spread
of activity to the brain structure anatomy. The combination of these static frames into animations
enhances the understanding of how the activity evolves in time and may help reveal the
interaction/coupling between distinct sources of activations. A distributed source model, like the
one used in MFT is a prerequisite to the analysis of spatiotemporal spread of activity [Ossenblock,
et al. 1999].
It is worth noting that the slice representation of the cortical map is not directly comparable to the
surface activation map produced with SAV. The later one presents only sources in the immediate
proximity to the cortex. However the two representations are strongly correlated, especially for
superficial sources.
For deeper structures the segmented structures themselves can be used for mapping the surface
data and the activation values may be normalized to the maximum value over both time and space.
These results shall always be compared with the global view of the activation map, for it is the
main goal to establish the relationship between distinct area activations, the way they interact in
time and the path of communication.
The activation analysis is more computationally demanding if it is done for the whole cortex
[Figure 5.3.5, a -d] than if performed for the postcentral gyrus [Figure 5.3.5, e] or central sulcus
[Figure 5.3.5, f] only but is essential for identifying all the activated regions on the cortical
surface, especially in cognitive experiments, where the number of sources is not known
beforehand. The maximal activity vector gives an indication on the location and orientation of the
source but the surface representation is important for appreciating the spatial extent of the sources.
The cerebrum is thus always used in the analysis of activation data, even if a selected structure is
of particular interest. It was found that this surface based analysis works best for superficially
located and focal sources.
Distinct generators are displayed as spatially separated sources of activation, as those in the SI and
SII areas in the hand stimulation example shown in [Figure 5.3-3]. The separation power is limited
Chapter 5 Visualization of surface activation
100
by the solution grid spacing and the spatial resolution of the input data describing the distribution
of activity.
The computation time is large and the surface representation can be noisy if the data are kept in the
original form. Smoothing, mesh decimating or using a low quality rendering can all be used to
increase the speed of analysis.
In this thesis a surface representation was used, in order to help identify the relationship between
surface topography and function and the interactions between cortical areas. However even if these
methodology would be able to generate a single story on a particular aspect of brain function one
could not say that visualization has done all it could to help understand the particular problem. An
integrated view would probably need to combine information from multiple sources. The use of
color, contours, and other visual clues to differentiate receptor densities, metabolic rates, electrical
potentials, magnetic fields, and other attributes of structure and function could be used to produce
multimodal views of the brain of increased complexity with overall patterns or relationships which
may tell a different story from that told individually [Toga and Mazziotta, 1996; Roland et al.,
2001].
Chapter 6. Applications in neurophysiology
101
Chapter 6. Applications in Neurophysiology
Chapter 6. Applications in Neurophysiology........................................................................... 101
6.1. Introduction .............................................................................................................. 101
6.1.1. The somatosensory system................................................................................... 102
6.1.2. Background on the early somatosensory evoked potentials/fields....................... 105
6.2. Methods.................................................................................................................... 107
6.3. Results ...................................................................................................................... 108
6.1.3. Electrical stimulation of nerves in the limbs of normal subjects ......................... 108
6.1.4. Electrical stimulation of the limbs for a paraplegic subject ................................. 115
6.1.5. The primary visual cortex - a combined fMRI and MEG analysis ...................... 117
6.1.6. Use of anatomical constraints for EEG dipole localization. Application to central
sulcus .............................................................................................................................. 118
6.4. Discussion ................................................................................................................ 121
6.1 Introduction
The present chapter is concerned with the use of SAV in studying the early sensory processing
in the somatosensory and visual systems. The novelty in comparison with early studies which
use activation data projected onto MR slices consists in the ability to expose the relationship
between time development of activation parameters and the cortical topography.
Our main goal is to study in parallel the temporal and spatial development of the main response
in the primary somatosensory cortex (SI) as well as of lesser activations surrounding the
primary focus.
Section 6.1.1 summarizes the principles of the neurophysiology of the somatosensory system.
Our methods are described in 6.2 and our results in 6.3. Our methodology in SEF studies on
several normal subjects and a paraplegic patient are described in sections 6.3.1 and 6.3.2
respectively. Additionally SAV was used in a study on the visual system (Section 6.3.3).
The information on the anatomy which can be obtained with SAV can be useful in many other
ways and we give an example where anatomical constraints were used to aid in EEG dipole
localization (Section 6.3.4).
Section 6.3.4 consists in discussions.
Chapter 6. Applications in neurophysiology
102
6.1.1 The somatosensory system
Although sensory perception differs for each of the senses, there are three steps common to all:
1) a physical stimulus, 2) a set of events which leads to the transduction of the former into a
neuronal signal 3) a response of the nervous system which leads to perception or inner
representation of the sensation [Kandel, 1995]. A similar principle of organization is present in
all sensory systems, but the somatosensory system appears to be simpler and it was therefore
preferred as the initial study done with SAV. The general organization of the sensory system is
reflected by the sensory pathway for touch, which crosses to the brain side contralateral to the
stimulation side (at the medulla level), as shown in Figure 6.1-1.
Figure 6.1-1 a) The somatosensory pathways for touch illustrates the general organization of the sensory
system. From the peripheral receptor to the cortex the signal is transmitted to the cortex: i) the primary
sensory neuron ending transduces the stimulus into a patter of action potentials which is conducted via its
axons through the spinal cord (the dorsal column) to the medulla ii) the secondary neuron’s axon carries
the signal through the brain stem (the medial lemniscus) to the thalamus iii) the tertiary neuron carries the
information from thalamus to the cortex. b) the somatosensory cortex has been localized and divided in
several areas. The primary somatosensory cortex is located in the posterior bank of the central sulcus and
the postcentral gyrus and includes Brodmann areas 1,2,3. In the posterior parietal cortex areas 5 and 7b
are also somatosensory regions. The second somatosensory cortex (SII) is located in the upper bank of the
lateral sulcus, at about the lower end of SI. [a and b figures are from Johansenberg, 2002]. c) SI contains
four distinct sub regions organized in parallel strips, each of which contains a somatotopic representation
of the body [from Bear and Connors, 1995]
b)
a) c)
Chapter 6. Applications in neurophysiology
103
Besides touch the somatosensory system is concerned with proprioception, temperature and
nociception.
The sense of touch is mediated by mechanoreceptors in the skin, which are either: a) superficial
or deep, b) with large or small receptive fields and c) fast or slowly adapting. [Table 6.1-1 ].
Small receptive fields
cutaneous
Large receptive fields
Subcutaneous
Fast adapting Meissner Pacini, Golgi
Slowly adapting Merkel Ruffini
Table 6.1-1. Receptor types and characteristics of afferent fibers from the skin of the hand (adapted from
Frackowiak, 1997)
Different combinations of receptors are usually stimulated at once and generate distinct
sensations according to the spatiotemporal pattern of the stimulus and the type of fibers which
is stimulated.
The somatic sensory information is processed in the human brain in several distinct cortical
areas, defined by criteria related to their cytoarchitecture, pattern of connectivity, neuronal
response properties, receptive field size, and the effect of lesions on perceptual capabilities
[Kaas, 1983]. Nine areas are considered to have primarily somatosensory function and these
include [Halgren, 1990; Frackowiak et. al., 1997; Mountcastle, 1998].
a) the primary sensory areas (SI) located the posterior gyri of the central sulcus. The SI is
organized in an orderly somatotopic way, the “homunculus” representation of the body surface.
SI consists of Brodmann areas (BA) 3a, 3b, 1 and 2. These areas are arranged into parallel
strips, perpendicular to the somatotopic map. Each of them contains the orderly representation
of the body.
b) the secondary sensory areas (SII) in the ventrolateral region of the parietal cortex, along the
upper bank of the lateral sulcus. The SII overlies the insula and because of this is considered as
being part of the parietal operculum [Woolsey 1946; Maeda, 1999].
c) the areas 5 and 7b of the somatosensory association cortex, located in the posterior parietal
cortex. Neurons in area 5 are responsive to passive and active limb movements and specific
combinations of positions of joints .The anterior part of area 7, area 7b, is involved with higher
order integration. These areas may have a role in visual guidance of movement [Frackowiak et
al., 1990].
d) the granular insular and retroinsular cortex [Schneider 1993; Mountcastle, 1998; Hendry et.
al., 1999;]. The insula is an anatomical target of area SII and appears to be important in
establishing associations that involve tactile information. The posterior insula contains a
Chapter 6. Applications in neurophysiology
104
granular-isocortical area which is functionally linked to somatomotor systems, with
connections to neocortical areas, thalamus, and basal ganglia.
In addition to the well-documented motor representation, the precentral gyrus receives also
somatosensory input. Tactile and proprioceptive maps have been reported in the anterior bank
of the central sulcus (areas 4p and a) [Tanji and Wise 1981; Strick and Preston 1982; Geyer at
al. 1995]. Penfield and Rasmussen [1950], Gentilucci at al. [1998], reported that BA6 also
possesses a tactile map.
The somatosensoy areas receive information via afferents from ventro-basal thalamus (nucleus
ventro posterior lateralis) coming from cutaneous receptors (primary for the areas 3b and 1)
and deep receptors in muscles and joints (primarily for areas3a and 2). A schematic
representation of the discriminative touch path is given in Figure 6.1-1, showing how the
information from the receptor reaches the spinal cord and travels along the dorsal columns in
the spinal cord to the dorsal column nuclei in medulla, where from reaches the ventroposterior
thalamus which projects to the sensory areas in the parietal cortex.
Sensory processing in cortex is organized in cortical columns which span all layers
[Mountcastle, 1998].
All somatosensory sub modalities are finally represented at the contralateral SI, although the
level of crossing may differ from that shown in Figure 6.1-1 for discriminative touch.
The different areas differ in the proportions of cells associated with specific kinds of channels
and modalities however they are not totally distinct in function. Area 3b, for example, contains
both slowly adapting and rapidly adapting neurons, and evidence exists of their being present
in alternating columns (like the ocular dominance columns in the visual cortex) [Sur et al.
1981]. Also the input to the cortical layers of the somatosensory system is segregated, with
different channels being associated with different layers.
SII receives input from both sides of the body via afferents from the ventro-basal nuclei of the
thalamus. The SI and SII receive independent input from thalamus but they also have rich
interconnections. SI has mainly “forward” connections to both contralateral and ispilateral SII,
whereas SII has “feedback” connections to the same side SI and calossal connections with SI
and SII of the other side [Felleman and Van Essen, 1991].
As a consequence the SI is activated mainly contralaterally by unilateral touch stimulation
while in SII and insula it is common to see bilateral activation.
The evoked responses appear later and are of increasing complexity starting from the SI area 3
to 2 and 5 and then to the SII area. This supports the view that part of the somatosensory
Chapter 6. Applications in neurophysiology
105
information is serially processed from SI to SII, considered a higher ordered sensory node
compared to SI. However SII receives also independent input from thalamus [Jones, 1985].
The two posterior parietal regions associated with somatosensory processing both contain cells
responsive to tactile input. Area 5 is immediately posterior to area 2. The receptive fields of
area 5 cells can be distinguished by their size, which represent larger areas of the body surface
than the cells of SI. The area 5 cells are responsive to particular movements.
Area 7, considered an endpoint of a spatial processing system, is responsive to sensory input
from more than one modality (e.g. visual stimuli as well as somatosensory) and is considered
to be a multimodal site. It contains a lateral region, area 7b which is responsive to
somatosensory stimuli.
6.1.2 Background on the early somatosensory evoked potentials/fields
A main idea of the study using evoked potentials/fields is to extract from the recorded signal
(EEG/MEG) the activity of a functional entity and the relationship of this activity to the
applied stimulus and its properties. This activity is considered to be either: a) the result of
activation of a group of interconnected neurons in response to the event, which must be
segregated from the ongoing activity (background) or b) the reorganization part of the ongoing
activity [da Silva, 1999]. These two points of view on the nature of the evoked activity gave
rise to two directions of studying them. The traditional approach relies on averaging over many
trials the signals recorded for a given time interval before and after the stimulus in order to
emphasize the response which appears consistently following the stimulus application and thus
increase the “signal to noise ratio”. The second approach attempts to preserve information
found in single trials and study the reorganization of phase spectra of the activity. The recent
work of Ioannides and coworkers [Ioannides et al., 2002a] emphasizes the importance of
studying information present in single trial activity.
The pattern of activation allows the division of the evoked response into components according
to topography, polarity, latency [Halgren, 1990], frequency or statistical properties
SEPs are produced by synaptic relays or tracts in the specific somatosensory pathway: dorsal
column in the spinal cord, medial lemniscus in brainstem, nucleus posteromedialis and nucleus
ventralis posterolateralis in thalamus, and the rolandic cortex.
The stimuli for evoking SEP are usually 0.2ms shocks repeated several hundred times for
producing the average signal with its salient features The median nerve is stimulated at the
wrist and the posterior tibial nerve at the knee or ankle.
Early work using direct electrical stimulation of the cortical surface [Foerster, 1936; Penfield
and Boldray, 1937] demonstrated the orderly representation on the cerebral cortex (SI) of
Chapter 6. Applications in neurophysiology
106
sensation coming from various parts of the body. Newer work done by Kakigi and coworkers
[Kakigi et al, 2000] found this somatotopic representation in SI but also a more diffuse one in
SII using SEF. They reported SII components at 80-100ms after stimulation and a bilateral
function of SII.
The EEG evoked responses are described in [Halgren, 1990] and classified as: short latency
subcortical components (latency earlier than 20ms), middle latency components from the
primary cortex and supplemented by related areas (20-80 ms), and long latency components,
which overlap with multimodal cognitive components. It is considered that the early
components are not affected by repetition, attention or sleep while the later are affected to
some degree.
The EEG components and generators for the SEP following median nerve stimulation are
illustrated in Figure 6.1-2
Figure 6.1-2. SEP time course of activation for the pre and postcentral sources [from
Halgren, 1990]. For a detailed explanation see text.
The ascending action potential arrives at the brachial plexus at a latency of about 9ms. The first
clearly identified scalp component appears at 13 ms, probably generated by the termination of
dorsal column fibers in the cuneate nucleus. It is followed by N18.
The predominant potentials are a parietaly negative, frontally positive potential peaking at
about 20 ms, followed by a parietaly positive, frontally negative potential at 30ms. The parietal
and frontal peaks may not be exactly synchronous, and smaller peaks may be present in
between those two parietal: N20-P30, or frontal P20-N30. All peaks are contralateral at the
cortex [Halgren, 1990]. Because of the latency differences some authors refer to the parietal
N20-P27 and frontal P22-N30.
The generators of these potentials have been the subject of discussion. Some consider a single
generating dipole located in the posterior bank of the CS, resulting in opposite polarities frontal
versus parietal. Another theory postulates separate sources: one radial dipole in the crown of
Chapter 6. Applications in neurophysiology
107
the postcentral gyrus accounting for the parietal peaks and one radial dipole in the crown of the
precentral gyrus for the precentral
MEG studies identify many of these sensory areas [Hari and Forss, 1999; Kakigi et al., 2000;
Kakigi et al., 2002] and reveal a complex cortical network, widely distributed, activated in
response to simple somatosensory stimuli. The first SI activation is robustly identified at about
20ms [Hari and Fross 1999], and this peak is identified as N20/M20 in EEG/MEG
respectively.
A major advantage of MEG is than it can easily identify through SEF the activities in SII,
where it is difficult for SEP to detect them because of the location and orientation of dipole
sources [Hari e al., 1990, Hari e al 1993, Mima et al. 1997]. The SII responses are detected
usually 70-100 ms after stimulation [Simoes and Hari 1999, Kakigi 2000; Disbrow et al, 2001,]
but some earlier responses have been reported as well [Karhu and Teshe, 1999; Korvenoja et
al. 1999]. The relative timing, interactivity and characteristics as plasticity in SI and SII areas
have been recently studied using single trial SEF [Ioannides et al. 2002a]. The pathophysiology
of somatic sensation also started to be explored [Ioannides et al. 2002b].
6.2 Methods
Our main goal is to study in parallel the temporal and spatial development of the main
somatosensory response as well as the smaller activations surrounding the primary focus in the
rest of SI area bilaterally. For this purpose we analyze the spatiotemporal distribution of
current density following arm and foot stimulation around BA3b and for the second
somatosensory area. We examine the activations at latencies between 10 ms and 31 ms after
stimulus onset for the arm stimulation and up to 100 ms for foot stimulation.
