Universidad Nacional de Colombia Nacional de Colombia Facultad de Medicina Method to detect the Default Mode Network in fMRI data corrupted with head motion Katherine Andrea Baquero

Universidad Nacional de ColombiaFacultad de Medicina

Method to detect the Default Mode Network infMRI data corrupted with head motion

Katherine Andrea Baquero Duarte

Tutor:Dr. Eduardo Romero Castro

Maestrıa en Ingenierıa BiomedicaFacultad de Medicina

Bogota D.C., Septiembre 2013

Method to detect the Default Mode Network infMRI data corrupted with head motion

Metodo para la deteccion de la Red Neuronal por Defecto en imagenesde fMRI con movimientos fuertes de cabeza.

Katherine Andrea Baquero Duarte

Tesis o trabajo de grado presentada(o) como requisito parcial para optar al tıtulo de:Magister en Ingenierıa Biomedica

Director(a):Tıtulo (MD-M.Sc.-Ph.D.) Eduardo Romero Castro

Lınea de Investigacion:Procesamiento y Analisis de Informacion Medica

Grupo de Investigacion:Computer Imaging and Medical Applications Laboratory - CIM@LAB

Universidad Nacional de ColombiaFacultad de Medicina, Departamento de Imagenes Diagnosticas

Bogota, Colombia2013

Agradecimientos

Quiero agradecer al profesor Eduardo Romero por su guıa, su apoyo constante, su pacienciay por creer en mı durante todo este proceso. A Francisco Gomez quien me introdujo eneste mundo de la neuroimagen, con quien he trabajado hombro a hombro y quien me haapoyado con su retroalimentacion y su amistad. A todas las personas del grupo cim@lab conquienes he tenido experiencias muy gratas y quienes me han dado su valioso apoyo en diversasformas. Tambien al grupo COMA Science Group de la universidad de Liege, Belgica, quienesaportaron los datos y sus valiosos comentarios durante mi proceso; en especial a AndreaSoddu, Athena Demertzi, Steven Laureys, entre otros, quienes me brindaron la oportunidadde trabajar conjuntamente en su grupo de investigacion.

Tambien quiero agradecer a mi familia, quienes siempre han estado apoyandome de todocorazon, me ayudan a estar consciente de cada momento y siempre han estado presentes encada paso que voy dando. A Pieter por su amor, apoyo y por darle un color especial a mivida, a Francoise por su gran amistad y con quien he aprendido tantas cosas y a muchas otraspersonas que no nombro aquı pero igual han estado presentes durante este proceso.

A todos mil gracias por aportar parte de sı mismos durante mis estudios de maestrıa, cadaquien ha dejado un granito de arena en mi ser.

ii

Contents

Abstract ix

1 The Brain’s Default Mode Network 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Generalities of the DMN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2.1 functional Magnetic Resonance Imaging . . . . . . . . . . . . . . . . . . 21.2.2 Functional connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2.3 The components of the DMN . . . . . . . . . . . . . . . . . . . . . . . . 51.2.4 DMN in brain diseases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Methodology for DMN detection in fMRI . . . . . . . . . . . . . . . . . . . . . 71.3.1 Preprocessing of fMRI data . . . . . . . . . . . . . . . . . . . . . . . . . 71.3.2 Model based methods (Seed method) . . . . . . . . . . . . . . . . . . . . 71.3.3 Data driven methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.4 Head motion and related artifacts in fMRI . . . . . . . . . . . . . . . . . . . . . 121.4.1 Artifacts due to head motion . . . . . . . . . . . . . . . . . . . . . . . . 121.4.2 Methods proposed to deal with artifacts . . . . . . . . . . . . . . . . . . 14

1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 A multiscale analysis to detect the DMN in fMRI data corrupted with headmotion 172.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.2.1 Subjects and acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.2.2 Data processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.2.3 The Multiscale Approach (MA) . . . . . . . . . . . . . . . . . . . . . . . 202.2.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.3.1 Motion indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.3.2 Multiscale compared with other processing methods . . . . . . . . . . . 252.3.3 Multiscale results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

A Presented conferences 34

Bibliography 46

iii

List of Figures

1.1 Consistent resting state networks (RSNs) in rs-fMRI for 10 healthy subjects.These networks reflect the “high” and “low” order cognitive functions of thebrain. (Taken from [46]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Regions of the DMN that share a functional connectivity. The most importantnodes of the DMN are the PCC/prec, MPFC and IPL. (Taken from [75]). . . . 6

1.3 BOLD signal of the PCC/prec and MPFC corresponding to the time compo-nents of the DMN; both signals are correlated indicating the functional con-nectivity. (Taken from [15]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.4 fMRI data is represented as a mixture of ICs that are represented by spatialmaps with their corresponding temporal components (Taken from [65]). . . . . 9

1.5 Illustration of the decomposition using spatial ICA. The 4D fMRI data is firstreduced to a 2D matrix (1-spatial and 1-temporal). This 2D data is decom-posed into components that are spatially independent. Each of the ICs has anassociated time course (Modified from [54]). . . . . . . . . . . . . . . . . . . . . 10

1.6 Representative IC-fingerprints of the six different classes explained in the text.Each IC is represented by a fingerprint with 11 parameters. The fingerprintsare visualized in a polar diagram (Modified from [24]). . . . . . . . . . . . . . . 11

1.7 fMRI artifacts caused by motion in the six motion rigid parameters (transla-tions Tx, Ty, Tz and rotations Rx, Ry, Rz). The figures show the misalignmentscaused by each motion parameter. The type of artifacts inserted for each pa-rameter are listed (taken from [27]). . . . . . . . . . . . . . . . . . . . . . . . . 13

2.1 Description of the three preprocessing methods: Standard Preprocessing (SP),ArtRepair (AR) and Multiscale Approach (MA). All methods are based onthe SP. The AR adds correction for large motion and the MA creates dataat different scales depending on the size of the smoothing kernels and thenprocesses each scale separately with the SP. . . . . . . . . . . . . . . . . . . . . 20

2.2 The proposed MA. For each scale a smoothing gaussian kernel is applied tothe raw fMRI data with sizes of 2, 4, and 8mm. Each fMRI scale is processedindependently and the last DMN is a weighted mean among the contributionof all scales. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

iv

2.3 Motion indices of control, EMCS and MCS subjects. Graphs of mean vs.maximum values are displayed for (a) displacement and (b) speed. Each dotrepresents the motion indices of one subject. The mean values comparing themotion indices of control and MCS are illustrated for (c) displacement and (d)speed. The MCS subjects had larger displacement and speed variations, whencomparing with control subjects. Likewise, the motion indices of the EMCSsubject were around the displacement and speed means, when comparing withthe MCS group, indicating that the motion curve of the EMCS is representativefor the MCS subjects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.4 The DMN spatial z-maps of an EMCS subject were compared with the threemethods: SP, AR and MA. Here the SP did not detect a consistent DMN, i.e.,it only preserved the MPFC with very minor contributions of the PCC/precand the left IPL. The AR and MA methods had a stronger DMN with similarGoF values. However, compared to AR, MA did detect the MPFC region. . . . 25

2.5 DMN Statistical spatial maps of (A) control (FDR, p < 0.001), (B) selectedMCS (uncorrected with p < 0.05, for illustration purposes) and (C) all MCSsubjects (uncorrected, p < 0.01, for illustration purposes) were compared forSP, AR and MA methods. The control subjects (A) had a consistent DMNfor the three methods, but the MA a higher t-value in the PCC/Prec. Theselected MCS subjects (B) preserved the DMN in the MA and AR methods,although the MA had significant results in all the DMN regions. The groupof all subjects (C) the DMN was preserved for the MA and AR methods, butthose were more perturbed by noisy components and the DMN regions weresmaller and splitted. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.6 GoF value of control and MCS subjects with the SP, AR and MA methods.The mean between AR and MA for all subjects was similar. The MA methodhad less variation in control subjects meanwhile AR had it for MCS subjects. . 29

2.7 GoF value of control and MCS subjects with the scales S2, S4 and S8. . . . . . 31

v

List of Tables

2.1 Peak t-values of the control subjects group for the main DMN clusters, com-pared with the SP, AR and MA methods. Results had FDR correction withp < 0.001. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.2 Peak t-values for the seven selected MCS subjects group for the main clustersof the DMN. The SP, AR and MA methods are compared. Results had a smallvolume correction and these were significant with FWE p < 0.05 . . . . . . . . 27

2.3 Peak t-values of all the MCS subjects group at the main clusters of the DMN.SP, AR and MA methods are compared. Results were corrected with FDR,significant at p < 0.05 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.4 Peak t-values of the control subjects group for the main DMN clusters, com-pared between scales S2, S4, S8 and the MA method. Results had FDR cor-rection with p < 0.001. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.5 Measurements of the α-weights in each scale for control and MCS subjects.The percentage is the ratio of the standard deviation w.r.t. the mean value. . . 31

vi

List of Symbols

ACC Anterior Cingulate Cortex

AD Alzheimer Disease

AR ArtRepair (Artifact Repair)

BOLD Blood Oxygenation Level Dependent

CCA Cross-Correlation Analysis

deoxyHb Deoxygenated Haemoglobin

DMN Default Mode Network

DoC Disorders of Consciousness

ECN External Control Network

EEG Electroencephalography

EMCS Exit MCS

FDR False Discovery Rate

fMRI functional Magnetic Resonance Imaging

FWE Family Wise Error

FWHM Full Width at Half-Maximum

GoF Goodness of Fit

IC Independent Component

ICA Independent Component Analysis

IPL Inferior Parietal Cortex

LIS Locked-in Syndrome

MA Multiscale Approach

MCS Minimally Conscious State

vii

MEG Magnetoencephalography

MPFC Medial Prefrontal Cortex

oxyHb Oxygen Haemoglobin

PCC/prec Posterior Cinculate Cortex/precuneus

PET Positron Emission Tomography

ROI Region of Interest

rs-fMRI Resting-state fMRI

RSN Resting State Network

sICA spatial ICA

SNR Signal to Noise Ratio

SP Standard Preprocessing

SPM Statistical Parametric Mapping

SVM Super Vector Machine

tICA temporal ICA

UWS Unresponsive Wakefulness Syndrome

VS Vegetative State

viii

Abstract

The Default Mode Network (DMN) is a neuronal network widely used for grading the neu-rological state in patients with disorders of consciousness (DoC). It is detected in functionalmagnetic resonance imaging (fMRI) when the subject is in resting state conditions. However,provided that DoC patients present a high degree of head motion, the fMRI imaging pro-cess will end up contaminated and the DMN will be completely lost or highly blurred. Thiswork presents a novel multiscale method that improves DMN measurement in healthy controlsubjects and DoC patients whose data has been perturbed by head motion. The multiscaleapproach consists in finding the relevant neuronal information at each scale so that the finalDMN measurement is improved, considering a weighted compensation of the DMN at eachscale. This method was compared with two baseline fMRI data processing methods: (1) abasic pipeline without special motion artifact reduction steps; and (2) a method that includessteps that aim to remove artifacts caused by motion.

Keywords: Default Mode Network (DMN), functional Magnetic Resonance Imaging(fMRI), Disorders of Consciousness (DoC), large head motion, multiscale.

Resumen

La Red Neuronal por Defecto (DMN, por sus siglas en ingles), es una red neuronal usadapara determinar el nivel de consciencia en pacientes con desordenes de consciencia (DoC, porsus siglas en ingles). Esta red neuronal se mide con imagenes de resonancia magnetica fun-cional (fMRI) cuando los sujetos se encuentran en estado de reposo. Sin embargo, dado queeste tipo de pacientes presentan un alto movimiento involuntario de cabeza, el procesamientode imagenes de fMRI resultara contaminado y la medida de la DMN se podra perturbar totalo parcialmente. Este trabajo presenta un novedoso metodo multiescala que mejora la medidade la DMN en sujetos de control y en pacientes DoC de los cuales los datos fueron pertur-bados por movimientos de cabeza. El analisis multiescala consiste en detectar la informacionneuronal relevante en cada escala y la medida de la DMN resultante es mejorada teniendo encuenta la compensacion ponderada de la DMN en cada escala. Este metodo fue comparadocon dos metodos de preprocesamiento de linea base: (1) un metodo basico de preproceamientode fMRI sin reduccion de artifactos de movimientos; y (2) un metodo que incluye etapas queremueven artefactos causados por movimiento.

Palabras clave: Red Neuronal por Defecto (DMN), Imagenes de Resonancia Magneticafuncional (fMRI), Desordenes de Consciencia (DoC), movimientos de cabeza, repre-sentacion multiescala.

ix

Chapter 1

The Brain’s Default Mode Network

1.1 Introduction

The brain is a complex network composed of distant anatomical components that are func-tionally connected [83]. This connectivity is defined as the correlation of the temporal activityamong different anatomical components [17, 39]. The components that are connected definesome sub-networks of the brain. These components define the resting state or large-scalefunctional-anatomical networks [69]. One of the most important is the Default Mode Net-work (DMN) [46].

The DMN is the large-scale network which activates at the resting state and deactivateswhen performing specific tasks [30,72,74]. The resting state is defined as a mind state whereina person does not perform any (externally-oriented) task and is not specifically thinking onsomething [39], but instead this is in a state of mindwandering, i.e., internal awareness ismore prominent than awareness of the external world [15]. Early neuroimaging experimentsfound DMN regions to become more active in periods when the subject was not involved inactive tasks [74, 79]. These regions were used as a baseline of the brain activity, in contrastwith the regions activated by execution of one specific task [45]. Nowadays, the resting stateis used as a measure of brain spontaneous activity [29, 73], wherein the DMN indicates thebrain intrinsic functional activity [75]; it is also associated with internal mental processes [59]and its study is of major importance for a better understanding of mental disorders [14].