The activation data are provided by MFT analysis [Ioannides et al, 1990] of the MEG signals
or BESA analysis [Scherg, 1990] of EEG signals. Tomographic reconstructions of the current
source distributions are obtained in general from MEG but also a dipole solution is used for
comparison purposes. Independent software was used also for generating dipole solutions from
EEG data, as well as for generating the MEG cortical and statistical parametric maps.
The cortical segmentation is performed using CORTSEG.
Areas normally not visible on the surface representation of the segmented brain can be exposed
if instead of the cortical surface, selected brain structures or cortical patches are used. Such
selected structures were the central sulcus (CS), the pre- and post- central gyri (PreCG and
PostCG), the area around the calcarine sulcus, etc. For the first study, the STRUCTSEG
module has been used to segment the sensory and motor strip, as well as the central sulcus. For
Chapter 6. Applications in neurophysiology
108
the second study the area around the calcarine sulcus, including the V1 visual cortex has been
segmented. The VISIO module has been used to display and analyze the activation data,
resulting from MFT analysis of MEG data or BESA analysis of EEG data.
The activation maps may consist in: a) either cortical current density maps or b) statistical
parametric maps, basically the p factors obtained following t tests or Kolmogorov-Smirnov
tests between two conditions. The values of the modulus for the current density vectors is
shown in a color scale ranging from yellow to red, where red are the most significant activation
values. A different colorscale is used to convey information on the direction of the cortical
density, by means of color coding. Current having the same direction with the surface normal
are shown in the yellow to red scale. Those in opposite direction are shown in the green to blue
scale. The same color scale is used for the statistical parametric maps. The red and the blue
colors correspond to the most significant values for positive or negative values respectively.
At first we examine the electrical activity on the background of the segmented cortex surface
and then on the background of the segmented brain structures. We will use a very low
threshold for the activations to emphasize the distributed nature of the solutions. Note however
that in all the reported investigations a much higher threshold was used and statistical
significance was computed for each grid point activations. Since we are primarily concerned
with the display capabilities of SAV we chose the lower threshold to allow the reader to
inspect the ability of SAV to extract widely distributed forms of activations. We will then
discuss the relative advantages of these alternative visualization modes for different case
studies.
6.3 Results
6.3.1 Electrical stimulation of nerves in the limbs of normal subjects
Numerous studies have demonstrated that the first significant volley of activity after
somatosensory stimulation arrives about 20 ms after the onset of median nerve stimulation and
in about twice as long for foot stimulation [see section 6.2]. The main activation is in
Brodmann area 3b (BA3b), on the anterior bank of post central gyrus, buried inside the central
sulcus, so usually it is difficult to study with precision its topography and temporal evolution.
In earlier studies it was enough to show that the activity was in the vicinity of BA3b at roughly
the right place according to the homunculus representation of the body.
In the first application we study the activation maps obtained following stimulation of the
median nerve at the wrist. The current density map overlapped on the cerebrum surface [Figure
6.3-1] shows that in response to left arm (LA) stimulation of the hand at the wrist a strongly
Chapter 6. Applications in neurophysiology
109
activated area is centered around the contralateral SI hand area and that there is also a lower
activated area, possibly corresponding to SII in subject 1. For subject 2 the statistical maps
following right arm (RA) stimulation at the hand indicate again high activity in the SI at early
latencies (B), hint to the possibility of an early activation in SII. At later latency (87ms) strong
activations appear in both SI and SII contralaterally, while some activation of the lower
postcentral gyrus, possibly corresponding to the ipsilateral SII, may be activated (C) in addition
to the contralateral somatosensory cortex (D).
Figure 6.3-1. (a) Current density map evoked by left arm (LA) stimulation at the median nerve for subject
J.R., 22ms after stimulus onset, threshold 25%, normalized per time slice. The activation values are
expressed in arbitrary units (au). The outline of the CS is shown in pink. b), c) and d) statistical
(Kolmogorov-Smirnov) KS parametric maps obtained for RA of subject R.B. at latencies 27ms, (b) and 87
ms (c from a right lateral view, and d from the left lateral view). The colorscale indicates the hyperactivity
in the “warm” color range and the hypoactivity in the “cold” color range. The signed, modified p factors
are displayed. A low threshold is used: 10% corresponding to p < 0.09.
We expect the main generator to be buried within the sulcus, probably in area BA3b and
therefore not directly visible from such a view. The projected activation does not appear very
focal at this threshold and the deeper the source is the larger the cortical area where it projects.
For this reason we aim to obtain a better insight by studying the same activation data in the
background of anatomy restricted to the segmented somatosensory strip. Specifically the
anatomy represents the anterior wall of the postcentral gyrus, the central sulcus and the
posterior wall of the precentral gyrus, segmented as shown in Figure 6.3-2 a. This figure shows
features extracted from the time course of activity following median nerve stimulation at four
Chapter 6. Applications in neurophysiology
110
latencies, selected from averaged activation curves. Figure 6.3-2 b shows that at 15 ms there is
no significant (> 20%) activation present on the postcentral gyrus (PostCG) wall facing the
central sulcus (CS). The first response above 20% is visible at 22 ms after stimulus onset
(Figure 6.3-2 c). The hand area and particularly the area BA3b appear activated (Figure 6.3-2 c
and d), in the contralateral primary somatosensory cortex (SI). At 31 ms the response is
stronger and more widespread though still highly localized in area BA3b on the anterior wall of
the PostCG (Figure 6.3-2 d). The current direction is approximately perpendicular to the
surface of the post central gyrus. The color coding helps notice the current reversal (blue-green
for currents entering and red-yellow for currents exiting the cortex) for the 31ms activation
peak relative to the one at 22ms.
An analysis using lower activity threshold (15%) reveals that a more posterior parietal source
appears active at 22 ms (Figure 6.3-2 e), while no activation is noticeable on the anterior wall
at this threshold. Since the posterior wall of the pre central gyrus is activated at 22 ms
following left arm stimulation (threshold at 25%), this indicates that there may be an anterior
activation source (Figure 6.3-2 f).
A smaller activity is noticed in the postcentral gyrus at 26 ms and indicates that the source of
activation in the hand area is located at the top of the gyrus, in the Brodmann area BA1 (Figure
6.3-2 g). A posterior view of the postcentral gyrus, at 31 ms, exposes the two distinct,
separated sources of activation located in SI and SII (Figure 6.3-2 h).
We present results for the left arm stimulation but similar studies have been done for the right
arm stimulation and the last two figures (Figure 6.3-2 i and j) show the central sulci activated
contralaterally to the stimulation. The inferior view is chosen to expose the area BA3a.
Chapter 6. Applications in neurophysiology
111
Current density maps (in arbitrary units (au)) over the
postcentral gyrus (PostCG), precentral gyrus (PreCG) or
central sulci (CS), elicited by arm stimulation at median nerve
from averaged data. The data are normalized across all time
moments and for 4 runs for the left arm (LA) and right arm
(RA) stimulation individually. For the latency corresponding
to the maximal activation value, obtained at 31ms the
maximal activity vector is shown (Subject 1). The color scale
of current density at the bottom applies to all images.
a) Details on the segmentation of the PostCG, CS and PreCG;
b) PostCG in anterior, top and right view show no significant
activation (threshold at 20%) at 15ms following left arm
stimulation. The thalamus is included for orientation
purposes.
c) PostCG in anterior view at 22 ms following LA stimulation
show activation (threshold at 20%) in the right hand area,
BA3b
d) PostCG in anterior view at 31 ms following LA stimulation
show maximal activation (threshold at 63%) in the right hand
area. Notice the current reversal for the 31ms activation peak
relative to the one at 22ms.
e) PostCG in posterior view at 22 ms following LA stimulation
(threshold at 15%). A more posterior parietal source appears
active while no activation is noticeable on the anterior wall at
this threshold.
f) PreCG in posterior view at 22 ms following LA stimulation
(threshold at 25%) indicates that there is an anterior
activation source.
g) PostCG in anterior view at 26 ms following LA stimulation
(threshold is 9%) indicate that Brodmann area BA1 is
activated.
h) PostCG in anterior view at 31 ms following LA stimulation
(threshold is 12%) show two distinct, separated sources of
activation located in SI and SII.
i) Central sulcus in inferior view at 31 ms following left arm
stimulation (threshold is 63%) indicate activation in the
contralateral hand area, here the area BA 3a is exposed.
j) Central sulcus in inferior view at 31 ms following right arm
stimulation (threshold is 70%) indicate activation in the
contralateral hand area, here the area BA 3a is exposed.
Figure 6.3-2. Early SEF over the PostCS,
PreCG and CS (subject J.R.).
Chapter 6. Applications in neurophysiology
112
Figure 6.3-3 shows the same time sequence of activations (current density maps) as Figure
6.3-2 but from the same (anterior) view and with constant threshold (5%). The data refer to left
arm stimulation at median nerve from averaged data, from 10 to 30 ms after stimulation. The
data are normalized across all time moments and for 4 runs. For the latency corresponding to
the maximal activation value (at 31 ms) the maximal activity vector is shown in Figure 6.3-3.
Figure 6.3-3. Spatio temporal course of activation over the PostCG following LA
stimulation. The time latencies are 1, 15, 22, 26, and 31 ms after stimulation and the
threshold is 5%. Above 20% the first significant activation is seen at 22ms. (Subject J.R.)
Figure 6.3-3 and Figure 6.3-4 show distinct foci of activations after stimulation of the median
and tibial nerves. The foci at the early latencies are at the expected locations for SI for the hand
and foot, on the contralateral post-central gyrus to the stimulated limb. In both figures
activations are seen in locations where the second somatosensory area is expected, at the base
of the postcentral gyrus.
R
L
Chapter 6. Applications in neurophysiology
113
Figure 6.3-4. Current density maps (modulus only), following hand and leg stimulation.
The first row shows the results for left arm (LA) stimulation at 22ms after stimulus onset
and 48 and 95 ms for the foot stimulation (LF) respectively. The second row shows the
results for right arm (RA) stimulation at 36ms after stimulus onset and 46 and 75 ms for
the foot stimulation (RF) respectively. (Subject R.B.).
These SII activations are seen either ipsi- or contra-laterally to the simulated limb and usually
at longer latencies, starting at 31 ms for arm stimulation and at 75 and 95 ms for right and left
foot stimulation respectively.
Several studies have addressed the question of serial versus parallel activation of SII (see
[Ioannides et al. 2000a] and references therein) allowing the visualization of very early
activation (Figure 6.3-5).
Using the complementary views available in SAV the SI and SII activations at the earlier
latency of 22 ms are better delineated, showing a distinct SII activation which is much weaker
than the SI activation. The whole brain view on the left column provides only a hint of separate
activations in SI and SII since the main SI focus is deep in the sulcus while secondary
activations appear on the brain surface in a more wide spread fashion. Strategic placing of
coronal, axial and sagittal slices provide an accurate alternative representation of brain
LA 24 ms LF 48 ms LF 95 ms
RA 36 ms RF 46 ms RF 75 ms
1.98*10(-3) au 2.60*10(-4) au 4.80*10(-4) au
2.36*10(-4) au 3.03*10(-4) au 4.06*10(-4) au
4.88*10(-6) au 4.88*10(-6) au 4.88*10(-6) au
4.88*10(-6) au 4.88*10(-6) au 4.88*10(-6) au
Chapter 6. Applications in neurophysiology
114
activation data and show conclusively coactivation of SI and SII at the early latency of 22 ms.
The cuts placed to capture best the SI activity are in the middle column of Figure 6.3-5. The
slices placed over the SII area are in the right column of Figure 6.3-5. They show activation in
SII which is weaker than the one in SI but clearly discernible if using a low threshold (25%).
Figure 6.3-5. 3D view with coronal (A), horizontal (B) and sagittal (C) cuts at the level of
SI at the hand area detailed in the middle image column. The current density maps are
elicited by left arm stimulation (the median nerve) at the wrist. The right column shows
slices cut at the level where SII activation may be expected. At this early latency (22ms)
the SII activation is much weaker than the SI but of distinct localization. The threshold is
25%. For color scale see Figure 6.3-3 (Subject J.R.).
SI
SI
SI SII
SII
SII
A1
B1
C1
A2 A3
B2 B3
C2 C3
Chapter 6. Applications in neurophysiology
115
6.3.2 Electrical stimulation of the limbs for a paraplegic subject
In a recent study [Ioannides et al., 2002b] the MEG signal was recorded following electrical
stimulation of upper and lower limbs in both normal and paraplegic subjects, with clinically
diagnosed complete paraplegia.
MFT analysis identified foci just behind the central sulcus consistent in location with SI for
foot and arm and SII areas. Activation of the SI foot area was identified in normal and
paraplegic subjects when the arms (above the lesion) were stimulated. Activations were also
identified following electrical stimulation of the lower limbs in both normal and paraplegic
subjects. The surprising activations with stimulation below the lesion in paraplegic subjects
previously diagnosed as complete were found in the SI foot area and/or nearby cortical areas,
they were weak and not well time-locked. Statistical analysis of the arm and foot activations
from patients identified statistically significant activations in extended areas which included
the expected location of the SI area for the arm and foot. In general, the activations in the
paraplegic subjects were extended outside the primary area and for foot stimulation were
identified when large time windows were used in the statistics.
We used SAV and displayed activation data imported from the above study [Ioannides et al.,
2002b] in the background of the segmented post- and pre-central gyri. These displays show the
spread of the activations beyond the postcentral sulcus. The contours bounding statistically
significant activations (p<0.05 or p<0.001) along perpendicular cuts for the arm (red contours)
and foot (yellow contours) are displayed on the whole brain and on the pre-and postcentral gyri
segmentations. The displays show the results at the latencies where the most significant
activations for each case are. For arm stimulation (median nerve, above the lesion) the
significant activation is at similar latencies as for normal subjects, 24ms, but the activated area
extends beyond the expected area for SI. For foot stimulation (tibial nerve, at the ankle) the
significant activation is delayed by 30 to 50ms compared to normal subjects (Figure 6.3-6 ).
Chapter 6. Applications in neurophysiology
116
Figure 6.3-6. Contours outlining areas of statistically significant activations are shown
following arm and foot stimulation in a paraplegic patient (subject EP). The first row (A,
D) shows the contours relative to the cortical anatomy (the contours intersect a
transparent brain surface). The central sulcus has been outlined by placing a blue
marking point at the bottom of the sulcus in each MRI slice. Depending on the aspect of
viewing the 3D rendered brain parts of this blue outline may appear as covered by
cortical folding. The lower rows (B, C, E, F) show the same contours overlaid on the
Chapter 6. Applications in neurophysiology
117
central sulcus walls (pre- and postcentral gyri segmented together). Shown in red are the
contours obtained at 24ms following right arm (RA) stimulation (p<0.05) and left arm
(LA) stimulation (p<0.001). In yellow are the contours obtained following left foot (LF)
stimulation (75ms, p<0.05) and right foot (RF) stimulation (99ms, p<0.001). The contours
were traced in three planes (sagittal, axial and coronal) for LA; two planes (axial and
coronal) for each of the two activated areas (approximating BA3a and BA3b) for RA and
two planes (axial and sagittal) for LF and RF. B shows the gyri viewed from an anterior-
right perspective with contours from LF and LA stimulation. C presents the same data,
from a posterior-right perspective; the gyri were made transparent to allow visualization
of the areas where the contours intersect the gyri. E shows the gyri viewed from an
anterior-left perspective with contours from LF and LA stimulation. F presents the same
data on transparent gyri, from a posterior-left perspective. Note that the contours extent
into and also posteriorly to the postcentral gyrus. An animation of above figure can be
found in http://nucleus.med.upatras.gr/~alex/sav_results.html.
6.3.3 The primary visual cortex - a combined fMRI and MEG analysis
In a recent study [Moradi et al 2002] the tomographic localization of activity within human
primary visual cortex (striate cortex or V1) was examined using whole-head MEG and 4-Telsa
functional magnetic resonance imaging (fMRI). Circular checkerboard pattern stimuli confined
to part of the lower quadrant of the visual field were used designed to excite the dorsal part of
V1 which is distant from the V1/V2 border and from the fundus of the calcarine sulcus. Both
fMRI and MEG identified spatially well-overlapped activity within the targeted area in each
subject. The mean separation between V1 activation centers identified by fMRI and the earliest
MEG activations around 40 ms after stimulus onset was only 3-5 mm.
SAV was used with activation data from the above study [Moradi et al, 2003] and made
apparent the remarkable agreement between the two tomographic reconstructions in both
location and shape when the MEG and fMRI solutions were displayed together with the V1/V2
boundary on the background of anatomy, as shown in Figure 6.3-7.
The figure shows the activation following the presentation of the stimulus on the lower right
visual field for one subject. For both MEG and fMRI maps of statistical significance are
displayed compared to a baseline condition when no stimulus is presented. The MEG result is
displayed as a filled red (blue) contour for increase (decrease) of activity at a p<0.01 level.
The computation of the statistical significance compares 6 ms windows with similar windows
from the baseline. The comparison is made throughout the post-stimulus period by advancing
the center of the active window in 3.2 ms steps. The latency used in the display is 40 ms and it
Chapter 6. Applications in neurophysiology
118
corresponds to the first significant increase in activity in V1. The fMRI activation at p<0.01 is
bounded by the yellow contour which surrounds the red area (MEG area). The yellow line
above the calcarine surface corresponds to the V1/V2 boundary. The display demonstrates that
the fMRI and MEG estimates of increase activity coincide and that the area of precise
correspondence is on the contralateral upper part of the calcarine entirely within V1. Note also
the hypoactivation seen in the MEG statistical comparison on the ipsilateral V1.
Figure 6.3-7.Post-MFT statistical map of current density for the onset of M50 (at 40 ms,
stimulus size 5.2 degrees). Data from [Moradi et al, 2003]. The map is superimposed on
3D rendering of the upper banks of the calcarine sulcus. The highly significant increase of
activity (p < 0.001) is on the left V1 (shown in red), contralateral to the stimulated visual
field. Note also the hypoactivation (shown in blue) on the ipsilateral side. The yellow
outlines represent fMRI contours for statistically significant increases in activity and
V1/V2 contours (Subject V.P.).
6.3.4 Use of anatomical constraints for EEG dipole localization. Application to
central sulcus
We have used SAV to display EEG dipole solutions from plain EEG or simultaneous
MEG/EEG recordings, using BESA and MFT [Figure 6.3-8]. This helps to conceptualize the
relationship between the direction and location of the dipoles to the local anatomy i.e. that of
the somatosensory evoked dipoles in posterior central sulcus to the convoluted surface of the
sulcus.
Chapter 6. Applications in neurophysiology
119
Figure 6.3-8. Dipole sources resulting from analysis of MEG and EEG signals presented
in the background of the CS anatomy. The posterior view is at the left, right view at
center and top view at the right. The sources correspond to activation of the BA3b and
BA1 areas. Subject J.R.
In preliminary one case study [Zainea et al., 2002], EEG somatosensory evoked responses
following electrical stimulation of the left median nerve have been recorded using a 64 channel
Neuroscan machine. SAV was used in an effort to provide restrictions of the solutions in terms
of locations and orientations [Figure 6.3-9] for an independent analysis. One thousand epochs
recorded from -50 to 200 ms around the stimulus were filtered (1-290Hz), cleaned from eye
movement and muscle artifacts using ICA (independent component analysis) and averaged.
The electrodes locations were digitized with a Polhemus FastTrack device and converted to
MRI coordinates. The postcentral gyri contralateral to the stimulation site have been segmented
and a smooth mesh of the surface has been obtained using SAV. The locations of the surface
vertices and the normals characterizing the local anatomy have been brought into the same
coordinate system as the functional data. Source localization has been performed linearly
fitting over the middle latency interval (20-50) ms of the average SEPs the amplitudes of any
two dipoles taken from the set of surface normals. Two major sources of mid-latency electrical
activity were found as a result of an exhaustive search over the dipole candidates provided by
the local normals. The sources locations (Figure 6.3-10) correspond to the anatomical brain
areas BA3b and BA1 of contralateral primary somatosensory cortex.
EEG
EEG MEG
EEG
MEG
EEG
ME
EEG
Chapter 6. Applications in neurophysiology
120
Figure 6.3-9:Segmented postcentral gyrus with associated local normals. Subject O.Z.
Figure 6.3-10: The dipole solutions corresponding to sources in area BA3b (upper row) and area BA1
(lower row). Images produced with independent software produced by Marc Schellens and modified by
Ovidiu Zainea. Subject O.Z.
We made use of the detailed knowledge of anatomy to refine the accuracy of the localization
process by confining the search within the area of interest, the contralateral postcentral gyrus,
and the orientation to the columnar orientations characteristic to the cortical sheet. The use of
an exhaustive search method instead of a local nonlinear search avoids the local minima,
leading to a more accurate result.
Chapter 6. Applications in neurophysiology
121
6.4 Discussion
High resolution imaging of structures extracted from MRI is helpful in analyzing activation
patterns within the hand area of the central sulcus, or at the larger scale of the somatosensory
strip, because the distinct cortical areas involved can span less than 1cm in the anterior-
posterior plane [White et al., 1997]. The 3D visualization offsets the ambiguity introduced by
the curved path of the central sulcus and the neighboring gyri, which penetrates 2D slices at a
variety of angles, and make the precise assignment of activation to specific sulcal or gyral
regions difficult [Gelnar, 1998] .
Because cortical morphology is organized along circumvolutions one can use SAV to avoid
restricting the analysis of the functional topography along the surface of the gyrii and exploit
the information located on deeper structures like sulcus walls and fundii. These have potential
to reveal some stronger inter-individual regularity in their spatial organization rather than
according to their traces along the cortical surface only [Sastre-Janer et al. 1998].
We have used SAV to help the conceptualization of the spatial relationship between equivalent
current dipoles (ECD) and the local anatomy and we give an example from the combined
MEG/EEG study on the somatosensory system. The local anatomy refers in this case to the CS
anatomy.
The previous studies using the slice representation of activity data projected onto the anatomy
often fail or have difficulty to convey the spatial relationship and the time interactions between
various areas involved in the sensory processing. The separation of activations in adjacent
areas requires careful selection of slices, but even the best choice of slice views often fails to
represent the large scale relationship between the activated brain areas. In recent studies of SI
and SII activations in normal [Ioannides et al., 2002a] and paraplegic [Ioannides et al., 2002b]
subjects early activation of SII was in evidence but the distinction and relationship between the
SI and SII activations so identified was not easy to make in terms of images of activations in
any one or more slices. We demonstrate here the advantage of the SAV representations for the
visualization of separate activations in SI and SII somatosensory areas and in the parts of SI
representing arm and foot from these two studies.
The idea of serial processing in the somatosensory cortex is supported by the organization of
the cortex into large, somatotopically organized primary receiving cortices which project
information to smaller association areas. In support of this idea come the patterns of responses
to median nerve stimulation recorded by MEG sensor arrays. Initial activation in SI about 20
ms after stimulation, followed by processing in SII at approximately 100 ms. However, the
Chapter 6. Applications in neurophysiology
122
anatomical connectivity of the somatosensory system also suggests simultaneous participation
of widely separated cortical areas in the early processing of sensory input.
We have analyzed the early evoked somatosensory fields and shown focal activations in the
hand area BA 3b with peaks at 22ms and 31ms, in between which happens a reversal of current
The SI activation is accompanied by a weaker activation in the SII area or the parietal
operculum as seen from a lateral view of the cortical surface. These areas, usually viewed as
higher order processing areas for somatosensory perception appear coactivated with SI during
the early processing of somatosensory input.
While SII appears activated bilaterally [Simoes and Hari, 1999; Ioannides et al, 2002a] it
maintains a clear contralateral dominance, supporting the view that it plays a role in supporting
a unitary body scheme by integrating information from the two body halves.
The combination of high resolution MR imaging with the MEG data, as implemented in SAV,
has shown the possibility to discriminate between closely situated but highly localized, distinct
sources of activity in the somatosensory (SI/SII and BA3b/BA3a) and visual system (V1).
Finally, we examined the potential of SAV in pathophysiological studies, specifically in a
study on paraplegia.
Mainly the current density maps but also statistical maps are used with SAV. Multimodal
information, extracted for example from fMRI, can be incorporated and allows comparing the
localization of activities across these modalities. We use the contour options in different ways
to demonstrate how different types of functional information can be superimposed on the
relevant 3D anatomical context.
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
123
Chapter 7. A Software Tool for Interactive Determination of
the Plane of Cut through the Rat Brain
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain ......................................................................................................................................... 123
7.1. Introduction .......................................................................................................... 124
7.2. Methods and materials ......................................................................................... 124
7.2.1. Reconstructing the rat brain (structures) based on atlas images .................. 124
7.2.2. Search protocol............................................................................................. 126
7.3. Results .................................................................................................................. 127
7.4. Practical solution .................................................................................................. 128
7.5. Discussion ............................................................................................................ 129
7.6. Conclusions .......................................................................................................... 130
The aim of the present work is to use visualization tools to help planning an
electrophysiological experiment on rat brain slices. Namely we aim to provide an interactive
tool for finding the plane of cut through the rat brain which yields the maximum section
through a selected brain structure or a combination of structures, i.e. a path of nerve fibers and
the area they are known to connect to.
A set of two-dimensional coronal images from a rat brain atlas is preprocessed. The structures
of interest, landmark structures, and the bulk of the brain are assigned specific gray levels, and
the images are registered.
Any rat brain atlas can be used to provide the desired set of two-dimensional data used to
reconstruct the brain volume. Selected rat brain structures can be analyzed with this method,
provided that the preprocessing of the atlas images is performed. As a first application of this
method we have identified a set of parameters for which a 0.5 mm thick slice can be cut
through the rat brain so as to include as many as possible intact connections between the fornix
fibers and the mammillary bodies.
The same type of application could be extended to visualization of histochemical and
autoradiographic data.
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
124
7.1. Introduction
The classical rat brain atlases provide images at either coronal, sagittal, or horizontal planes.
Frequently a different orientation of sectioning is desired for optimal visualization of specific
brain areas and pathways stained with specific markers (including Horse Radish Peroxidase
staining of axons, receptor autoradiography, etc.). This need is also currently emerging with the
advent of experimental electrophysiology studies in vitro [Agmon and Conners, 1991;
Dingledine et al., 1980].
We aimed to develop an interactive tool for finding the plane of cut through the rat brain which
would yield the maximum size section through a selected brain structure or a combination of
structures, i.e. a path of nerve fibers along with the area they are known to connect to. Such a
tool would assist in the planning of in vitro electrophysiological studies in brain slices, by
determining and graphically illustrating the optimal plane of brain sectioning. The latter should
be better defined interactively and expressed in an operational way, i.e. the position and angle
of the plane in reference to the three planes of stereotaxic coordinates as well as to visible brain
landmarks.
An important circuitry within the limbic system relies on the transmission of information from
hippocampus via fornix fibers to the mammillary bodies (MB), a paired nuclear mass in the
basal part of the brain [Kostopoulos and Phillis, 1977]. In vivo studies established the
excitatory nature of the fornix input to the MB; but were unable to investigate the detailed
synaptic mechanisms and possible plasticity of this connection. Such precise studies demand
the experimental convenience of brain slices maintained in vitro. However, due to the tortuous
course of fornix (see Figure 7.3-1), slices cut at any of the three conventional stereotaxic planes
do not yield functional connections between fornix and MB neurons. We therefore tested the
ability of the developed software to help finding the plane of cut through the rat brain, which
would preserve as much as possible intact connections between the fornix fibers and the MB.
7.2. Methods and materials
7.2.1. Reconstructing the rat brain (structures) based on atlas images
The present approach involves two initial steps, the preprocessing of the atlas images and the
volume reconstruction in order to determine the proposed plane of cut, offered to the
experimenter for further interactive refinement of the optimal plane.
The resources employed consist of a stereotaxic rat brain atlas [Paxinos and Watson, 1997], an
image processing software (Paint Shop Pro) and the IDL computing environment.
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
125
76 coronal images spaced at an average distance of 0.3mm along the anterior to posterior axis,
from the Paxinos and Watson rat brain atlas are the input data. A prerequisite for using these
images is that they should be in digital format and registered. The first condition is fulfilled by
the atlas images, while registration of the images is achieved with reference to grid points of
known stereotaxical coordinate, which are identified in all used atlas images. The images are
assumed to be aligned.
Figure 7.2-1. The reconstruction of the brain and selected brain structures starts with: (a) the original atlas
image (from [Paxinos, 1997]); (b) This is preprocessed and the structures of interest are assigned specific
gray levels (F = fornix, MB = mammillary bodies, H= hippocampus, R = the reference point used for
registration; The results consist in (c) reconstructed cerebrum and , (d) mamillary bodies and part of the
fornix. A=anterior, P=posterior.
In the next preprocessing step the brain contour and the structures of interest (MB, fornix, and
hippocampus) are isolated and specific gray levels are assigned to each of them [Figure 7.2-1 a
and b].
The resulting contours are stacked and used for creating two volumetric data sets: the whole
brain (with a low resolution), and the volume of interest comprising the fornix and MB (with
high resolution).
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
126
Paint Shop Pro is used in the preprocessing step. IDL programming is used to reconstruct the
outline of the selected structures (Fig. 7.2-1 c and d) and planes of arbitrary orientations
through the volume.
Bilinear interpolation is used to generate intermediary slices in order to obtain isotropic voxels.
We developed software to interactively find the plane of cut that contains the maximum area of
selected structures, in this application the MB and fornix.
7.2.2. Search protocol
The process of finding the plane that preserves the maximum intact connections between the
fornix and the MB is summarized in Figure 7.2-2. The problem is reduced to finding the plane
that (a) contains the maximum area of fornix, and (b) intersects the MB. Another search can be
made for obtaining a plane that contains the maximum cross-section with the MB.
Figure 7.2-2 The search protocol
The data set containing the whole brain is used in the beginning, for selecting the structure of
interest. The purpose is to find the sectioning plane including a maximum area of this structure.
The viewing parameters can be varied to get information about the shape and location of the
structures
The actual search is performed within the volume of interest containing the fornix and the MB.
The first sectioning plane is selected intuitively, placing the mouse cursor at the desired
location. The software returns the plane parameters with reference to the bregma, a landmark
Automatic orthogonalscanning
Selected plane with max area
Mouse selectrotation center
Select varyingparameter
Automatic oblique scanning
Stop
Select thestructure of interest
Volume of interest
Draw plane through brain volume
Brain volumeStart
Area=max? (Y/N)
YN
Oblique scanning
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
127
on the skull. These parameters are the three coordinates of the plane normal origin, and the
angles made by the plane normal with the axes.
The volume is scanned with a fixed step, parallel to one axis, and the plane containing the
maximum area of the structure is redrawn. The maximum area is held in memory for future
reference.
An oblique scan is performed starting from the plane previously found: the origin of the plane
normal on the cut plane is selected with the mouse and the plane can be rotated around either
axis or translated with a fixed step. The cut areas are compared with the maximum value and
the plane with maximum area is held in memory, with its associated parameters.
In case the suggested sectioning plane is considered not feasible or convenient to be cut by the
experimenter in practice the search procedure is restarted.
Once the best orientation of the sectioning plane is found in the volume of interest, the same
plane is drawn through the whole brain volume.
The sectioning plane parameters can be varied as follows: a) varying the plane orientation,
while keeping constant the distance from the plane to the origin of the coordinates system
(bregma in the present case); b) varying the distance from the plane to the origin of the
coordinates system, while keeping the same plane orientation; c) automatically scanning the
volume along a chosen axis; d) automatically scanning by rotation around a chosen axis.
7.3. Results
The proposed software tool was used for three-dimensional visualization of the brain and
structures within the brain: MB, hippocampus and fornix fibers, providing visual clues about
the spatial orientation of the selected structures (Figure 7.3-1).
Figure 7.3-1: View of the volume of interest (A =
anterior, P = posterior, L = left, R = right, D = dorsal, V
= ventral; F = fornix, MB = mammillary bodies, H =
hippocampus)
H
F MB
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
128
We have applied the method to analyze the fornix and MB. We have identified the parameters
for which a 0.5mm thick slice can be cut through the rat brain so as to include as many as
possible intact connections between the fibers in the fornix and the MB. Suggested parameters
of the cut plane are given relative to the bregma and with reference to the right-to–left (RL),
dorsal-to-ventral (DV), posterior to–anterior (PA) axes: the coordinates of the plane normal
origin: 1.4mm left of bregma, 8.8mm ventral and 4.8mm posterior, angles made by the plane
normal with the axes: 113o relative to RL, 148o relative to DV, 71o relative to PA. These angles
correspond to the angles made by the cut plane with the sagittal plane, the horizontal plane and
the coronal plane respectively.