The DMN connectivity has been shown to be altered in a number of mental disordersand has been used for diagnostic and research purposes [14, 15, 43]. Changes in the DMNhave been associated with Alzheimer’s disease (AD) [16, 40], schizophrenia [37], disorders ofconsciousness (DoC) [11, 12, 43, 85], among others. This network has been widely used fordiagnosis of DoC because it gives an indication of the remaining level of consciousness inthose patients [43]. However, DoC patients present both, lack of neuronal signal and largeinvoluntary movements, at the moment of acquisition [46]. Therefore in these cases, the DMNdetection is a challenging task.

This document presents a proposal of a multiscale method to improve the measurementof the DMN in functional Magnetic Resonance Imaging (fMRI) data that has been perturbedwith head motion. The main hypothesis underlying this work is that as the DMN is composedof large regions scattered around the brain, a scale-space representation would enhance theDMN at certain scales and would decrease the effects caused by head motion. This idea was

1

tested in control subjects and DoC patients. The method was compared with a well estab-lished baseline including the usual steps for fMRI data preprocessing. The proposed methodimproved DMN measurements compared with this baseline for control and DoC subjects.This method also provided comparable results in terms of DMN detection capacity comparedwith other state of the art alternatives, but with more robustness.

This document is organized as follows: in this chapter hereafter the definition of the DMNis introduced, its differences in presence of brain diseases, the methods already proposed forits detection and the challenges due to head motion artifacts applied in fMRI. In chapter 2,the main product of this work is presented, which is the paper that will be submitted to anindexed journal. In the appendix A partial products of this work are presented.

1.2 Generalities of the DMN

The DMN is a specific, anatomically defined brain system preferentially active when individu-als are not focused on the external environment [15]; the DMN is identified principally as theanatomical components that are activated and functionally connected in the resting state, andare deactivated in task-directed states [39]. The functional connectivity is measured with thecorrelation between the temporal signal of each anatomic component [69]. This correlationmeans that those regions are linked together, having a certain communication and makingthe DMN a large scale network of the brain [69].

The DMN has been detected with different modalities, such as fMRI, magnetoencephalog-raphy (MEG), Positron Emission Tomography (PET) and Electroencephalography (EEG)[14]. The fMRI modality has a broader use, because this provides a non-invasive representa-tion of brain activity, conserving high spatial resolution and a medium temporal resolution(around 2-3 sec.) [54].

1.2.1 functional Magnetic Resonance Imaging

fMRI is a technique that measures the changes in brain haemodynamics which are linkedto local neuronal activity. These changes are measured with the blood oxygenation leveldependent (BOLD) signal [61]. The origin of the associated BOLD fMRI signal changes lies inthe different magnetic properties of haemoglobin-carrying oxygen (oxyHb) and deoxygenatedhaemoglobin (deoxyHb). DeoxyHb is slightly paramagnetic relative to brain tissue, whereasoxyHb is isomagnetic [61]. By measuring BOLD single changes, fMRI detects functional activeregions around the brain [54].

The BOLD fMRI signal allows an image spatial resolution in the order of a few millimeters,with a temporal resolution of a few seconds (limited by the haemodynamic response itself) [61].Additionally, fMRI does not rely on the use of radiation exposure and is becoming increasinglyavailable. Both factors make fMRI a modality feasible for the diagnosis of patients [42].

With fMRI, there are two different paradigms to locate active regions of the brain de-pending of the task that a person is performing [42]: the active and passive paradigms. Withthe active paradigm, the subject receives an external stimulus at certain periods of time [83]to examine which regions are activated in response to the stimulus. Meanwhile, with thepassive paradigms, the subject does not perform any task [45, 64]. This passive paradigm iswell known as resting state fMRI (rs-fMRI) and is used to measure the spontaneous activityof the brain [29]. This method looks at oscillating patterns of spontaneous neuronal activity

2

at the low-frequency range, visualizing networks of brain areas that show similar spontaneousactivity (i.e., that represent a functional connectivity) [42].

The DMN has been located in both paradigms. With the active paradigm the DMN hadan activation at the moment of non-stimulus and deactivation at the moment of stimulusapplication [15]. Nowadays, the DMN is mainly measured with the resting state paradigmbecause it is highly associated with spontaneous brain activity, as are the other resting stateneural networks [83]. Further description of these neuronal networks are shown in Sect. 1.2.2.

Another important feature of the DMN is the anti-correlated behavior with the ex-ternal control network (ECN) [30], meaning that their low-frequency fluctuations are ac-tivated/deactivated inversely [14, 30, 32]. The DMN is characterized as the task-negativenetwork, because this has an increased deactivation in attention-demanding cognitive tasks[14, 30], while the ECN is characterized as the task-positive network, because this has anincreased activation on this task [14,30,32].

1.2.2 Functional connectivity

Functional connectivity is defined as the temporal dependence of neuronal activity patternsbetween anatomically separated brain regions [83]. This functional connectivity has been wellstudied with rs-fMRI [14, 83]. When a subject is at rest, the brain is not idle, but rathershows a vast amount of spontaneous activity with functionally connected regions [83].

The functional connectivity is characterized by low-frequency spontaneous correlations(≈ 0.01 - 0.1 Hz) of rs-fMRI time-series [83]. The brain regions that share a functionalconnectivity in resting state are well known as the resting state networks (RSNs) [14,15, 80].The RSNs provide a way to relate the functional connectivity to human behavior. Abnormalconnectivity of RSNs has been found in a wide range of mental disorders [83], meaning thatthose could be used as neuronal biomarkers for diagnosis purposes [80].

Several robust RSNs have been identified across different subjects [23, 80], and those canbe detected in the same rs-fMRI dataset. Although they have similar low frequency patterns,those have a different temporal shift from each other. For this reason, the RSNs could berecognizable in one single fMRI dataset [80].

Resting state networks

The RSNs are differentiated by the brain regions that are functionally connected and havean interpretable function category [80]. RSNs are principally differentiated by those thatreflect the “high order” cognitive functions (the DMN, the ECN and the salience network)and those that reflect a “low order” function (sensorimotor, auditory and visual) [46]. Thoseare represented in Fig. 1.1. Among all RSNs, the most widely studied is the DMN [46, 83].The anatomical and related functions of the DMN will be explained with more detail in Sect.1.2.3.

The ECN (left and right) encompasses bilateral middle, inferior and superior frontal cor-tices, bilateral inferior parietal lobes, Anterior cingulate cortex (ACC)/supplementary motorarea, and bilateral insular cortices. The left ECN is thought to be more involved in cogni-tive and “language” paradigms while the right ECN relates to perceptual, somesthetic, andnociception processing [46,80].

The salience network encompasses fronto-insular and ACC with connections to subcortical

3

Figure 1.1: Consistent resting state networks (RSNs) in rs-fMRI for 10 healthy subjects.These networks reflect the “high” and “low” order cognitive functions of the brain. (Takenfrom [46]).

and limbic structures [46]. The functions associated to this network are: bottom-up detectionof salient events, the switch between other networks (as between the DMN and ECN) to facil-itate access to attention and working memory when a salient event is detected and functionalcoupling with the ACC to facilitate an access to the motor system [66].

The sensorimotor network is composed of supplementary motor area, sensorimotor cortex,and secondary somatosensory cortex. This network corresponds to activations seen in motortasks [46,80].