Figure 7.3-2: Suggested cut plane containing MB and the maximum area of the right fornix column
connected to MB: a) ventral view, b) dorsal view, exposed plane (F = fornix, MB = mammillary bodies, H =
hippocampus)
7.4. Practical solution
In order to facilitate the precise sectioning at the desired orientation we can now suggest:
a) To construct two prisms from Plexiglas, of 25mm height. Both should have as bases right
angled triangles, and two sides, while perpendicular to the bases, should form between them
angles of 320, and 190 respectively for the second prism. These angles are derived taking into
account: a) the fact that the cut plane is perpendicular to the local normal and b) the position of
the brain as it is going to lye on the prisms.
b) To perform first a midsagittal cut, and for the right hemisphere (to which our results pertain)
another parasagittal cut (Figure 7.3-2 a).
Lay the right half of the brain on the parasagittal plane, the medial plane facing upwards as
shown in Figure 7.3-2 b.
Place the brain, oriented as described previously, on the first prism with angle equal to the
angle made by the cut plane with the ventral to dorsal axis, 320 in this case (see Figure 7.3-2 b).
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
129
The cut is made vertically, parallel to the edge of the prism whose direction coincides now with
the posterior to anterior axis. Using the second prism, placed upon the first, the sagittal plane is
rotated around the left to right axis by an angle equal to 900 minus the angle made by the cut
plane with the posterior to anterior axis (710), 190 in this case. The dorsal part of the brain is
made parallel with the hypotenuse of the prism, as shown in Figure 7.3-2 b. Several slices are
to be cut, as indicated by the direction of the dashed line (i) in Figure 7.3-2 b in order to find
the plane with maximum right fornix area, and passing through MB.
Figure 7.4-1.a) Initial cuts through the rat brain are made midsagittaly and parasagittaly (image mdifed
from [Paxinos, 1995]); b) The right half of the brain is laid on a prism whose angle is the angle the
sectioning plane makes with the ventral to dorsal axis (the midsagittal plane (M) facing upwards, and the
parasagittal plane downwards), and rotated around the left to right axis by placing the dorsal part of the
brain tangent to the hypotenuse of the prism whose angle is equal to 90o minus the angle the sectioning
plane makes with the anterior to posterior axis. The dashed line (i) indicates how the sectioning is done.
7.5. Discussion
An interactive software tool is proposed to find the plane of cut through the rat brain, which
would preserve as much as possible intact connections between the fornix fibers and the MB.
The interactivity is advantageous since there might be no unique solution to this problem, and
the solution needs to be practical for the experimenter.
Te sectioning plane coordinates are given relative to bregma but, since this landmark is not
visible anymore after extracting the brain from the skull, the results would be better computed
relative to anatomical landmarks visible on the brain surface.
a) b)
Chapter 7. A Software Tool for Interactive Determination of the Plane of Cut through the Rat
Brain
130
Insufficient data and unequally spaced images have been acknowledged as sources of
inaccuracy. The coronal slices are spaced at an average of 0.3 mm, leading to a poor resolution
on the anterior-posterior axis compared to the coronal plane. Small structures are therefore
likely to be misrepresented. Atlases often provide unequally spaced slices. This reduces the
accuracy, since one has to interpolate between existing slices thus introducing interpolation
errors.
7.6. Conclusions
The proposed software tool would be of help in electrophysiological studies of synaptic
relationships in brain slices maintained in vitro. Before a real experiment, the plane of cut
would be determined and illustrated using the proposed software tool.
Any rat brain atlas can be used to provide the set of two-dimensional data used to reconstruct
the brain volume. The limitations of a 2D viewing system are overcome by the use of software
tools. Planes of arbitrary orientations can be reconstructed, offering visual clues not evident in
classical two-dimensional atlases.
Selected rat brain structures can be analyzed with this method, if the preprocessing of the atlas
images is performed. This includes segmentation and cleaning of the background. A further
application would be to construct devices to support the rat brain in the desired orientation
[Dingledine et al, 1980], improving the precision of sectioning at the specified parameters.
Further development of the application will focus on automating the process and it is possible
to extend it to other structures.
Actual experiments on rat brain slices are planned in order to validate the approach. We also
plan to extend the use of the present software tool to visualization of combined histochemical
and receptor autoradiography data.
Chapter 8.General Discussion
131
Chapter 8. General Discussion
Chapter 8. General Discussion..................................................................................................131
8.1 Programs design ........................................................................................................132
8.2 Contributions of SAV to understanding brain function, complementing and
integrating the relevant techniques........................................................................................135
8.2.1 Source localization and extent...........................................................................135
8.2.2 Source separation ..............................................................................................135
8.2.3 Spatial resolution of electrophysiological techniques.......................................136
8.2.4 Comparison of multimodal data........................................................................136
8.2.5 Morphometric studies........................................................................................136
8.3 Applications ..............................................................................................................136
8.4 Outlook......................................................................................................................137
The problem of understanding brain function and its relation to the cortical topography has
challenged people with different backgrounds and various methods have been developed to
approach this question. The development of these methods was boosted by improvements in
techniques like EEG, MEG, fMRI, or PET. The use of increasingly powerful computers
reduced the time required for analysis of large amounts of data pertaining to brain activation.
Historically EEG was the first technique used for monitoring brain function, either invasively
or noninvasively using electrodes pasted onto the head. At a different level the electrical
potentials recorded from brain slices give useful information regarding the normal or
pathological functioning of specific circuits. In the minor part of this thesis we developed an
application (Chapter 7) designed for facilitating such an experiment on the rat brain by
providing an optimal plane of cut which would preserve as much as possible intact connections
from the fornix to the mammilary bodies. Even if the practical realization of this slices proved
to be difficult it was useful to construct a tool for visualizing the tortuous circuitry of the fornix
and the relationship to the mammilary bodies. A similar effort like the one we have put into
analyzing the circuitry from the hippocampus, via fornix to the mammilary bodies can be put
Chapter 8.General Discussion
132
into constructing a complete 3D rat brain. Eventually the same methodology we proposed for
the mammillary bodies can be applied for the study of other brain structures’ roles.
Besides time series analysis one needs visualization tools to be able to relate the time
development of activation to the cortical topography. In the major part of this thesis we present
methods for imaging and analyzing the human brain based on structural information from MRI
and functional data from MEG and EEG signal analysis. These methods are implemented into a
software package called SAV, its name being the acronym for Surface Activation Visualization.
SAV has been designed specifically, but not exclusively, for the display of vector fields like the
current density vector extracted by MFT analysis of MEG data, allowing the observation and
comparative evaluation of the space-time characteristics of activations in specified brain areas.
The following section discusses aspects related to the software design (Section 8.1) which
implements methods for brain structure segmentation and visualization of surface activation.
The choice of the methods and the relative advantages and disadvantages are discussed. Section
8.2 discusses how SAV can contribute to enhancing the understanding of brain function
through the use of information on the brain topography. Finally Section 8.3 discusses the
specific applications where we have used SAV.
8.1 Programs design
The present thesis introduces a collection of software programs (SAV) for analysis of brain
activation. SAV is intended to comply with three main requests: a) portability across platforms
and this was achieved by selecting IDL as the programming language; b) an object oriented
design and a modular structure to allow reusability of functions and the easy addition of new
modules; c) high interactivity so that a user can adjust the parameters of analysis and expose the
results in a way that provides significant information. A multitude of options for analysis exists
and they can be combined in various ways for a maximum flexibility.
Two of the software modules are dedicated to brain segmentation: (i) semi-automated
segmentation of the brain from MRI scans with CORTSEG and (ii) manual or semi-automated
segmentation of identifiable cortical or subcortical structures with STRUCTSEG.
The problem of brain segmentation remains challenging due to the complexity of the structure
of interest and the imperfect nature of images. Manual segmentation is extremely tedious, prone
to operator bias and has relatively low reproducibility. Still it is considered the gold standard in
most cases when the accuracy of a segmentation algorithm is evaluated. We have adopted an
Chapter 8.General Discussion
133
interactive approach, which enables to successfully segment the brain from MRI. The
combination with the editing program allows correction of eventual errors from the
semiautomatic segmentation. Also, the editing procedures available in STRUCTSEG could be
used for eliminating topological defects (handles or holes) in view of further processing of the
surface, for example for flattening.
The third SAV module is VISIO. This allows visualization of electrophysiological activation
data on the background of the anatomy provided by the other modules. We normally employ
the gray matter surface for visualization but also the white matter visualization has advantages
when the sources are buried within cortical folds. To extract the white matter surface (the
WM/GM interface) we use a fuzzy c means approach, modified so as to include neighborhood
information and compensate for inhomogeneity artifacts. The negative effects of the
inhomogeneity on the segmentation can also be alleviated by the use of anisotropic diffusion
filtering.
We chose a surface based approach since it works faster compared to a volume based approach.
The data enclosed by the brain boundary are not used for 3D rendered but are available for
inspection in slices cut through the brain.
SAV has been tested and used in two laboratories and its results have been compared with the
parallel 2D representation of activations produced with other software.
A couple of limitations should be acknowledged. First, it is in the nature of this method to
display just the surface representation and therefore relatively deeper lying sources (provided
they are strong enough to “reach” the surface) will project on a larger area, in a more diffuse
way, compared to equally strong but superficial sources. It is therefore acknowledged that the
accuracy is best for superficial sources, like those we demonstrate in the calcarine gyrus. On the
other hand, SI activations displayed on the surface of the 3D-rendered whole brain can be
misleading. Complementary approaches like the display on the segmented posterior wall of CS
or the display in orthogonal slices can resolve the ambiguities.
Depending on the nature of the deeper sources SAV affords the use of a set of complementary
approaches towards an accurate representation of their activity. Important for the performance
of these approaches are the functions of normalization, thresholding and color coding of
activity across different sets of data, along with the possibility to set the depth from which a
source can still project onto the surface, so that one can compare surface and deep activations.
Chapter 8.General Discussion
134
A second limitation is the fact that the surface activations are mapped onto a dense mesh, of
resolution comparable to this of the MRI and therefore greater than the one of functional data.
The interpolation causes smearing of the data. The amount of smearing can be controlled by
adjusting the depth parameter, which represents the maximum distance from the surface vertex
where from activity can be picked up. Different interpolation options allow for a sharper or
smoother point spread function.
Among the advantages offered by SAV we count the multitude and complementarity of
approaches allowing a more global view on the relationships – both biophysical and
physiological – between the displayed activations.
Activation data consisting in cortical current density maps can be mapped onto the surface
taking into consideration the local surface orientation or just the scalar properties. Also
statistical parametric maps can be displayed on the brain.
The current density vectors can be displayed as they are, or restricted to their projection along
the local normals. The latter approach carries potential in several directions: exploring the
biophysical basis of evoked potentials generators, deciding whether the displayed source is a
pickup from a remote strong source or is indeed generated in the analyzed region, and
differentiating sets of independent components of the response possibly linked to independent
generators etc.
The analysis of cortical activation data is helpful in identifying the multiple sources involved in
a specific task but requires a lengthy analysis. This can be complemented with the analysis
restricted to areas of interest, occasionally buried within sulci or lying deeper in identifiable
subcortical areas or even lying outside cerebrum (cerebellum, brain stem, etc.).
The anatomical information contained in SAV can be used to generate models for the conductor
and for restricting the solutions to the inverse problem.
SAV demonstrates the feasibility of a diverse range of display capabilities of structure and
function through the features and options implemented, while offering a high level interactivity
with the user. We believe that careful manipulation of the visualization and analysis parameters
may reveal information which otherwise will be lost if using standard, fixed parameters and
analyzing only the statistical maps.
Chapter 8.General Discussion
135
8.2 Contributions of SAV to understanding brain function,
complementing and integrating the relevant techniques
The main focus in SAV is on using information on the brain topography to enhance the
understanding of its function in several ways: a) source localization; b) separation of
concurrently, simultaneously activated sources (i.e. SI and SII); c) evaluation of the spatial
resolution of electrophysiological techniques; d) comparison of multimodal data, i.e. MEG,
EEG and fMRI .
8.2.1 Source localization and extent
The source localization problem can be addressed by examining the cortical current density or
the statistical maps in the background of the segmented brain or for better access to areas
hidden otherwise from the viewer in the background of a selected anatomical region. The two
separate modules dedicated to segmentation, CORTSEG and STRUCTSEG can be used for
best results alternatively (i.e. the ventricles can be segmented based on morphology, using
CORTSEG) or in combination (i.e. to edit the segmented brain using STRUCTSEG).
The anatomical information which can be extracted from SAV can be used: a) into the model
used for the inverse/forward problem; b) to incorporate constraints into the inverse problem.
The color coding helps identify the areas of high activation and thresholding enhances the
appreciation of the source location and extent. Superficial and focal sources are most accurately
represented using SAV in comparison with deeper sources which are more imprecisely
localized, and project their activity onto the cortex in a more diffuse way, spread over an area
increasingly larger.
8.2.2 Source separation
The separation of sources benefits from the use of high resolution representation of the
anatomy, as is the one obtained from MRI. Various interpolation functions help discriminating
between close sources. Two distinct activation sources can be separated based on information
on the source orientation. This is possible since the activation data can be mapped from the
volume onto the surface taking into consideration the scalar and also the orientation with
reference to the local surface normals.
Chapter 8.General Discussion
136
8.2.3 Spatial resolution of electrophysiological techniques
One can directly compare the location and extent of sources in the background of anatomy and
decide on the power of different techniques to discriminate between distinct sources.
8.2.4 Comparison of multimodal data
Besides the use of distributed or dipoles solutions for the MEG or EEG sources we imported
into SAV contours delimiting areas of statistically significant activations from fMRI studies.
These can be presented in the background of the segmented brain structures (i.e. central sulci or
the area around the calcarine fissure) and overlaid onto the MEG based activation maps. These
studies complement each other and give indications on the possibility of achieving high
resolution in localizing sources with MEG (see Chapter 7).
Additionally landmark points or structures derived from the anatomy can be generated within
SAV and kept in the background as a reference (the use of central sulcus outline or nerve exits
from the brain stem to point to the relevance of the activated areas).
8.2.5 Morphometric studies
Simple morphometric studies on the segmented brain structures are feasible using SAV and
allow qualitative but also quantitative comparison between normal and pathology affected
brains in terms of the total volumes or surface areas of the segmented structures.
8.3 Applications
The areas where we have applied SAV encompass both normal and pathological aspects of
brain functioning. We addressed the issue of normal brain processing of sensory information in:
a) somatosensory evoked fields/potentials elicited by median nerve and tibial nerve stimulation
and b) visual evoked fields. We used SAV to illustrate aspects of somatosensory processing in a
paraplegic patient.
It is essential to be able to use complementary approaches, as those offered by SAV. It is often
the case that the high values of the activations are hidden beneath the visible cortical surface. In
such cases a combination of displays on surfaces in all three of (a) 3D whole brain renderings,
(b) appropriately cut brain slices and (c) segmented specific brain structures may be needed in
order to expose the gyri, sulci or nuclei whose view may be otherwise partially obstructed and
Chapter 8.General Discussion
137
thus visualize the separate sources. We have shown this with the example of the SI and SII
activations.
Another application makes use of the surface description in terms of the local geometry and the
orientation of the patches.
SAV was found as potentially useful in studying questions of pathophysiology. Paraplegic
patients clinically evaluated as having suffered a complete spinal cord injury surprisingly do
display evoked potentials following stimulation of the body below the lesion, and the responses
reach statistical significance in the appropriate contralateral SI leg area [Ioannides et al.,
2000b]. The responses are however much weaker than normal and appear rather irregularly,
possibly representing post lesion plasticity mechanisms of the brain operating on scarce
remaining original or alternative afferents. Displaying the space-time development of stimulus
evoked activations with SAV offers therefore an opportunity for studying how human brain
copes with massive deafferentation of a specific cortical node and more specifically the extend
of reorganization and partial takeover by projections from healthy parts of the body
[Bruehlmeier et al., 1998]. Such knowledge is critical for developing new approaches to
rehabilitation of paraplegic patients. Specifically, SAV is well suited to help such studies by
providing comprehensible visualization of the ms by ms extend of activations evoked by
stimulating body parts below and above the lesion along with the statistics of these activations
and aid their comparison on the background of the gross brain anatomy which may or may not
be deformed. One trend is evidently amenable to scrutiny with SAV, that of more wide spread
representation of feet responses in the posterior to CS parietal cortex in the paraplegic patients.
To the extend that this trend has prognostic value [Green et al., 1998], SAV may help in
patients screening and progress follow-up.
8.4 Outlook
Most of the problems in source localization arise from the fact that there are usually several of
them contributing to the field of the same brain area in overlapping time periods and therefore
besides their identification and separation in space one is called to disentangle the biophysical
and physiological interactions of these sources. Different modeling, statistical, correlational and
other neuroinformatics methodologies can be employed for this goal as long as a common
framework of reference and correlation is used. The latter can be provided by visualization of
all results on the anatomy of the brain, and here SAV can be useful.
Chapter 8.General Discussion
138
One of the possible applications of an integrated environment, comprising segmentation,
functional data analysis and visualization is based on the observation that it is possible to
discriminate (based on a mask of the segmented brain) between which source points fall within
the brain volume and which fall outside. We have developed within SAV a software routine
that can perform the intersection between the segmented anatomy and the source space and we
hope this could help define the source space boundaries.
Further refinement of our segmentation approach, like correcting for topological errors would
allow us to proceed to producing flat mesh representations of the surface, exposing thus the
activity buried in the sulci.
Starting from the initially segmented structures one may study morphological differences across
normal and pathological cases in diseases like Alzheimer’s disease, schizophrenia, epilepsy,
paraplegia, etc., as well as in longitudinal studies. Right now a simple volumetric (or surface)
analysis is straightforward using SAV.
We have used a description in terms of curvatures of the extracted anatomy (see Chapter 3).
These descriptions might be useful for registration or shape analysis purposes, in comparing
healthy versus pathology affected brains.
The normalization across multiple activation data files can help comparing studies across brain
areas, time slices, experiments and subjects, if the data are properly transformed.
The information on the vector properties of functional data could help deciding whether the
displayed source is a pickup from a remote source or is indeed generated in the analyzed region.
A further direction of work would be to examine the effect of using finer grids with current
density sources. SAV may also be found useful in the study of some physiological questions
which involve space-time relationships, like that of surround inhibition in sensory projections
[Mountcastle and Powell, 1959; Welker et al., 1993].
Among the complementary approaches, that of visualizing the projection of currents along the
local normals of segmented surfaces is the most dynamic approach to many source localization
problems and has been a major impetus in developing the present software tool. It is based on
the idea that most MEG dipoles are created by currents running in the same direction as the
apical dendrites of cortical pyramidal neurons, i.e. normal to cortical surface [Ioannides 2001;
Michel et al., 2001; Hari 1999; Stern and Silbersweig, 2001]. Given the columnar organization
of the cortex, this may be a natural constraint on the cortical MEG sources. Indeed in the
presented example, the direction of the main current flow at the peak activation latency in SI
Chapter 8.General Discussion
139
(entering the cortex perpendicularly to its surface, as seen in Figure 7.3.2) is consistent with an
intracellular current flowing from top to bottom of apical dendrites of pyramidal neurons in
Brodmann area 3b (tangential dipole in Fig. 2.2.1) [Halgren, 1990]. This could result from
synchronous excitatory postsynaptic potentials in the upper end of the apical dendrites, close to
the cortical surface. Given the columnar organization of cerebral cortex, the currents tangential
to the whole brain surface are supposed to generate the strongest MEG dipoles in the paradigm
used [Hari, 1999]. It would be challenging to investigate whether the EEG source at the same
place and moment in time would be in the direction exiting the surface as suggested by its
attribution to the returning extracellular branch of the same circuit, i.e. a dipole with negative
pole near the cortical surface. To the extent that this proposed biophysical constraint is
vindicated in many other evoked potential examples, it may acquire significance as an a priori
constraint in MEG reverse problem solutions. SAV may thus be added to the tools, which are
currently trying to decipher the exact biophysical basis of MEG.
Chapter 9.Conclusion
140
Chapter 9. Conclusion
In the framework of this thesis several methods have been developed for processing anatomical data
in view of their use for functional studies on the brain.
Chronologically, the first application has been the use of anatomical images from a digital rat brain
atlas to reconstruct the brain and selected brain structures. This anatomical information was used to
help design an electrophysiological experiment on slices in vitro, under the request to select a cut
through the brain which would preserve as much as possible intact connections to a selected structure.
The main goals of the first part of the thesis are: a) to increase the appreciation of the anatomy of
specific brain structures, namely the mammilary bodies and the fornix, in relationship with the
hippocampus and the whole rat brain; b) to help planning an experiment in vitro on rat brain slices by
providing the optimal slicing plane which would keep intact most of the connections between the two
structures.
The next task has been to develop a method allowing the realistic visualization of brain electrical
activity i.e. display it on the surface of the brain areas actually sustaining the electric current, We
developed an object-oriented software tool, SAV, for visualization of spatio-temporal brain activity,
which allows the interplay of geometry and vector properties of the current density directly in the
representations. The main goal in SAV is to increase the understanding of the spatiotemporal
characteristics of brain current sources associated with distinct phenomena.
We investigate the use of methods based on mathematical morphology for cerebrum segmentation
and make use of interactivity to alleviate some of the problems in this type of approach (setting
appropriate thresholds, establishing the number of times the morphological operators are applied,
etc). The combination of segmentation methods dedicated in principle to different tasks like
segmenting the cerebrum or segmenting selected brain structures can improve the performance in
terms of quality for the semiautomatic method on one hand and also helps improving the speed in the
second case on the other hand.
We have developed software tools displays of activation maps in the background of the anatomy to
gain insight into the wealth of information on brain function that MFT analysis has unlocked from the
MEG data. Especially our tools allow evaluating regional and directional specificity of the
unconstrained solutions, as they allow quick and comfortable browsing through the data, including
generation of activation curves. The software’s design and capabilities are consistent with the
Chapter 9.Conclusion
141
philosophy to introduce the minimum necessary constraints in the inverse problem, and to contrast
the minimally constrained solutions to the anatomy [Ioannides 1989].
We have added a wealth of options into SAV for various types of visualization and analysis, while
leaving as much freedom to the program user to change and adjust the parameters of the analysis.
This principles aim to maximize the potential of the technique for understanding the complex cortical
structure- function relationship.
The problem of efficient analysis of large amounts of data produced with high resolution imaging and
functional techniques remains challenging while their interpretation is still a matter of both art and
science. An interdisciplinary effort is highly valuable in advancing our knowledge on the always
intriguing brain.
i
Publications related to the present thesis
• Badea A1., Kostopoulos G.K. and Ioannides A.A., “Surface Visualization of
Electromagnetic Brain Activity”, accepted for publication in the Journal of
Neuroscience Methods, 2003
Meeting presentations related to the present thesis
• Kostopoulos G.K., Fenwick P.B.C., Schellens M., Badea A., Zainea O. and Ioannides
A.A., “The relationship between cortical, cerebellum and brainstem activity from
average and single trial MEG and EEG data”, proceedings of the 32nd Annual Meeting
of the Society for Neuroscience, Orlando, Florida, November 2-7, 2002
• Badea A., Kostopoulos G.K, Ioannides A. A., “White–gray matter segmentation and
visualization of cortical activations”, proceedings of the 3-rd European Symposium on
Biomedical Engineering and Medical Physics, Patras, 30.08-1.09, 2002
• Zainea O., Badea A., Kostopoulos G.K., Ioannides A.A., “Identification of primary
somatosensory cortex subdivisions using anatomically constrained dipoles”,
proceedings of the 3-rd European Symposium on Biomedical Engineering and Medical
Physics, Patras, 30.08-1.09, 2002
• Bocioaca A., Ioannides A. A., Kostopoulos G. K., "A Semiautomatic Method for
Segmentation and Volumetry of the Hippocampus from MRI"; Proceedings of the 15th
Meeting of the Hellenic Society for Neuroscience, Patras, Greece, October 27-29, 2000.
• Bocioaca A., Schellens M., Kostopoulos G.K., Ioannides A. A., "Visualization of
distributed source solutions in electrophysiology: Examples from MEG"; proceedings
of 12th International Conference on Biomagnetism, Biomag2000, Helsinki University
of Technology, Espoo, Finland, 13 to 17 August , 2000. An oral presentation with a
similar name was held at the European Symposium for Biomedical Engineering, Patras
2001.
• Bocioaca A., Badea C., Papatheodoropoulos C., Kostopoulos G. K., "A Software Tool
to Assist the Planning of Electrophysiological Experiments on Rat Brain Slices",
1 Badea A. and Bocioaca A. are one and the same person.
ii
Medical Informatics Europe 1999 (MIE '99) congress, Ljubljana, Slovenia, 22-26
August 1999.
• Bocioaca A., Badea C., Kostopoulos G., Ioannides A., "A software for brain
segmentation from MRI data - a morphology based approach", Proceedings of the 14th
Meeting of the Hellenic Society for Neuroscience, pp.3-4, Volos, Greece, 14-16 May
1999; (paper awarded the second prize for best poster presentations ); presented then at
the 20th European Winter Conference on Brain Research, 11-17 March 2000, Villars-
sur-Ollon , Switzerland).
• Bocioaca A., Badea C., Kostopoulos G. K., "A Software Tool for Interactive
Determination of the Plane of Cut through the Rat Brain", Proceedings of the First
European Symposium in Biomedical Technology and Medical Physics, 28-29 August
1998, Patras, Greece.
iii
• Abbreviation list
3v: third ventricle
a: alveus
ac: ambient cistern
AFCM: adaptive fuzzy c means
AG: amygdala
au: arbitrary units
BA: Brodmann area
BS: brain stem
CA: cornu ammonis
ca: cerebral aqueduct
cc: crus cerebri
CT: computed tomography
cc:crus cerebri, cerebral penduncles
cf: crus fornices
CM=callosomarginal sulcus
CS: central sulcus, rolandic sulcus
CSF: cerebro spinal fluid
EC=enthorinal cortex
ECD: equivalent current dipole
ECG: electrocardiogram
EEG: electro encephalogram
EM: expectation maximization
EOG: electrooculogram
EPSP: excitatory postsynaptic potential
f1: the superior frontal sulcus
f2: the inferior frontal sulcus
FDA: federal drug administration
FCM : fuzzy c means
fMRI: functional magnetic resonance tomography
GB: globus palidus
GM: gray matter
H: hypothalaus
HH: hippocampus head
iv
ic: internal capsule
ica: independent component analysis
ips: intraparietal sulcus
IPSP: inhibitory postsynaptic potential
KS: Kolmogorov-Smirnov
LA: left arm
LF: left foot
LH: left hippocampus
LGN: lateral geniculate nucleus
lv: lateral ventricle
mb: mammillary bodies
MEG: magnetoencephalogram
MFT: magnetic field tomography
MGN: medial geniculate nucleus
ML: medial lemniscus
MR: magnetic resonance
MRI: magnetic resonance imaging
MM: mathematical morphology
MS: multiple sclerosis
MTS: medial temporal sclerosis
MUSIC: multiple signal classification
NMR: nuclear magnetic resonance
PreCG: precentral gyrus
PostCG: postcentral gyrus
RA: right arm
RF: right foot
rf: radio frequency
RH: right hippocampus
ROI: region of interest
Rol: rolandic sulcus (central sulcus)
TLE: temporal lobe epilepsy
t1: superior temporal
P: putamen
PD: proton density
v
PET: positron emission tomography
PSF: point spread function
PVA: partial volume effects
pu: pulvinar nucleus
RH: right hippocampus
SI: primary somatosensory cortex
SII: secondary somatosensory cortex
s: splenium of corpus calossum
SAV: surface activation vizualizer
sc: superior colliculi
sn: substantia nigra
SEF: somatosensory evoked fields
SEP: somatosensory evoked potetials
SNR: signal to noise ratio
SPECT: single photon emission computed tomography
SQUID: superconducting cuantum interference device
T1: inversion time
TE: echo time
th: temporal (inferior) horn of lateral ventricle
V1: primary visual cortex
V2: secondary visual cortex
VA: ventralis anterior (nucleus of thalamus)
ZI: zona incerta
VRML: virtual reality markup language
WM: white matter
References
vi
Abrahams S., Morris R. G., Polkey C. E., Jarosz J. M., Cox T. C. S., Graves M., Pickering A.,
“Hippocampal Involvement in Spatial and Working Memory: A Structural MRI Analysis
of Patients with Unilateral Mesial Temporal Lobe Sclerosis”, Brain and Cognition, Vol.
41, No. 1, Oct 1999, pp. 39-65
Agmon A., Connors B.W., “Thalamocortical responses of mouse somatosensory (barrel) cortex
in vitro”, Neuroscience, 41, No. 2/3, pp. 365-379, 1991.
Ahmed MN, Yamany SM, Mohamed N, Farag AA, Moriarty T. A modified fuzzy C-means
algorithm for bias field estimation and segmentation of MRI data. IEEE Trans Med
Imaging 2002 Mar;21(3):193-9
Amir A. Amini, Saeed Tehrani, and Terry Weymouth, "Minimizing the Energy of Active
Contours in the Presence of Hard Constraints," Proceedings of the Second International
Conference on Computer Vision, Tarpon Springs, Florida, 1988 or A. A. Amini et al.,
Using Dynamic Programming for Solving Variational Problems in Vision, IEEE
Transaction n PAMI, Vol. 12, No.2, 855-866, 1996.
Andreasen, N. C., Arndt, S., Swayze, V., Cizadlo, T., Flaum, M., O’Leary, D. S., Ehrhardt, J.
C., and Yuh, W. T. C. 1994. Thalamic abnormalities in schizophrenia visualized through
magnetic resonance image averaging. Science 266: 294–298.
Arndt S, Swayze V, Cizadlo T, O'Leary D, Cohen G, Yuh WT, Ehrhardt JC, Andreasen NC.
Evaluating and validating two methods for estimating brain structure volumes: tessellation
and simple pixel counting. Neuroimage. 1994 Jun;1(3):191-8.
Badea A, Kostopoulos GK, Ioannides AA, submitted 2002, Surface Visualization of
Electromagnetic Brain Activity. J. Neuroscience Methods.
Baillet S, Garnero L. A bayesian approach to introducing anatomo-functional priors in the
EEG/MEG inverse problem. IEEE Trans. Biomed. Eng., 1997; 44: 374-85.
Baillet S., Mosher J.C., Leahy R.M., "Electromagnetic Brain Mapping", IEEE Signal Processing
Magazine, Vol.18, No 6, pp. 14-30, November 2001
Barra V, Boire JY. Automatic segmentation of subcortical brain structures in MR images using
information fusion. IEEE Trans Med Imaging. 2001 Jul;20(7):549-58
Bear MF, Connors BW, Paradiso MA.1996. Neuroscience: Exploring the Brain Lippincott,
Williams & Wilkins 1996.
Berger, H. 1967. On the electroencephalogram of man (trans. by P. Gloor).EEG Clin.
Neurophysiol., Suppl. 28:1-350
Bocioaca A, Schellens M, Kostopoulos GK, Ioannides A A. Visualization of distributed source
solutions in electrophysiology: Examples from MEG. Proceedings of 12th International
References
vii
Conference on Biomagnetism, Biomag2000, Helsinki University of Technology, Espoo,
Finland, 2000.
Bocioaca A., Badea C., Papatheodoropoulos C., Kostopoulos G. K., "A Software Tool to Assist
the Planning of Electrophysiological Experiments on Rat Brain Slices", Medical
Informatics Europe 1999 (MIE '99) congress, Ljubljana, Slovenia, 22-26 August 1999.
Bocioaca A., Badea C., Kostopoulos G. K., "A Software Tool for Interactive Determination of
the Plane of Cut through the Rat Brain", Proceedings of the First European Symposium in
Biomedical Technology and Medical Physics, 28-29 August 1998, Patras, Greece.
Bowyer SM, Aurora KS, Moran JE, Tepley N, Welch KM. Magnetoencephalographic fields
from patients with spontaneous and induced migraine aura. Ann Neurol. 2001
Nov;50(5):582-7.
Brodal P, “The central nervous system”, Oxford Univerisy Press, 1992
Bruehlmeier M, Dietz V, Leenders KL, Roelcke U, Missimer J, Curt A. How does the human
brain deal with a spinal cord injury? Eur J Neurosci. 1998 Dec;10(12):3918-22.
Bushberg JT, Leidholdt EM, Boone JM, Seibert JA. The Essential Physics of Medical Imaging,
Williams & Wilkins, Baltimore, 1994
Cho ZH, Jones JP, Singh M. 1993. Foundations of Medical Imaging. John Wiley and Sons, Inc.
New York.
Chung R and Ho C, 2000. 3-D reconstruction from tomographic data using 2D active contours.
Computers and Biomedical Research, 33, pp 186-210
Clarke CJS, Janday BS. Probabilistic methods in a biomagnetic inverse problem. Inverse
Problem, 1989; 5: 483-500.
Cohen D., Cohen I, 1990 “A finite element method applied to the new active contour models
and 3D reconstruction from cross sections”, Proc. 3rd Int. Conf. Comput. Vision, IEEE
(or: “Using a Finite Element Method for Active Contour Models and 3-D Reconstruction
from Cross Sections”. In: Artificial Intelligence and Computer Vision, Editors: Y.A.