The auditory network encompasses primary and secondary auditory cortices, includingHeschl’s gyrus, bilateral superior temporal gyri and posterior insular cortex. This network isimportant in audition, such as tone or pitch discrimination, music, and speech [46].

The visual network is associated with three independent networks: the lateral, occipitaland medial visual networks. The lateral visual network includes the middle temporal visualassociation area at the temporo-occipital junction. This network is most important in complex

4

(emotional) stimuli. The occipital and medial visual networks are important in simple andhigher-order visual stimuli [46, 80].

The last RSN is composed of the cerebellum and is associated with action-execution andperception-somesthesis-pain domains [46,80]

1.2.3 The components of the DMN

The DMN represents a brain system (or a composition of interacting subsystems) consisting ofinteracting brain areas that are anatomically connected [15]. The main regions composing theDMN are the Posterior Cinculate Cortex/precuneus (PCC/prec), the medial prefrontal cortex(MPFC) and the inferior parietal cortex (IPL) [14, 15, 39, 69], see Fig. 1.2. The functionalconnectivity among DMN regions is shown in Fig. 1.3.

The PCC/prec is a major node within the DMN [39, 52], although its precise functionremains uncertain [39]. The PCC/prec has a dense structural connectivity with many brainregions. This suggest that the PCC/prec has a role as a cortical hub [52]. Additionally,the PCC/prec is one of the most metabolically active regions during rest and cognitive tasks[52,74]. Some studies implicate the PCC/prec in retrieval of episodic memories [39], regulationof cognition [52] and is also the general sensory information monitor that gathers informationabout the external and internal world [74].

The MPFC is involved in the integration of emotional and cognitive processes [74], selfreferential mental activity and emotional processing [44]. This is divided in ventral and dorsalMPFC (vMPFC and dMPFC). The vMPFC is suggested to be involved in the integrationof emotional and cognitive processes [44], while the dMPFC is associated with introspection[14,44].

The IPL consists of the angular gyrus, supramarginal gyrus and borders the superiortemporal gyrus at a region often referred to as the temporoparietal junction [10]. The IPLhas a variety functional segregation: rostral IPL seems to be involved in motor planning andaction-related functions and is part of the human mirror neuron system. The left caudal IPLis active during language-related tasks with focus on semantic and phonological issues, whilethe right caudal IPL was found to be involved in spatial and nonspatial attention as well asmotor preparation [22].

These are the main features of the DMN in healthy subjects, but this network is disruptedin some mental diseases. A brief introduction of the DMN in mental diseases is shown below.

1.2.4 DMN in brain diseases

As the DMN is an important brain network related to self-relevant mental simulations, withfunctions that include remembering, thinking about the future, among others [15]. Somestudies show that brain diseases influence DMN connectivity. In this section, we present abrief description of the differences of DMN integrity in AD, schizophrenia and DoC.

Alzheimer’s disease

In the context of the DMN, the AD is highly studied because there is an association betweenthe DMN and working memory [14]. Greicius et al. [40] indicated that DMN activity isdeficient in AD compared to healthy elderly controls, with the PCC/prec being the mostcommon site of metabolic and perfusion abnormalities. Furthermore, there is a disrupted

5

Figure 1.2: Regions of the DMN that share a functional connectivity. The most importantnodes of the DMN are the PCC/prec, MPFC and IPL. (Taken from [75]).

Figure 1.3: BOLD signal of the PCC/prec and MPFC corresponding to the time componentsof the DMN; both signals are correlated indicating the functional connectivity. (Taken from[15])

connection between the PCC/prec and other regions, such as hippocampus/entorhinal cortex,medial temporal lobe and MPFC [14].

Moreover, AD’s patients had reported to show decreased deactivation of the DMN in theprocessing of attentional information, suggesting decreased resting state activity and decreasedadaptation of the DMN in comparison to healthy controls [83].

Schizophrenia

Schizophrenia is a psychotic disorder that alters patients’ perception, thought processes, andbehavior as evidenced by hallucinations, delusions, disorganized speech or behavior, socialwithdrawal, and varied cognitive deficits. Garrity et al. [37] found that schizophrenia hadsignificant spatial differences in the DMN, most notably in the frontal, anterior cingulate, andparahippocampal gyri. The medial frontal, temporal, and cingulate gyri showed a temporalcorrelation when the patients had positive symptoms. Additionally, this disorder presented ahigher response in frequency of the DMN (0.08 - 0.24 Hz), and a higher correlation betweenthe DMN and the ECN [14].

6

Disorders of Consciousness

Consciousness is a concept that is divided in two main components: arousal (i.e. wakefulnessor vigilance) and awareness (e.g. awareness of the environment and of the self); [49, 51];in normal states both are normally correlated [11]. Each component involves different brainstructures: arousal involves the activity of subcortical structures encompassing the brain stemreticular formation, hypothalamus, and basal forebrain; awareness is related to the activity ofa widespread set of frontoparietal associative areas. Awareness is divided into self awarenessand external awareness, which normally are anti-correlated [11].

Both consciousness components are disturbed in DoC, and in some situations, wakefulnessand arousal are dissociated [11]. The DoC are characterized as follows [49,85]: coma definedas unarousable unresponsiveness; the unresponsive wakefulness syndrome (UWS, previouslyknown as vegetative state (VS)) [42,50] has preservation of sleep-wake cycles and reflexive butnot purposeful behaviors; the minimally conscious state (MCS), where patients are unable tocommunicate, but show non-reflexive behaviors, which are interpreted as signs of awareness;and the locked-in syndrome (LIS), in which patients are awake and conscious, but have nomeans of producing speech, limb or facial movements.

The DMN has been studied in DoC because this could possibly be a reliable indicator of apatient’s level of consciousness. The DMN connectivity was found to be negatively correlatedwith the level of DoC, ranging from healthy controls and LIS to MCS, VS/UWS and thencoma patients [85]. Additionally, PCC/prec connectivity strength could distinguish betweenMCS and VS/UWS patients [42,85].

1.3 Methodology for DMN detection in fMRI

To detect the functional connectivity in fMRI, first a preprocessing step is required to adjustthe data for its analysis. The basic preprocessing methodology for fMRI is shown in Sect.1.3.1. After preprocessing, for the detection of functional connectivity in fMRI, there are twomain approaches proposed: model based methods and data driven methods [54]. Model basedmethods (Sect. 1.3.2) aim to find the neuronal regions that have a functional connectivitywith predefined neuronal regions of interest (ROI). Data driven methods (Sect. 1.3.3) lookfor independent components (IC) that are linearly mixed and distributed around the wholebrain. The ICs that do not result from artifacts represent neuronal networks.

1.3.1 Preprocessing of fMRI data

The preprocessing step on fMRI data is composed of [18]: correction of head motion byusing least-squares minimization [34], coregistration of functional onto structural data, nor-malization of images with coordinates of Talairach and Tournoux [82], resampling using sincinterpolation and smoothing with a Gaussian kernel to decrease spatial noise [18]; after thisprocess, the 4D fMRI data could be used for the functional connectivity measurement andthen the DMN detection.