Feldman and A. Bruckstein. Elsevier Science 1991; 237–247.
Cohen L. D., Cohen I.. 1993. Finite-Element Methods for Active Contour Models and Balloons
for 2D and 3D Images. IEEE Transactions on Pattern Analysis and Machine Intelligence,
15(11):1131-1147, 1993.
Csernansky JG, Joshi S, Wang L, Haller JW, Gado M, Miller JP, Grenander U, Miller MI.
Hippocampal morphometry in schizophrenia by high dimensional brain mapping. Proc
Natl Acad Sci U S A. 1998 Sep 15;95(19):11406-11.
References
viii
Collins D. L. and Evans A. C., 1997 ``Animal: validation and applications of non-linear
registration-based segmentation,'' International Journal and Pattern Recognition and
Artificial Intelligence, vol. 11, pp. 1271-1294, Dec 1997.
Convit A, McHugh P, Wolf OT, de Leon MJ, Bobinski M, De Santi S, Roche A, Tsui W. MRI
volume of the amygdala: a reliable method allowing separation from the hippocampal
formation. Psychiatry Res. 1999 Apr 26;90(2):113-23.
Cootes T.F., Hill A.,Taylor C.J, Haslam J., 1994. "The Use of Active Shape Models for
Locating Structures in Medical Images." Image and Vision Computing Vol.12, No.6,
pp.355-366.
Crosson, B., and Hughes, C. W. 1987. Role of the thalamus in language: Is it related to
schizophrenic thought disorder? Schizophrenia Bull. 13: 605–621.
Chung MK, Worsley KJ, Paus T, Cherif C, Collins DL, Giedd JN, Rapoport JL, Evans AC. A
unified statistical approach to deformation-based morphometry.Neuroimage. 2001
Sep;14(3):595-606.
Cohen D. Magnetoencephalography: detection of the brain's electrical activity with a
superconducting magnetometer. Science 1972;175:664-66.
Cohen D. Magnetoencephalography. [review]. In: Encyclopedia of Neuroscience, Adelman G,
ed. Elsevier, vol.II, 1999: 1079-83
Collins DL, Zijdenbos AP, Kollokian V, Sled JG, Kabani NJ, Holmes CJ, Evans AC. Design
and construction of a realistic digital brain phantom. IEEE Trans Med Imaging. 1998
Jun;17(3):463-8.
Collins DL, Peters TM, Dai W, Evans AC. Model based segmentation of individual brain
structures from MRI data. In: Robb, R.A. (Ed.): Visualization in Biomedical Computing
II, Proc. SPIE 1808. Chapel Hill, NC, 1992:10-23.
Crouzeix, A., Yvert, B., Bertrand, O., and Pernier, J. 1999. An evaluation of dipole
reconstruction accuracy with spherical and realistic head models in MEG. Clin.
Neurophysiol. 110(12): 2176–2188.
Dale, A. M., and Sereno, M. I. 1993. Improved localization of cortical activity by combining
EEG and MEG with MRI cortical surface reconstruction: A linear approach. J. Cogn.
Neurosci. 5(2): 162–176.
Dale, A. M., Fischl, B., and Sereno, M. I. 1999. Cortical surface-based analysis I: Segmentation
and surface reconstruction NeuroImage 9: 179–194.
Dallas WJ. Space solution to the magnetostatic imaging problem. 1985. Applied Optics; 24:
4543-4546.
References
ix
Damasio, H. & Frank, R. 1992. Three-dimensional in vivo mapping of brain lesions in humans.
Arch. Neurol. 49, 137-143.
Davatzikos C. Spatial transformation and registration of brain images using elastically
deformable models. Comput Vis Image Underst. 1997 May;66(2):207-22.
David O, Garnero L . Time-coherent expansion of MEG/EEG cortical sources. Neuroimage.
2002 Nov;17(3):1277-89.
Dawant BM, Hartmann SL, Thirion JP, Maes F, Vandermeulen D, Demaerel P., Automatic 3-D
segmentation of internal structures of the head in MR images using a combination of
similarity and free-form transformations: Part I, Methodology and validation on normal
subjects.IEEE Trans Med Imaging. 1999 Oct;18(10):909-16.
Dhawan AP, Arata L. 1992. Knowledge-based multi-modality three-dimensional image analysis
of the brain Am. J. of Physiol. Imag. 7(3-4):210-219
Dimitrov LI. Texturing 3D-reconstructions of the human brain with EEG-activity maps.Hum
Brain Mapp. 1998;6(4):189-202.
Dingledine R, Dodd J., Kelly J.S., “The in vitro brain slice as a useful neurophysiological
preparation for intracellular recording”. J. Neurosci. Methods 1980 Aug;2 (4):323-362.
Di Russo F, Martinez A, Sereno MI, Pitzalis S, Hillyard SA. Cortical sources of the early
components of the visual evoked potential. Hum. Brain. Mapp., 2002;15: 95-111.
Disbrow E, Roberts T, Poeppel D, Krubitzer L. Evidence for interhemispheric processing of
inputs from the hands in human S2 and PV. J Neurophysiol. 2001 May;85(5):2236-44.
Duchesne S, Pruessner J, Collins D. Appearance-based segmentation of medial temporal lobe
structures. Neuroimage. 2002 Oct;17(2):515.
Elbert T, Sterr A, Flor H, Rockstroh B, Knecht S, Pantev C, Wienbruch C, Taub E. (1997).
Input-increase and input-decrease types of cortical reorganization after upper extremity
amputation in humans. Exp Brain Res. 1997 Oct;117(1):161-4.
Engel SA, Glover GH, Wandell BA.1997. Retionotopic organization in human Visual Cortex
and the Spatial precision of functional MRI. Cereb Cortex 7: 181-192
Essen DC Van., Drury H. A. 1997. Structural and functional analyses of human cerebral cortex
using a surface-based atlas. The Journal of Neuroscience, 17: 7079–7102.
Essen DC Van, Drury HA, Joshi S, Miller MI. 1998. Functional and structural mapping of
human cerebral cortex: solutions are in the surfaces.Proc Natl Acad Sci U S A. 1998 Feb
3;95(3):788-95
References
x
Essen DC Van, Drury HA, Dickson J, Harwell J, Hanlon D, Anderson CH. 2001. An integrated
software suite for surface-based analyses of cerebral cortex.J Am Med Inform Assoc. Sep-
Oct;8(5):443-59.
Foerster, O. (1936). In Handbook der Neurologie, vol. 6, ed. Bumke, O. & Foerster O.,
Springer, Berlin pp 50-51.
Evans A. C. Collins D. L Milner B., "An MRI-based stereotactic atlas from 250 young normal
subjects", Journal Soc. Neurosci. Abstr. 18: 408, 1992
Evans AC, Collins DL , Mills SR, Brown ED, Kelly RL, Peters TM. 3D statistical
neuroanatomical models from 305 MRI volumes. Proc. IEEE-Nuclear Science Symposium
and Medical Imaging Conference, 1813-1817, 1993.
Faugeras O, Clément F. Deriche R., Keriven R., Papadopoulo T, Roberts J., Viéville T,
Devernay F., Gomes J., Hermosillo G., Kornprobst P, Lingrand D. The inverse EEG and
MEG problems: The adjoint state approach . I: The continuous case. Research Report, N°
3673, May 1999
Felleman DJ, Van Essen DC. (1991). Distributed hierarchical processing in the primate cerebral
cortex. Cereb Cortex. 1991 Jan-Feb;1(1):1-47.
Filippi M, Mastronardo G, Rocca MA, Pereira C, Comi G. Quantitative volumetric analysis of
brain magnetic resonance imaging from patients with multiple sclerosis.J Neurol Sci. 1998
Jun 30;158(2):148-53.
Fischl, B., Sereno, M. I., Tootell RBH, Dale A. M. 1999 a. High Resoltion Intersubject
Averaging and a Coordinate System for the Cortical Surface, Human Brain mapping,
8:272-284
Fischl, B., Sereno, M. I., and Dale, A. M. 1999 b. Cortical surface-based analysis II: Inflation,
flattening, a surface-based coordinate system. NeuroImage 9: 195–207.
Frackowiak R, Friston K. J., Frith C. D., Dolan R., Mazziotta J. C. editors. 1997, Human Brain
Function, Academic Press, chapter 10 Maps of somatosensory cortex.
Fuster, J. M. 1997. The Prefrontal Cortex: Anatomy, Physiology,and Neuropsychology of the
Frontal Lobe, 3rd ed. Lippincott–Raven, Philadelphia.
Gallen CC, Sobel DF, Waltz T, Aung M, Copeland B, Schwartz BJ, Hirschkoff EC, Bloom FE.
Noninvasive presurgical neuromagnetic mapping of somatosensory cortex. Neurosurgery. 1993
Aug; 33(2):260-8
Gelnar PA, Krauss BR, Szeverenyi NM, And Apkarian AV. Fingertip representation in the
human somatosensory cortex: an fMRI study. Neuro-image7: 261–283, 1998
References
xi
George JS, Aine CJ, Mosher JC, Schmidt DM, Ranken DM, Schlitt HA, Wood CC, Lewine JD,
Sanders JA, Belliveau JW. Mapping function in the human brain with
magnetoencephalography, anatomical magnetic resonance imaging, and functional
magnetic resonance imaging. J. Clin. Neurophysiol., 1995; 12:406-31.
George MS, Scott T, Kellner CH, Malcolm R. Abnormalities of the septum pellucidum in
schizophrenia: Two case reports and a discussion. J Neuropsychiatry Clin Neuro 1989;
1:385-390
Ghanei, H. Soltanian-Zadeh, J. P. Windham, “Segmentation of the hippocampus from brain
MRI using deformable contours”, Comp. Med. Imag. Graph., 22, 1998, 203-216
Giedd JN, Blumenthal J, Jeffries NO, Castellanos FX, Liu H, Zijdenbos A, Paus T, Evans AC,
Rapoport JL. Brain development during childhood and adolescence: a longitudinal MRI
study. Nat Neurosci. 1999 Oct;2(10):861-3.
Gross J, Ioannides AA. Linear transformations of data space in MEG. Phys. Med. Biol., 1999;
44:2081-97.
Gross J, Kujala J, Hamalainen M, Timmermann L, Schnitzler A, Salmelin R. Dynamic imaging
of coherent sources. 2001: Studying neural interactions in the human brain. Proc Natl
Acad Sci U S A. 16;98(2):694-9.
Goldszal AF, Davatzikos C, Pham DL, Yan MX, Bryan RN, Resnick SM. An image-processing
system for qualitative and quantitative volumetric analysis of brain images. J Comput
Assist Tomogr. 1998 Sep-Oct;22(5):827-37.
Golland P, Grimson EL, Shenton ME, Kikinis R. 2000. Small Sample Size Learning for Shape
Analysis of Anatomical Structures. In Proc. MICCAI'2000, LNCS 1935, pp. 72-82
Good CD, Johnsrude IS, Ashburner J, Henson RN, Friston KJ, Frackowiak RS. A voxel-based
morphometric study of ageing in 465 normal adult human brains. Neuroimage. 2001
Jul;14(1 Pt 1):21-36.
Gunn S. R., Nixon M. S.. A Dual Active Contour. BMVC 94, September, York, U.K, 305-314,
1994.
Gur, R. E., Maany, V., Mozley, P. D., Swanson, C., Bilker, W., and Gur, R. C. 1998.
Subcortical MRI volumes in neuroleptic-naı¨ve and treated patients with schizophrenia.
Am. J. Psychiatry 155: 1711–1717.
Halgren E. 1990. Human Evoked Potentials. In Neurophysiological Techniques, Applications to
Neural Systems. Neuromethods 15 , Eds.: Alan A Boulton, Glen B Baker, Case H.
Vanderwolf, Humana Press, Clifton, New Jersey
References
xii
Haller JW, Christensen GE, Joshi S, Miller MI, Vannier M W., "Digital Atlas Based
Segmentation of the Hippocampus". Computer Assisted Radiology: Proceedings of the
International Symposium on Computer and Communication Systems for Image Guide
Diagnosis and Therapy, 1995.
Haller JW, Banerjee A, Christensen GE, Gado M, Joshi S, Miller MI, Sheline Y, Vannier MW,
Csernansky JG. 1997. Three-dimensional hippocampal MR morphometry with high-
dimensional transformation of a neuroanatomic atlas. Radiology Feb;202(2):504-10
Hämäläinen MS, Ilmoniemi RJ. 1984. Interpreting measured magnetic fields of the brain:
Estimates of current distributions. Helsinki Univ. Technol., Helsinki, Finland, Tech. Rep.
TKK-F-A559.
Hämäläinen M, Hari R, Ilmoniemi, RJ, Knuutila J, Lounasmaa OV. 1993.
Magnetoencephalography—theory, instrumentation, and applications to noninvasive
studies of the working human brain. Rev. Mod. Phys. 65, 413–497
Hämäläinen M.S. and Ilmoniemi R.J.. 1994. Interpreting magnetic fields of the brain:
Minimum-norm estimates, Med. Biol. Eng. Comp., 32: 35 - 42, 1994.
Hari R, Hamalainen H, Hamalainen M, Kekoni J, Sams M, Tiihonen J. 1990. Separate finger
representations at the human second somatosensory cortex. Neuroscience. 1990;37(1):245-
9.
Hari R, Karhu J, Hamalainen M, Knuutila J, Salonen O, Sams M, Vilkman V. 1993. Functional
organization of the human first and second somatosensory cortices: a neuromagnetic study.
Eur J Neurosci. 1993 Jun 1;5(6):724-34.
Hari R, Forss N. 1999. Magnetoencephalography in the study of human somatosensory cortical
processing. Philos Trans R Soc Lond B Biol Sci. 1999 Jul 29;354(1387):1145-54.
Helmholtz H. 1853. Über einige gesetze der vertheilung elektrischer strome in körperlichen
leitern, mit anwendung auf die tierisch-elektrischen versuche. Ann. Phys. Chem. 1853,
89:211-233, 353-377. (Some laws about the distribution electrical currents in volume
conductors, with application to animal electric experiments)
Hendry SH, Hsiao SS, Brown MC. 1999. Fundamentals of sensory system. In : Zigmond Mj,
Bloom FE, Landis SC, Roberts JL, Squire LR, eds. Fundamental Neuroscience. New
York. Academic Press. pp 657-670.
Höhne KH, Bomans M, Riemer M, Schubert R, Tiede U, Lierse W. A 3D anatomical atlas based
on a volume model. IEEE Comput. Graphics Appl. 1992(a);12:72-78.
Hohne K.H., Hanson W.A.1992b. Interactive 3-D segmentation of MRI and CT volumes using
morphological operations. J. Comp. Assist. Tom., 16:285–294, 1992.
References
xiii
Hornak J.P., 2002 . The Basics of MRI, a hypertext book on magnetic resonance imaging.
www.cis.rit.edu/htbooks/mri
Hui F, Cavazos JE, Tien RD. 1997. Hippocampus: Normal Magnetic Resonance Imaging
Anatomy with Volumetric Studies. Neuroimaging Clin N Am 1997;7:11–30 (also in
Imaging the Limbic System: Clinical Implications. Neurology Network Commentary,
2:92-110, 1998
Hurdal M. K., Bowers P. L., Stephenson K., Sumners D. W. L., Rehm K., Schaper K.,
Rottenberg D. A., 1999. Quasi-conformally flat mapping the human cerebellum, in C.
Taylor and A. Colchester (eds), MICCAI'99, Vol. 1679 of Lecture Notes in Computer
Science, Springer, Berlin, pp. 279-286. (available online at
http://web.math.fsu.edu/~mhurdal/papers/index.html#miccai99).
Ioannides AA, Bolton JPR, Hasson R, Clarke CJS. Localised and distributed source solutions
for the biomagnetic inverse problem II. In: S. J. Williamson et al., editors. Advances in
Biomagnetism. Plenum Press: New York, 1989; 591-94.
Ioannides AA, Bolton JPR, Clarke CJS. 1990. Continuous probabilistic solutions to the
biomagnetic inverse problem. Inverse Problems, 1990; 6: 523-42.
Ioannides, A. A., Singh, K. D., Hasson, R., Baumann, S. B., Rogers, R. L., Guinto, F. C.,
Papanicolaou A. C., Brain Topography 6(1), 27-34, 1993.
Ioannides AA, Real Time Human Brain Function: Observations and Inferences from Single
Trial Analysis of Magnetoencephalographic Signals, Clinical EEG 32:98-111, 2001.