1.3.2 Model based methods (Seed method)

The seed-voxel approach methods look for regions that have functional connectivity with theBOLD signal of a region of interest (ROI) [46,54]. For the detection of the DMN, Greicius et

7

al. [39] used the ROI that decreased during paradigm tasks (such as voxels of the PCC/precand MPFC). These ROI are the main clusters of the DMN. Afterwards the regions thathad a functional connectivity with the ROI were selected with a given correlation threshold,forming the whole DMN. The functional connectivity can be measured using cross-correlationor coherence analysis between the seed ROI with the whole brain voxels [54].

Cross-correlation analysis

The cross-correlation analysis (CCA) is a widely used analysis employed to examine the sim-ilarity of two waveforms. As denoted before, the brain regions with a functional connectivityhave a correlation between their BOLD components [39]; for a fMRI BOLD time course Fx(k)and a seed Fy(k), the CCA estimates the correlation as [54]:

Corrx,y(µ) =Covx,y(µ)√

V ar(x)× V ar(y)(1.1)

where V ar(x) and V ar(y) are the variances of Fx(k) and Fy(x), respectively; Covx,y(µ) is thecross variance of Fx(k) and Fy(x) at lag µ:

Covx,y(µ) = E{(Fx(k)− E(Fx(k)))× (Fy(k)− E(Fy(k)))} (1.2)

where E is the expected value and E(Fx(k)) and E(Fy(k)) are the mean values of Fx(k) andFy(k), respectively. When the CCA is higher than a given threshold it is considered thatthe time BOLD components are functionally connected. For the analysis of BOLD data it isusual to consider a zero lag (µ = 0).

Coherence analysis

The coherence analysis is the spectral representation of correlation in the frequency domain.For the time courses defined above (Fx(k) and Fy(k)), the coherence is defined as [54]:

Cohx,y(λ) =|Fx,y(λ)|2

Fxx(λ)Fyy(λ)(1.3)

where Fx,y(λ) is the cross spectra, defined by the Fourier transform of the cross covariance asfollows:

Fx,y(λ) =∑

u

Covx,y × exp−jλu (1.4)

and Fx,x(λ) and Fy,y(λ) are the power spectrum defined as:

Fx,x(λ) =∑

u

Covx,x × exp−jλu (1.5)

Fy,y(λ) =∑

u

Covy,y × exp−jλu (1.6)

With the coherence analysis, sources that have different frequency bands (e.g. functionalconnectivity with frequencies below to 0.1 Hz and cardiac activity with frequencies around1.25Hz), are distinguished. However, this method does not consider the time-shift presentedin the BOLD signal.

8

1.3.3 Data driven methods

Data driven methods look for functional connectivity but independently of prior informa-tion [54]. The most common method used in the literature is the independent componentanalysis (ICA). ICA is a statistical technique that separates a set of signals into independent,uncorrelated and non-Gaussian spatio-temporal components [65]. The technique applied tofMRI data was first proposed by McKeown et al. [65], for the detection of functional connec-tivity and other sources. The main hypothesis is that fMRI is composed of a linear mix ofspatial independent components of neuronal and other sources, as shown in Fig. 1.4 [18, 19].ICA aims to find the linear combinations of the components that are as statistically indepen-dent as possible without the necessity of a priori information [54]; the only parameter neededto be given is the number of independent components (ICs) to decompose the data into. TheICA model is expressed as:

X = AC =N∑

i=1

AiCi (1.7)

where X represents the observed fMRI signals, Ci is the ith IC signal source and A is themixing matrix. The independence of the different sources can be viewed as:

P (C1, C2, . . . , Cn) =N∏

i=1

P (Ci) (1.8)

where P (Ci) is the probability of each independent source. With the calculation of theunmixing matrix (W ), which is the pseudo reverse matrix of A, the ICs can be found by:

C = WX (1.9)

(a) Independent Components (b) Mixed components

Figure 1.4: fMRI data is represented as a mixture of ICs that are represented by spatial mapswith their corresponding temporal components (Taken from [65]).

The solution of ICA is found through the minimization of mutual information of com-ponents Ci [65]. This solution has been proposed with two algorithms [20, 28, 54]: the Info-max [7, 65] and the FastICA [47]. The Infomax algorithm proposes to adaptively maximize

9

the output entropy of a neuronal network with as many outputs as ICs to be estimated, whilethe FastICA algorithm applies the same principle but maximizing the negentropy [54]. Es-posito et al. [28] found that both methods have accurate results. They found that FastICAsurpassed Infomax with spatial and temporal accuracy, but Infomax had better results interms of global estimation of the ICA model and noise reduction [28].

Considering that fMRI data is composed of 4D data (3D spatial-1D temporal), the ICs canbe analyzed as spatially (sICA) or temporally (tICA) independent [54]. Most applications ofICA in fMRI aim to find the components that are maximally independent in space (sICA), as itfinds systematically non-overlapping, temporally coherent brain regions without constrainingthe shape of the temporal response [20]. With sICA, the 4D fMRI data is decomposed in nnumber of spatial ICs that have an associated time course, as shown in Fig. 1.5.

Figure 1.5: Illustration of the decomposition using spatial ICA. The 4D fMRI data is firstreduced to a 2D matrix (1-spatial and 1-temporal). This 2D data is decomposed into compo-nents that are spatially independent. Each of the ICs has an associated time course (Modifiedfrom [54]).

The ICs found by ICA are composed of neuronal or noise independent sources [65]. Todetect the IC that corresponds to the DMN in one fMRI dataset, it is necessary to firstidentify those components that have a functional connectivity (with low-frequency spectraaround 0.01− 0.1 Hz) [83], afterwards selecting the one that matches to the spatial regions ofthe DMN [40].

For the selection of the ICs that correspond to functional connectivity, Greicius et al.,proposed to first remove the ICs in which high-frequency signals (> 0.1 Hz) constituted 50%or more of the total power in the Fourier spectrum, then the remaining components shouldcontain the neuronal connectivity information [40]. Meanwhile De Martino et al., proposeda multidimensional representation of ICs in fingerprints, where the neuronal ICs could beclassified depending on their spatial and temporal parameters [24] (further explanation isfound below). When the neuronal ICs are classified, then the IC of interest is selected by a

10

template-matching procedure, such as the Goodness of Fit (GoF) approach.

Fingerprint

When the data is decomposed into ICs, there is no convention on whether an IC belongs to aneuronal component or to another source. This problem was assessed with the representationof ICs in fingerprints. The IC-fingerprint is a representation of the IC in a multidimensionalspace of descriptive measures, which can be visualized as a polar diagram [24]. The under-lying assumption is that similar types of ICs have a similar fingerprint representation. Thenneuronal ICs could be separated from ICs of other sources.

For each IC, values of 11 descriptive measures were derived from: IC’s voxel values dis-tribution (kurtosis, skewness, entropy), spatial layout (degree of clustering in the anatomicalspace), as well as their temporal (one-lag autocorrelation, entropy) and spectral (power con-tribution in five different frequency bands) properties [24]. Each fingerprint is represented asa point within a multidimensional space of parameters which are visualized using a polar plot.

De Martino et al. (2007), proposed to classify the ICs in six classes of sources [24]: (1)the ’BOLD’ class includes components that are thought to reflect consistently task-related,transiently task-related and non task- related (e.g., resting state) neuronal activity, i.e. neu-ronal connectivity components; (2) the second class (MOT) includes residual motion artifacts;(3) the third class (EPI) includes the typical EPI susceptibility artifact which is maximal inthe frontal part of the brain; (4) the fourth class (VESSEL) includes physiological noise withhighly localized peaks (e.g., large vessels); (5) and (6) the fifth and sixth class include noise athigh spatial (SDN, spatially distributed noise) or temporal (tHFN, temporal high-frequencynoise) frequency. These fingerprints are illustrated in Fig. 1.6.

BOLD MOT EPI VESSEL SDN tHFN

Fingerprints

Spatial

Maps

Figure 1.6: Representative IC-fingerprints of the six different classes explained in the text.Each IC is represented by a fingerprint with 11 parameters. The fingerprints are visualizedin a polar diagram (Modified from [24]).

With the fingerprint representation, the neuronal ICs can be automatically classified fromthose of non-neuronal sources [24]. The classifier should initially be trained with fingerprintslabeled manually by an expert. The classification algorithms already used are the SuperVector Machine (SVM) [24] and Perceptron Neural Network [25]. With this fingerprint repre-sentation, new ICs can be successfully classified by applying the same trained classifier. There

11

is noneed of re-training the algorithm for fMRI datasets that have differences in the fMRIacquisition parameters, stimulation modality and timing considerably different from thoseused for training [24].

Goodness of Fit

The GoF approach is a template-matching measurement that shows the degree of correspon-dence of the ICs along with a spatial template of the neuronal network of interest [40]. Inthe case of the DMN, the spatial template is mainly composed of the PCC/prec, MPFC andIPLs brain regions. To measure the GoF, first all spatial ICs are normalized into z-voxels.This means that the spatial maps are subtracted from the mean and divided by the standarddeviation [65]. Then the GoF value is the average z-score of voxels falling within the templateminus the average z score of voxels outside the template [40]. The IC component that has thehighest GoF value is selected as the component of the neuronal network of interest; in ourcase, the DMN.

1.4 Head motion and related artifacts in fMRI

Despite the frequent use of fMRI for detecting the connectivity among brain regions, thismodality suffers from low signal to noise ratio (SNR) [26]. The main sources of noise are prin-cipally caused by subject’s motion followed by physiological effects. Among motion artifacts,the main artifact is caused by head motion [84], followed by body/arm, eye, tongue and jawmotion [6, 8]. Physiological effects like breathing and heart beating induce noise fluctuationsaround previously known frequencies in fMRI [9, 31, 48], but to a lower degree than headmotion. In this work, we will focus on head motion, which is one of the major sources of noisein fMRI data.

1.4.1 Artifacts due to head motion

Head motion is detrimental to fMRI in two ways: first, it may induce false-positive activa-tions when motion artifacts are correlated with the stimulus, events of interest or neuronalactivation; and second, it inserts false-negative activations when the sensitivity is reduceddue to motion noise [53]. Unfortunately, there is no a way to deduce when a false-positive or-negative activations are induced in the data.

Even the slightest subject motion (less than one degree of rotation or one mm of trans-lation) during a scan can change the initial fixed voxel position in the present scans as insubsequent volumes [36]. Motion could produce misalignments in scan volumes [36] and slices.Slice misalignments vary depending on the motion parameters (translation and rotation inx,y,z) [27], as illustrated in Fig. 1.7. The change of the signal intensity caused by motion canbe much greater than the BOLD effect (the change of BOLD signal is around 1− 2%, whilethe signal change caused by motion could be 5 − 10%), this effect is even increased at theboundaries of tissues, next to the ventricles or edges of the brain [27].

Additional to position displacements, motion induces other types of artefacts, such as:blurring in the case of motion in plane [27], inhomogeneity magnetic field interactions duringrotation of the head [1] and spin-history effects [27].

12

A head motion could cause two types of inhomogeneity distortions with the B0 magneticfield [1]: (i) inhomogeneities that are fixed with respect to the object and (ii) inhomogeneitiesthat change with subject movement [1]. The first case is when the deformation field movesperfectly with the object, a case we would expect to occur when the deformations are causedby the object itself and where rotations only occurred around an axis parallel to the magneticfield [1]. This happens when the structure and the position of the tissue-air boundaries in theobject does not change (with respect to the direction of the B0-field) with object’s motion(rotation in z−axis) [27]. The second case is when the distorting field changes as the subjectmoves in the scanner [1], i.e. the tissue-air boundaries within the object change with respectto the direction of the B0-field (rotation in x− or y−axis) [27].

Spin-history effects are affected by motion that causes movement in the through-slicedirection (translation in z, rotation in x, rotation in y) [27, 36]. It occurs when a spin getsexcited by the RF pulse during the acquisition of one volume and if then the head moves inthe z-direction, the frequency of the precession of the spin will change possibly matching thefrequency of the RF pulse which was planned to happen for the following slice. The followingRF pulse will then excite that same spin again sooner (or later) than when the planned RFpulse was supposed to happen. This will result in a spike in the image intensity. [27]

Slice misalignmentPosition changeBlurring

Slice misalignmentPosition changeBlurring

Slice misalignmentPosition changeSpin history

Slice misalignmentPosition changeBlurringB0 Inhom-1

Slice misalignmentPosition changeBlurringB0 Inhom-2Spin history

Slice misalignmentPosition changeBlurringB0 Inhom-2Spin history

Figure 1.7: fMRI artifacts caused by motion in the six motion rigid parameters (translationsTx, Ty, Tz and rotations Rx, Ry, Rz). The figures show the misalignments caused by eachmotion parameter. The type of artifacts inserted for each parameter are listed (taken from[27]).

13

Measurement of head motion

Knowing the artifacts that a head motion inserts in a dataset, it is important to measure thehead motion presented during the scanning process. For online motion correction, where themotion parameters should be detected during the scanning process, the use of head markerswith compatible hardware is necessary [58]. For offline motion correction, the motion param-eters used are usually obtained in the registration step of preprocessing [36,70]. The methodsproposed to deal with motion artifacts are listed in Sect. 1.4.2.

For experts, a dataset that has motion translations and rotation amplitudes > 1mm/deg(measured with respect to a reference volume) inserts significant confounds in the data [53].Therefore, motion parameters are classified as low motion < 1mm/deg, high motion 1 − 3mm/deg and severe motion > 3 mm/deg [62]. Unfortunately, most datasets have the presenceof high (33%) and severe motion (55%) [62]. However, discarding data due to the maximumvalue of these parameters is not enough, as motion between scans affects the detection ofneuronal connectivity [77,84].

Several indexes for total motion and relative motion have been proposed [70,77,81,84], toevaluate how much the motion affects the neuronal connectivity. Unfortunately, there is noconsensus on which motion index should be used and the results between experiments cannotbe directly comparable. Furthermore, different patterns could be presented in a motion curve.The patterns associated so far are: slow drifts [65], presenting themselves when the head slowlymoves, influencing the total amplitude; peak motion [53], resulting in large variations betweenscans; periodic motion, where low or large peaks are repeated or motion has a frequencypattern; and motion correlated with the stimulus, which induces false-positive activations [58].With all of these different motion patterns presented in a dataset, there is still no a consensuson how much a motion presented in a dataset perturbs the neuronal activity. With thisconcept, Baquero et al. proposed a perturbation index (motion index) based on the changeof the DMN measurement of a given motion curve [5].