Ioannides AA, Kostopoulos GK, Laskaris NA, Liu L, Shibata T, Schellens M, Poghosyan V,
Khurshudyan A. (2002a). Timing and connectivity in the human somatosensory cortex
from single trial mass electrical activity. Hum Brain Mapp. 2002 Apr;15(4):231-46.
Ioannides AA, Liu L, Khurshudyan A, Bodley R, Poghosyan V, Shibata T, Dammers J, Jamous
A. (2002b) Brain activation sequences following electrical limb stimulation of normal and
paraplegic subjects. Neuroimage. 2002 Jul;16(1):115-29.
Iosifescu DV, Shenton ME, Warfield SK, Kikinis R, Dengler J, Jolesz FA, McCarley RW. 1997.
An automated registration algorithm for measuring MRI subcortical brain structures.
Neuroimage. 1997 Jul;6(1):13-25.
Ishii R, Shinosaki K, Ikejiri Y, Ukai S, Yamashita K, Iwase M, Mizuno-Matsumoto Y, Inouye
T, Yoshimine T, Hirabuki N, Robinson S.E., Takeda M.: Theta rhythm increases in left
superior temporal cortex during auditory hallucinations in schizophrenia: a case report.
NeuroReport 2000; 11(14); 3283-3287.
References
xiv
Iwasaki N, Hamano K, Okada Y, Horigome Y, Nakayama J, Takeya T, Takita H, Nose T.
Volumetric quantification of brain development using MRI. Neuroradiology. 1997;
39(12):841-6.
Jack CR Jr, Petersen RC, Xu YC, Waring SC, O'Brien PC, Tangalos EG, Smith GE, Ivnik RJ,
Kokmen E. Medial temporal atrophy on MRI in normal aging and very mild Alzheimer's
disease. Neurology. 1997 Sep;49(3):786-94.
Jernigan, T. L., Archibald, S. L., Berhow, M. T., Sowell, E. R., Foster,D. S., and Hesselink, J. R.
1991. Cerebral structure on MRI, Part I: Localization of age-related changes. Biol.
Psychiatry 29: 55–67.
Johansen-Berg H. 2001 Reorganisation and modulation of the human sensorimotor sytem:
Implications for recovery of function after stroke. DPhil Thesis. Univ of Oxford. Available
online at http://www.fmrib.ox.ac.uk/~heidi/thesis.
Jones, E. G. 1985. The Thalamus. Plenum, New York.
Joshi, S. C.,Wang, J., Miller, M. I., Van Essen, D., and Grenander, U.1995. On the differential
geometry of the cortical surface. In Proceedings of SPIE’s 1995 Geometric Methods in
Applied Imaging,San Diego, CA, July 9–14, 1995.
Kass M, Witkin A, Terzopolous D. 1987. `Snakes: Active contour models'. In Proc. First
International Conference on Computer Vision, pp. 259--268. IEEE Computer Society
Press.
Kaas JH, 1983. What, if anything, is SI? Organization of first somatosensory area of cortex.
Physiol Rev 63: 206–231, 1983
Kakigi R, Hoshiyama M, Shimojo M, Naka D, Yamasaki H, Watanabe S, Xiang J, Maeda K,
Lam K, Itomi K, Nakamura A. The somatosensory evoked magnetic fields.Prog
Neurobiol. 2000 Aug;61(5):495-523.
Kandel E. R., Schwartz JH, Jessell TM. (Eds.). 1995. Essentials of neural science and behavior.
Appleton and Lange, Stamford.
Kapur T, Grimson WE, Wells WM 3rd, Kikinis R. Segmentation of brain tissue from magnetic
resonance images.Med Image Anal. 1996 Jun;1(2):109-27.
Karhu J, Tesche CD. Simultaneous early processing of sensory input in human primary (SI) and
secondary (SII) somatosensory cortices.J Neurophysiol. 1999 May;81(5):2017-25.
Kelemen A, Szekely G, Gerig G. Elastic model-based segmentation of 3-D neuroradiological
datasets. IEEE Trans Med Imaging. 1999 Oct;18(10):828-39
Kikinis R, Shenton ME, Iosifescu DV, McCarley RW, Saiviroonporn P, Hokama HH, Robatino
A, Metcalf D, Wible CG, Portas CM, Donnino RM, Jolesz FA. 1996. A digital brain atlas
References
xv
for surgical planning, model driven segmentation, and teaching. IEEE: Visualization and
Computer Graphics 2(3): 232-241.
Kim, J.S. Lee, Y.K. Kim, H.Y. Lee, S.K. Lee, D.S. Lee. Voxel-Based Morphometry of
Temporal Lobe Epilepsy with Hippocampal SclerosisProceedings of the 8th International
Conference on Functional Mapping of the Human Brain June 2 - 6, 2002, Sendai, JAPAN,
available online at http://www.apnet.com/www/journal/hbm2002/15270.html#15270
Kincses WE, Braun C, Kaiser S, Elbert T. Modeling extended sources of event-related potentials
using anatomical and physiological constraints.Hum Brain Mapp. 1999;8(4):182-93.
Kincses, W. E., Braun, C., Kaiser, S., and Mathiak, K. 2000. Modeling extended sources using a
maximum likelihood estimator. In Biomag2000 12th Int. Conf. on Biomagnetism,
Helsinki, Finland.
Korvenoja A, Huttunen J, Salli E, Pohjonen H, Martinkauppi S, Palva JM, Lauronen L, Virtanen
J, Ilmoniemi RJ, Aronen HJ.Activation of multiple cortical areas in response to
somatosensory stimulation: combined magnetoencephalographic and functional magnetic
resonance imaging. Hum Brain Mapp. 1999; 8(1):13-27.
Kostopoulos G.K. and Phillis, J.W., “Mamillothalamic neurons activated antidromically and by
stimulation of the fornix”, Brain Research, 122 (1977) 143-149.
Kriegeskorte N, Goebel R. An efficient algorithm for topologically correct segmentation of the
cortical sheet in anatomical mr volumes. Neuroimage. 2001 Aug;14(2):329-46.
Kullmann W, Dallas WJ. Fourier imaging of electrical currents in the human brain from their
magnetic fields. IEEE Trans Biomed Eng., 1987; 34: 837-42.
Laaks M.P., Vaurio O., Savolainen L, Repo E., Soininen H, Aronen H.J., 2000. “A volumetric
study of the hippocampus in type I, II of alcoholics”, Behavioural Brain Research 109,
177-186
Lawrie SM, Abukmeil SS. Brain abnormality in schizophrenia. A systematic and quantitative
review of volumetric magnetic resonance imaging studies. Br J Psychiatry. 1998
Feb;172:110-20.
Lawrie, S. M., Whalley, H., Kestelman, J. N., Abukmeil, S. S., Byrne, M., Hodges, A.,
Rimmington, J. E., Best, J. J. K., Owens, D. G. C.,and Johnstone, E. C. 1999. Magnetic
resonance imaging of brain inpeople at high risk of developing schizophrenia. Lancet 353:
30–33.
Lawrie SM, Whalley HC, Abukmeil SS, Kestelman JN, Miller P, Best JJ, Owens DG, Johnstone
EC. 2002. Temporal lobe volume changes in people at high risk of schizophrenia with
psychotic symptoms, Br J Psychiatry. Aug;181:138-43
References
xvi
Leahy RM, Mosher JC, Spencer ME, Huang MX, Lewine JD (1998): A study of dipole
localization accuracy for MEG and EEG using a human skull phantom.
Electroencephalogr Clin Neurophysiol 107:159–173.
Lehericy S, Baulac M, Chiras J, Pierot L, Martin N, Pillon B, Deweer B, Dubois B, Marsault C.
1994. Amygdalohippocampal MR volume measurements in the early stages of Alzheimer
disease. AJNR Am J Neuroradiol. May;15(5):929-37.
Leemput K Van, Maes F, Vandermeulen D, Suetens P. Automated model-based bias field
correction of MR images of the brain. IEEE Trans Med Imaging. 1999 Oct;18(10):885-96.
Liu AK, Belliveau JW, Dale AM. 1998. Spatiotemporal imaging of human brain activity using
functional MRI constrained magnetoencephalography data: Monte Carlo simulations.
Proc. Natl. Acad. Sci. USA. Vol. 95. pp. 8945-8950.
Llinas R.R., Ribary U., Jeanmonod D., Kronberg E., Mitra P.P., “Thalamocortical dysrhythmia:
A neurological and neuropsychiatric syndrome characterized by
magnetoencephalography,” Proc. Natl. Acad. Sci. USA, vol. 96, pp. 15222-15227, 1999.
Llinas R. I for the Vortex: From Neurons to Self, A Bradford Book. 2001
Maeda K, Kakigi R, Hoshiyama M, Koyama S. 1999. Topography of the secondary
somatosensory cortex in humans: a magnetoencephalo-graphic study. Neuroreport.
5;10(2):301-6.
MacDonald D, Kabani N, Avis D, Evans AC. Automated 3-D extraction of inner and outer
surfaces of cerebral cortex from MRI. Neuroimage. 2000 Sep;12(3):340-56.
Malmivuo J, Plonsey R, 1995, Biolecetromagnetism. Principles and Appiactions of Bioloectric
and Biomagnetic fields. Oxford Univeristy Press. New York.
Mazziotta J. C and Toga A. W.. Evans A and Fox P.and Lancaster J., "A Probablistic Atlas of
the Human Brain: Theory and Rationale for Its Development", NeuroImage 2:89-101,
1995
Mazziotta J, Toga A, Evans A, Fox P, Lancaster J, Zilles K, Woods R, Paus T, Simpson G, Pike
B, Holmes C, Collins L, Thompson P, MacDonald D, Iacoboni M, Schormann T, Amunts
K, Palomero-Gallagher N, Geyer S, Parsons L, Narr K, Kabani N, Le Goualher G,
Boomsma D, Cannon T, Kawashima R, Mazoyer B. A probabilistic atlas and reference
system for the human brain: International Consortium for Brain Mapping (ICBM). Philos
Trans R Soc Lond B Biol Sci 2001a Aug 29;356(1412):1293-322
Mazziotta J, Toga A, Evans A, Fox P, Lancaster J, Zilles K, Woods R, Paus T, Simpson G, Pike
B, Holmes C, Collins L, Thompson P, MacDonald D, Iacoboni M, Schormann T, Amunts
K, Palomero-Gallagher N, Geyer S, Parsons L, Narr K, Kabani N, Le Goualher G, Feidler
References
xvii
J, Smith K, Boomsma D, Hulshoff Pol H, Cannon T, Kawashima R, Mazoyer B. A four-
dimensional probabilistic atlas of the human brain.J Am Med Inform Assoc. 2001b Sep-
Oct;8(5):401-30.
McCarley RW, Wible CG, Frumin M, Hirayasu Y, Levitt JJ, Fischer IA, Shenton ME. MRI
anatomy of schizophrenia. Biol Psychiatry. 1999 May 1;45(9):1099-119.
Michel CM, Thut G, Morand S, Khateb A, Pegna AJ, Grave de Peralta R, Gonzalez S, Seeck M,
Landis T. Electric source imaging of human brain functions. Brain Res Brain Res Rev.
2001 Oct;36(2-3):108-18.
Miller DH, Barkhof F, Frank JA, Parker GJ, Thompson AJ. Measurement of atrophy in multiple
sclerosis: pathological basis, methodological aspects and clinical relevance. Brain. 2002
Aug;125(Pt 8):1676-95.
Mima T, Ikeda A, Nagamine T, Yazawa S, Kunieda T, Mikuni N, Taki W, Kimura J, Shibasaki
H. Human second somatosensory area: subdural and magnetoencephalographic recording
of somatosensory evoked responses.J Neurol Neurosurg Psychiatry. 1997 Oct;63(4):501-5.
Mitra PP and Pesaran B. 1999. Analysis of Dynamic Brain Imaging Data; Biophysical Journal
Volume 76 February. 1999. pp 691–708
Momenan R, Hommer D, Rawlings R, Ruttimann U, Kerich M, Rio D: Intensity adaptive
segmentation of single echo T1-weighted magnetic resonance images. Human Brain
Mapping 1997;5:194-205.
Mosher J.C., Lewis P.S., Leahy R., Multiple dipole modeling and localization from spatio-
temporal MEG data, IEEE Trans. Biomed. Eng., 39: 541 - 557, 1992.
Mogilner A, Grossman JA, Ribary U, Joliot M, Volkmann J, Rapaport D, Beasley RW, Llinas
RR. Somatosensory cortical plasticity in adult humans revealed by
magnetoencephalography. Proc Natl Acad Sci U S A. 1993 Apr 15;90(8):3593-7.
Moradi F., Liu L. C., Cheng K., Waggoner R. A., Tanaka K. and Ioannides A. A. 2003.
Consistent and precise localization of brain activity in human primary visual cortex by
MEG and fMRI. Neuroimage. 2003. March 18 (3): 595:609.
Mori E, Yoneda Y, Yamashita H, Hirono N, Ikeda M, Yamadori A. Medial temporal structures
relate to memory impairment in Alzheimer's disease: an MRI volumetric study. J Neurol
Neurosurg Psychiatry. 1997 Aug;63(2):214-21.
Mori E, Ikeda M, Hirono N, Kitagaki H, Imamura T, Shimomura T. Amygdalar volume and
emotional memory in Alzheimer's disease. Am J Psychiatry. 1999 Feb;156(2):216-22.
Mosher JC, Leahy RC, Shattuck DW, Baillet S MEG Source Imaging Using Multipolar
Expansions. Proceedings of the 16th Conference on Information Processing in Medical
References
xviii
Imaging, IPMI'99, Visegrád, Hungary, June/July 1999, Springer Series in Computer
Science, pp.15-28, A. Kuba, Sámal M., Todd-Pokropek A. (Eds.): June/July 1999.
(available online at http://neuroimage.usc.edu/publics.html)
Mountcastle VB. 1998. Perceptual neuroscience: The cerebral cortex.: Harvard University Press.
Cambridge, MA
Munck J.C. de. 1990. The estimation of time varying dipoles on the basis of VEPs
Electroenceph. Clin. Neurophysiol. 77:156-160
Nunez P.L., Electric Fields of the Brain. New York: Oxford, 1981.
Nunez P.L. Silberstein R.B., “On the relationship of synaptic activity to macroscopic
measurements: Does co-registration of EEG with fMRI make sense?,” Brain Topogr., vol.
13, pp. 79-96, 2000.
O’Leary, D. D. M., Schlaggar, B. L. and Tuttle, R. 1994. Specification of neocortical areas and
thalamocortical connections. Annu. Rev.Neurosci. 17: 419–439.
Ossenblock P., Fuchs M., Nevalis D., Veltman E., Pijn J., Silva F. L. da, “Source Analysis of
Lesional Frontal-Lobe epilepsy,” IEEE Eng. Med. Biol. Mag., 1999, 18(3):67-77
Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE Trans. Systems,
Man, and Cybernetics, 9(1): 62-66.
Pascual-Marqui R.D., Michel C.M., and Lehmann D., Low resolution electromagnetic
tomography: a new method for localizing electrical activity in the brain, Int. J.
Physchopsysiol., 1994, 18:49 - 65,
Paus T, Otaky N, Caramanos Z, MacDonald D, Zijdenbos A, D'Avirro D, Gutmans D, Holmes
C, Tomaiuolo F, Evans AC. In vivo morphometry of the intrasulcal gray matter in the
human cingulate, paracingulate, and superior-rostral sulci: hemispheric asymmetries,
gender differences and probability maps.J Comp Neurol. 1996 Dec 23;376(4):664-73.
Paus T, Zijdenbos A, Worsley K, Collins DL, Blumenthal J, Giedd JN, Rapoport JL, Evans AC.
Structural maturation of neural pathways in children and adolescents: in vivo study.
Science. 1999 Mar 19;283(5409):1908-11.
Paxinos G. (Ed.) The rat nervous system. Second Edition. Academic Press, San Diego, 1995.
Paxinos G, Watson Ch. The rat brain in stereotactic coordinates, Academic Press, San Diego,
1997.
Penfield W, Boldray E (1937). Somatic motor and sensory representation in the cerebral cortex
of man as studied by electrical stimulation. Brain. 60:389-443
Penfield W and Rasmussen T.. 1950: The Cerebral Cortex of Man. A Clinical Study of
Localization of Function. New York, The Macmillan Comp. 1950.
References
xix
Penhune VB, Zatorre RJ, MacDonald JD, Evans AC. Interhemispheric anatomical differences in
human primary auditory cortex: probabilistic mapping and volume measurement from
magnetic resonance scans. Cereb Cortex. 1996 Sep-Oct;6(5):661-72.
Pham, X., Prince J., “A survey of Current Methods in Medical Image Segmentation”, Annual
Review of Biomedical Engineering, 1998
Pham DL, Prince JL Adaptive fuzzy segmentation of magnetic resonance images. IEEE Trans
Med Imaging. 1999 Sep;18(9):737-52.
Pham DL, Xu C, Prince JL. 2000. Current methods in medical image segmentation. Annu Rev
Biomed Eng 2000;2:315-37
Phillips JW, Leahy RM, Mosher JC. MEG-based imaging of focal neuronal current sources.
IEEE Trans. Med. Imag., 1997; 16:338-48.
Portas CM, Goldstein JM, Shenton ME, Hokama HH, Wible CG, Fischer I, Kikinis R, Donnino
R, Jolesz FA, McCarley RW., 1998a Volumetric evaluation of the thalamus in
schizophrenic male patients using magnetic resonance imaging. Biol Psychiatry. May
1;43(9):649-59.
Portas, C. M., Rees, G., Howseman, A. M., Josephs, O., Turner, R., and Frith, C. D. 1998b. A
specific role for the thalamus in mediating the interaction of attention and arousal in
humans. J. Neurosci. 1: 8979–8989.
Powell, T. P. S. and Mountcastle, V. B, (1959) Some aspects of the functional organization of
the cortex of the postcentral gyrus of the monkey: a correlation of findings obtained in a
single unit analysis with cytoarchitecture, Bulletin of the Johns Hopkins Hospital, 105:
133-162.
Pruessner JC, Köhler S, Crane J, Pruessner M, Lord C, Byrne A, Kabani N, Collins DL , Evans
AC. 2002. Volumetry of Temporopolar, Perirhinal, Entorhinal and Parahippocampal
Cortex from High-resolution MR Images: Considering the Variability of the Collateral
Sulcus , Cerebral Cortex, Vol. 12, No. 12, 1342-1353.
Pucci E, Belardinelli N, Regnicolo L, Nolfe G, Signorino M, Salvolini U, Angeleri F. 1998.
Hippocampus and parahippocampal gyrus linear measurements based on magnetic
resonance in Alzheimer's disease. Eur Neurol.;39(1):16-25.
Ribary U, Ioannides AA, Singh KD, Hasson R, Bolton JP, Lado F, Mogilner A, Llinas R.
Magnetic field tomography of coherent thalamocortical 40-Hz oscillations in humans.Proc
Natl Acad Sci U S A. 1991 Dec 15;88(24):11037-41.
Rojas DC, Bawn SD, Carlson JP, Arciniegas DB, Teale PD, Reite ML. Alterations in tonotopy
and auditory cerebral asymmetry in schizophrenia. Biol Psychiatry. 2002 Jul 1;52(1):32-9.
References
xx
Roland P, Svensson G, Lindeberg T, Risch T, Baumann P, Dehmel A, Frederiksson J,
Halldorson H, Forsberg L, Young J, Zilles K. A database generator for human brain
imaging.Trends Neurosci. 2001 Oct;24(10):562-4.
Roland PE, Zilles K. 1996. The developing European computerized human brain database for all
imaging modalities. Neuroimage.4(3 Pt 2):39-47.
Rossi, A., Stratta, P., Mancini, F., Gallucci, P.M., Mattei, P.,Core, L., Di Michele, V.,
Casacchia, M., 1994. Magnetic resonance imaging findings of amygdala-anterior
hippocampus shrinkage in male patients with Schizophrenia. Psychiatry Research 52, 43-
53.
Sarvas J. Basic mathematical and electromagnetic concepts of the biomagnetic inverse problem.
Phys. Med. Biol., 1987; 32:11-22.
Sastre-Janer FA, Regis J, Belin P, Mangin J-F, Dormont D, Masure M-C, Remy P, Froulin V,
Samson Y. Three-dimensional reconstruction of the human central sulcus reveals a
morphological correlate of the hand area. Cereb Cortex 8: 641–647, 1998.
Scherg M.and Cramon D. von. 1985. Two bilateral sources of the late AEP as identified by a
spatio-temporal dipole model, Electroenceph. Clin. Neurophysiol., 62: 232 – 244.
Scherg, M. 1990. Brain electric source analysis: The importance of physiological constraints.
Brain Topography, 3: 268-269.
Schiemann T, Höhne KH, Koch C, Pommert A, Riemer M, Schubert R, Tiede U. Interpretation
of tomographic images using automatic atlas lookup. IEEE: Visualization in Biomedical
Computing. 1994;2359:457-465.
Schmidt DM, George JS, Wood CC. Bayesian inference applied to the electromagnetic inverse
problem. Hum Brain Mapp., 1999; 7:195-212.
Schnack HG, Hulshoff Pol HE, Baare WF, Staal WG, Viergever MA, Kahn RS. Automated
separation of gray and white matter from MR images of the human brain. Neuroimage.
2001 Jan;13(1):230-7.
Schneider RJ, Friedman DP, Mishkin M. 1993. A modality-specific somatosensory area within
the insula of the rhesus monkey.Brain Res. Sep 3;621(1):116-20.
Scott TF, Price TRP, George MS, Brillman J, Rothfus W. Midline Cerebral Malformations and
Schizophrenia. J Neuropsychiatry Clin Neuro 1993; 5:287-293.
Serra J. 1982. Image Analysis and Mathematical Morphology - Vol. 1. Ac. Press, London, 1982
Sekihara K, Nagarajan SS, Poeppel D, Marantz A, Miyashita Y. Reconstructing spatio-temporal
activities of neural sources using an MEG vector beamformer technique. IEEE Trans.
Biomed. Eng., 2001; 48:760-71.
References
xxi
Shen D, Moffat S, Resnick SM, Davatzikos C. 2002. Measuring size and shape of the
hippocampus in MR images using a deformable shape model. Neuroimage Feb;15(2):422-
34
Shenton ME, Kikinis R, Jolesz FA, Pollak SD, LeMay M, Wible CG, Hokama H, Martin J,
Metcalf D, Coleman M, McCarley RW. 1992. Abnormalities of the left temporal lobe and
thought disorder in schizophrenia: a quantitative magnetic resonance imaging study.
N.Engl.J.Med. 327:604-612
Shenton ME, Gerig G, McCarley RW, Szekely G, Kikinis R. Amygdala-hippocampal shape
differences in schizophrenia: the application of 3D shape models to volumetric MR data.
Psychiatry Res. 2002 Aug 20;115(1-2):15-35.
Spinks R, Magnotta VA, Andreasen NC, Albright KC, Ziebell S, Nopoulos P, Cassell M;
2002, Manual and Automated Measurement of the Whole Thalamus and Mediodorsal
Nucleus Using Magnetic Resonance Imaging. Neuroimage. Oct;17(2):631
Silva FL da. 1999. Event related potentials: methodology and quantification. in Nidermeyer E
and FL da Silva eds. Electroencephalography: Basic Principles, Clinical Applications, and
Related Fields. Lippincott, Williams and Wilkins, Baltimore, 1999. pp. 947-957.
Lopes de Silva, F.H. and Van Rotterdam, A., Biophysical aspects of EEG and MEG generation,
in: Electroencephalography: Basic principles, clinical applications and related fields. in
Nidermeyer E and FL da Silva eds. Electroencephalography: Basic Principles, Clinical
Applications, and Related Fields. Lippincott, Williams and Wilkins, Baltimore, 1999. pp.
93-109.
Simoes C., Hari R., Relationship between responses to contra- and ipsilateral stimuli in the
human second somatosensory cortex SII. Neuroimage. 1999 Oct;10(4):408-16.
Singh KD. 1995. Functional Imaging Of the Brain Using Superconducting Magnetometry.
Endeavour. 19(1), pp. 39-44.
Squire L, 1992. Memory and the hippocampus—a synthesis from findings with rats, monkeys,
and humans," Psychol. Rev. 99 [2]: 195-231, April 1992.
Staib L.H., Duncan J.S., 1996. “Model Based deformable surface finding for medical images”,
IEEE Trans. Med. Imag., 15, 5, 712-731.
Steriade, M., McCormick, D. A., and Jones, E. G. (Eds.) 1997. The Thalamus. Elsevier, New
York.
Stern E, Silbersweig DA. Advances in functional neuroimaging methodology for the study of
brain systems underlying human neuropsychological function and dysfunction. J Clin Exp
Neuropsychol. 2001 Feb;23(1):3-18.
References
xxii
Strick PL, Preston JB. 1982. Two representations of the hand in area 4 of a primate. II.
Somatosensory input organization. J Neurophysiol. 1982 Jul;48(1):150-9.
Suddath RL, Christison GW, Torrey EF, Casanova MF, Weinberger DR. Anatomical
abnormalities in the brains of monozygotic twins discordant for schizophrenia. N Engl J
Med 1990; 322:789-794.
Sullivan EV, Marsh L, Mathalon DH, Lim KO, Pfefferbaum A. 1995. Age-related decline in
MRI volumes of temporal lobe gray matter but not hippocampus. Neurobiol Aging. Jul-
Aug;16(4):591-606.
Sur M, Wall JT, Kaas JH . Modular segregation of functional cell classes within the postcentral
somatosensory cortex of monkeys. Science 1981 May 29;212(4498):1059-61
Tanji J, Wise SP. Submodality distribution in sensorimotor cortex of the unanesthetized
monkey.J Neurophysiol. 1981 Mar;45(3):467-81.
Tek, H. and Kimia, B. (1997). Volumetric segmentation of medical images by three-dimensional
bubbles. Computer Vision and Image Understanding, 65(2):246-258.
Tenke C.E., Schroeder C.E., Arezzo J.C., Vaughan H.G, “Interpretation of high-resolution
current source density profiles: A simulation of sublaminar contributions to the visual
evoked potential,” Exp. Brain Res., vol. 94, pp. 183-192, 1993.
Teo PC, Sapiro G, Wandell BA. Creating connected representations of cortical gray matter for
functional MRI visualization. IEEE Trans Med Imaging. 1997 Dec;16(6):852-63.
Tesche C.D., Karhu J., “Somatosensory evoked magnetic fields arising from sources in the
human cerebellum,” Brain Res., vol. 744, pp. 23-31, 1997.
Tesche C.D., Karhu J., “THETA oscillations index human hippocampal activation during a
working memory task,” Proc. Natl. Acad.Sci. USA, vol. 97, pp. 919-924, 2000.
Tiede U, Bomans M, Höhne KH, Pommert A, Riemer M, Schiemann T, Schubert R, Lierse W.
A computerized three-dimensional atlas of the human skull and brain. Am. J.
Neuroradiology. 1993;14:551-559.
Thompson, P.M., Schwartz, C., Lin, R.T., Khan, A.A., Toga, A.W. 1996. 3D statistical analysis
of sulcal variability in the human brain. Journal of Neuroscience. 16(13): 4261-4274.
Toga A. and Mazziota JC, 1996. “Introduction to Cartography of the Brain”, in “Brain
Mapping. The methods”, Toga A. and Mazziota JC (eds), Academic Press, San Diego.
Tsuchiya K, Kosaka K. Neuropathological study of the amygdala in presenile Alzheimer's
disease.J Neurol Sci. 1990 Dec;100(1-2):165-73.
Tzourio N., L. Petit Use of anatomical parcellation to catalog and study structure-Function
Relationships in the human brain. Human Brain Mapping, 1997, (5): 228-232.
References
xxiii
Uutela K, Hamalainen M, Salmelin R.. Global optimization in the localization of neuromagnetic
sources. IEEE Trans Biomed Eng. 1998 Jun;45(6):716-23.
Van Veen BD, Van Drongelen W, Yuchtman M, Suzuki A.Localization of brain electrical
activity via linearly constrained minimum variance spatial filtering. IEEE Trans. Biomed.
Eng., 1997; 44: 867-80.
Wandell BA, Chial S, Backus BT. Visualization and measurement of the cortical surface. J
Cogn Neurosci. 2000 Sep;12(5): 739-52.
Wang JZ, Williamson SJ, Kaufman L. Magnetic source images determined by a lead-field
analysis: the unique minimum-norm least-squares estimation. IEEE Trans Biomed Eng.,
1992; 39:665-75.
Wang L, Joshi SC, Miller MI, Csernanasky JG., 2001 Statistical Analysis of Hipocampal
Asymmetry in Schizophrenia, Neuroimage, 14, 531-545
Watson, C., Jack, C.R., Cendes, F., Volumetric magnetic resonance imaging: clinical
applications and contributions to the understanding of temporal lobe epilepsy. Archives of
Neurology .1997; 54:1521-1531.
Webb S. The physics of medical imaging. 1998. Institute of Physics Pub.Bristol.
Webb J, Guimond A, Eldridge P, Chadwick D, Meunier J, Thirion JP and Roberts N,.
Automatic detection of hippocampal atrophy on magnetic resonance images, Magnetic
Resonance Imaging. 1999; 17, (8):1149-1161
Welker E, Armstrong-James M, Van der Loos H, Kraftsik R. The mode of activation of a barrel
column: response properties of single units in the somatosensory cortex of the mouse upon
whisker deflection. Eur J Neurosci. 1993 Jun 1;5(6):691-712.
White LE, Andrews TJ, Hulette C, Richards A, Groelle M, Paydarfar J, Purves D. Structure of
the human sensorimotor system. I: Morphology and cytoarchitecture of the central sulcus.
Cereb Cortex. 1997 Jan-Feb;7(1):18-30.
Wieringa HJ and Peters MJ. 1993. Processing MRI data for electromagnetic source imaging.
Med. Biol. Eng. Comp., 1993, pp. 600-606.
Wikstrom H, Huttunen J, Korvenoja A, Virtanen J, Salonen O, Aronen H, Ilmoniemi RJ. Effects
of interstimulus interval on somatosensory evoked magnetic fields (SEFs): a hypothesis
concerning SEF generation at the primary sensorimotor cortex. Electroencephalogr Clin
Neurophysiol. 1996 Nov;100(6):479-87.
Williams DJ and Shah M., “A fast algorithm for active contours and curvature estimation”,
Image Understanding, Vol. 55, No. 1, pp14-26, 1992
References
xxiv
Wise SP, Tanji J. 1981. Supplementary and precentral motor cortex: contrast in responsiveness
to peripheral input in the hindlimb area of the unanesthetized monkey. J Comp Neurol.
1981; Jan 20;195(3):433-51.
Woods, R.P., Dapretto, M., Sicotte, N.L., Toga, A.W., Mazziotta, J.C. 1999. Creation and use of
a Talairach-compatible atlas for accurate, automated, nonlinear intersubject registration,
and analysis of functional imaging data. Human Brain Mapping. 8(2-3): 73-9.
Woods R , , “Correlation of bran structure and function”, in Brain Mapping. The methods, Toga
A and Mazziotta JC (Eds.), 1996; Academic Press, San Diego.
Woolsey CN, Fairman D. Contralateral, ipsilateral and bilateral representation of cutaneous
receptors in somatic areas I and II of the cerebral cortex of pig, sheep and other mammals.
Surgery 1946; 19: 684-702.
Wright IC, McGuire PK, Poline JB, Travere JM, Murray RM, Frith CD, Frackowiak RS, Friston
KJ. A voxel-based method for the statistical analysis of gray and white matter density
applied to schizophrenia. Neuroimage. 1995.2 (4):244-52.
Zainea O, Badea A, Kostopoulos G.K., Ioannides A.A. (2002). Identification of primary
somatosensory cortex subdivisions using anatomically constrained dipoles. Proceedings of
the 3-rd European Symposium in Biomedical Engineering and Medical Physics, 30.08-
01.09, 2002. Patras, Greece, pp.28
Zeng X, Staib LH, Schultz RT, Duncan JS. 1990. Segmentation and measurement of the cortex
from 3-D MR images using coupled-surfaces propagation. IEEE Trans Med Imaging. 1999
Oct;18(10):927-37.
Zeineh MM, Engel SA, Bookheimer SY, 2000. “Application of cortical unfolding techniques to
functional MRI of the human hippocampal region”, Neuroimage, 11,Jun. (6 Pt 1) 668-683
Zhang Y, Brady M, Smith S. Segmentation of brain MR images through a hidden Markov
random field model and the expectation-maximization algorithm. IEEE Trans Med
Imaging. 2001 Jan;20(1):45-57.
Zimmerman JE, Thiene P, Harding JT, Design and operation of stable rf-biased
superconductiong point-contact quantum devices and a note on perfectly clean metal
contacts. J. Appl. Phys. 41, pp 1572-1580