1.4.2 Methods proposed to deal with artifacts

Head motion could be corrected at the moment of data acquisition (online) or when thedata has already been acquired (offline). For online motion correction [58], a combinationof external hardware should be used for tracking the real-time head motion [88]. Theseare dependent on scanner compatible sensors, such as camera systems [88], and laser systems,among others [58]. The main drawbacks of using an online motion correction are the following:the tracking system should be as precise as the resolution of a voxel, hardware should becompatible with the scanner, user-friendly markers are needed, fast processing of motionparameters to RF pulses is necessary, among others [58].

Offline motion correction methods aim to reduce the effect of motion artifacts in theacquired data. The most important and widely used method to deal with head motion isthe registration process to align volumes [34]. This method finds the translation and rotationparameters that have the least squares difference between the volume to register and thereference volume, by linearizing the problem with the first order of the Taylor expansion [34].Unfortunately, even with the perfect alignment of volumes, motion artifacts could induce largeconfounds (in some cases 90% of the voxels had a correlation with the motion parameters afterregistration) [36] and additionally, the registration process inserts interpolation effects [41].

14

Additional to registration, an auto-regression model with the six parameters of registrationwas proposed by Friston et al. [36] to reduce the effects of motion. The voxels that had acorrelation with this regression model were substracted from the data. This motion modelis well known as the Volterra expansion [53]. Further approaches were applied, such as theaddition of ’scan nulling’ regressors. In these, other regressors could be modeled, like varianceestimation of between scans [26] or peak inter-scan events of > 0.2mm [53]; white matterand cerebrospinal fluid mean values could be included as well. Even though these regressionapproaches reduced the effects due to motion, they are not enough to deal with large motion,and in some cases, they inserted confounding effects in the neuronal activity detection [68,77].Furthermore, when increasing the quantity of ’scan nulling’ regressors the results could bestatistically inefficient, particularly if the scan nulling effects are of small amplitude [53].

Apart from regression of motion confounds, there are proposals to discard volumes thatare highly contaminated by motion. Power et al. (2012) proposed to “scrub” the scans thathave large motion variations (inter-scan displacements > 0.5mm) or variance larger than 0.5%within the scan signal [70]. Mazaika et al., in the ArtRepair toolbox, proposed to excludescans that are affected by large motion and make an interpolation with the nearest non-repaired scans [63]. Unfortunately, with interpolation, some characteristics of the rejectedscans cannot be totally modeled from the information of surrounding scans [71]. There aretwo main drawbacks of rejecting scans by scrubbing or interpolating. First, there could be asubstantial loss of neuronal information. Second, in the presence of persistent large motion,the majority of scans in the dataset would be candidate to scrubbing or rejection. However,there is a minimum number of scans that needs to be present in order to perform rs-fMRIanalysis. Therefore, some scans significantly affected by motion might still need to be includedin the analysis. Such data would continue to be perturbed by motion. This is the case withclinical data, where patients tend to move more than control subjects [84].

The reduction of motion artifacts and the detection of real neuronal connectivity is stillan open problem that needs more research. Present methods aim to remove artifacts causedby motion. However, an interesting new approach is to enhance the neuronal connectivityinstead of reducing the artifacts.

1.5 Conclusions

The research of the DMN and the RSNs has been increasing in the last decade for variousreasons: resting state is a paradigm that does not need direct interaction with the patient,which could have a broader use in clinical applications; RSNs allow a better understandingof the interactions among several brain regions for high and low levels of cognition; the DMNis a RSN that shows how the baseline processing of the brain is, associated with internalmental processes and the changes of the DMN spatio-temporal connectivity could be used asa biomarker for the diagnosis of neuronal diseases.

Although the DMN has been detected with modalities such as fMRI, PET and EEG,DMN connectivity has been especially well-studied with fMRI, as this modality measures thefunctional connectivity among the brain regions. However, fMRI is susceptible to artifactsthat could introduce confounds in the measurement, mostly because of head motion.

Methods to deal with motion artifacts in fMRI have been proposed so far, and althoughthese seem to reduce the effects caused by motion, total removal of the artifacts has not been

15

attained. This problem is increased in clinical data, because patients tend to move more thancontrol subjects.

Further research is necessary to decrease the artifacts caused by motion. We proposethat the use of methods that aim to enhance the neuronal connectivity will give importantinformation about the neuronal connectivity. For instance, for DoC diagnosis, it is of greatimportance to know whether a patient has a DMN or not, or which patterns of the DMN arestill preserved. In conclusion, the reduction of motion artifacts and the detection of neuronalconnectivity is still an open problem that should be a main focus of future research.

16

Chapter 2

A multiscale analysis to detect theDMN in fMRI data corrupted withhead motion

2.1 Introduction

Detection of specific patterns of neuronal activity is an important step towards diagnosis ofdisorders of consciousness (DoC) [42]. With the advent of modern brain imaging like restingstate functional magnetic resonance imaging (rs-fMRI) [14, 69], task-free paradigms enabledresearchers to find patterns of brain activity that can lead to determining whether a post-coma patient is conscious or not [46]. Task-free neuroimaging is currently a valuable additionto the diagnostic battery of tools in DoC patients, as an examiner is not solely dependenton patient cooperation. Special brain’s connectivity in neuroimaging-based DoC diagnosis,includes the default mode network (DMN) [46], which is known to be important for internallyoriented consciousness [85]. The most common methods used to measure this type of neuronalconnectivity are: seed methods [30] and independent component analysis (ICA) [18,39]. Thedrawback of seed methods is that the connectivity maps are dependent on the voxels selectedas region of interests (ROI) [83] while ICA has consistently reported activation of the DMNduring rest [23,83].

Yet rs-fMRI has confirmed potential of discovering signs of consciousness in unresponsivepatients [43, 49], it is very sensitive to patient movements in the scanner [46, 81] so thatneuronal information is perturbed at different levels [86]. This issue is partially overcomewith the use of foam head rests, but they are usually insufficient. Data resulting from headmovement in the order of several millimeters can be severely jeopardized [84] so that data isnot analyzable, effect that is critical for diagnosis of DoC patients.

Proposals to deal with motion artifacts have been already introduced, starting with thestandard registration process used to align volumes [34], this method yet popular do notremove all motion effects from data [1] and insert interpolation effects [41]. Likewise, otherapproaches included both motion curve or noise component regressions to reduce noise effects.Friston et al., proposed the use of an auto-regression strategy with the six registration pa-rameters to reduce the motion effects [36]. Diedrichsen et al. included a variance estimationof each scan and added it to the linear regression model [26]. Lemieux et al. used the six reg-

17

istration parameters and peak motion measurements for a General Linear Model matrix [53].Even though these approaches reduced the motion effects, these methods are not enough todeal with large motion [70]. In addition to the motion regression effects, it was also proposedto remove outlier scans [57,63,70]. Mazaika et al., in ArtRepair [62,63], replaces the rejectedscans by interpolating the nearest non rejected scans and Power et al. proposes to “scrub”the rejected scans [70, 71]. However, these methodologies can alter the neuronal componentswhile a great amount of information may be lost in presence of data with large head motion,especially in the case of DoC patients.

On the other hand, it has been shown that smoothing fMRI data improves the signal-to-noise ratio (SNR) [35, 67], increases the sensitivity to signals of specific sizes and shapesdepending on the filter size [67]. The neuronal connectivity measurement may vary dependingon the size of the smoothing kernel used as part of the preprocessing step [3, 67]. Ball,et al. presented differences of neuronal activation at different kernel sizes, varying around4− 16mm [3], and they proposed that the use of smoothing kernels with varying kernel sizesmight give more information about the activation of neuronal components and would decreasespatial noise [3]. Lindeberg et al. [56] introduced the use of linear scale-space concepts toanalyze brain activation patterns by using gaussian smoothing kernels. Nevertheless, a scale-space representation is not used in current functional brain imaging studies [3].

Here, we introduce a novel method for enhancing fMRI resting state network detectabilityin control subjects and DoC patients. We propose a ‘multiscale’ analysis using independentcomponent analysis (ICA) [40,65]. Three main points are proposed, first, the use of fMRI dataat different smoothing scale-spaces, second, detection of independent neuronal components ofthe DMN at each scale by using standard preprocessing methods and ICA decomposition atthe particular scale-level, and third, a DMN defined as a weighted contribution of each scaleby using Goodness of Fit (GoF) measurement [40].

This article is organized as follows: Sect. 2.2 presents the materials and methods, whereinthe multiscale approach proposed is presented in Sect. 2.2.3. The results are shown in Sect.2.3 and the discussion in Sect. 2.4.

2.2 Materials and Methods

2.2.1 Subjects and acquisition

Two groups of subjects were selected: 36 healthy volunteers with no neurological or psychiatrichistory (mean age = 43.93, SD = 17.05; 17 women), 18 MCS patients (mean age = 36.25; SD= 17.54; 6 women) and one Exit MCS (EMCS) (17 years old, man). For all patients, clinicalexamination was repeatedly performed considering the Coma Recovery Scale Revised [38]and the Glasglow-Liege Scale [13]. The study was approved by the Ethics Committee of theFaculty of Medicine of the University of Liege. Written informed consent was obtained fromthe healthy volunteers and from the legal representative of all patients.

For all subjects, resting state fMRI-BOLD data were acquired on a 3T MR scanner (TrioTim, Siemens, Germany), with a gradient echo-planar sequence using axial slice orientation(32 slices; voxel size = 3.0 × 3.0 × 3.75 mm3; matrix size = 64 × 64 × 32; repetition time =2, 000 ms, echo time = 30 ms, flip angle = 78◦; field of view = 192 mm). A protocol of 300scans lasting 600 s was performed, and the three initial volumes were discarded to avoid T1saturation effects. A T1 magnetization prepared rapid gradient echo sequence was acquired

18

in the same session for coregistration with functional data on each subject.

2.2.2 Data processing

fMRI data were preprocessed using the Statistical Parametric Mapping (SPM8) (http://www.fil.ion.ucl.ac.uk/spm/). Three types of preprocessing methodologies were compared, asillustrated in Fig. 2.1. The first method, considered as standard preprocessing (SP) in thispaper, included the typical steps of registration, smoothing, normalization and coregistra-tion. The second method used the ArtRepair toolbox (AR) [62, 63], it also used the SP butincluding attenuation of large motion artifact effects [63, 70]. The third method is the pro-posed multiscale approach (MA) which generated fMRI data at different scales and was thenpreprocessed.

Standard Preprocessing (SP)

SP consisted of realignment and adjustment for head movement-related effects using a least-square minimization without higher-order corrections [34], coregistration of functional ontostructural data, segmentation of structural data, spatial and functional normalization w.r.t.standard stereotactic Montreal Neurological Institute (MNI) space, and spatial smoothingwith a gaussian kernel of 8 mm full width at half-maximum (FWHM) to decrease spatialnoise.

Addition preprocessing with Artifact Repair (AR)

The second method used the SP but included attenuation of large motion artifact effects[21, 70] with the ArtRepair (AR) toolbox for SPM [62, 63] (http://cibsr.stanford.edu/tools/ArtRepair/ArtRepair.htm). The additional preprocessing with AR is as follows: first,the motion regression was applied to the registered data [41]. Next, the outlier volumes wereselected, i.e., scans whose average intensity is far apart from the mean or those volumes withlarge scan-to-scan motions [53]. The whole temporal sequence is repaired by interpolating theremoved scan with the nearest non-repaired-scans [62, 63]. Finally, peaks and slow variationare removed by clipping the data and applying a digital high pass filter.

Decomposition of fMRI data into spatial-independent components

For all preprocessing methods, fMRI data was decomposed into 30 independent components(IC) [65,81] by using the Infomax algorithm [7]. IC decomposition was performed at individuallevel using Group ICA fMRI toolbox (GIFT http://icatb.sourceforge.net/) [18]. Datawere centered by subtracting the spatial mean at each time and principal component analysiswas used for dimensionality reduction before applying ICA. Each IC was analyzed by usingthe fingerprint [24] to discriminate BOLD signal among artifactual or other non-neuronalcomponents. Classification of neuronal or non-neuronal ICs was performed with a PerceptronNeural Network classifier, as described in [25]. Among the neuronal components, the DMNcomponent was selected by using the highest value of the GoF [40] using a spatial templateof the DMN [85].

19

http://www.fil.ion.ucl.ac.uk/spm/

http://www.fil.ion.ucl.ac.uk/spm/

http://cibsr.stanford.edu/tools/ArtRepair/ArtRepair.htm

http://cibsr.stanford.edu/tools/ArtRepair/ArtRepair.htm

http://icatb.sourceforge.net/

fMRIdata

Standard8Preprocessing8}SPu

ArtRepair8}ARu

Realign828reslice

Coregistration Normalize82fwhm=8mm

Smoothfwhm=4mm

Motion8regression

Artifact8repair

Despikefilter

Smoothing8-8scalesfwhm={2,4,8}8mm

Multiscale8Approach8}MAu

Figure 2.1: Description of the three preprocessing methods: Standard Preprocessing (SP),ArtRepair (AR) and Multiscale Approach (MA). All methods are based on the SP. The ARadds correction for large motion and the MA creates data at different scales depending on thesize of the smoothing kernels and then processes each scale separately with the SP.

2.2.3 The Multiscale Approach (MA)

Resting state neuronal networks are produced by the low frequency activity of certain brainanatomical regions, within which groups of neurons are synchronically activated in scatteredareas of about 1 cm2 [60]. Interestingly, it has been shown that neuronal activity detectiondiffers with the smoothing kernel [3, 67, 87], the smaller the kernel the larger the observedscattered neuronal activity [67]. The signal to noise ratio (SNR) of fMRI data increasesaccording to the size of smoothing data with gaussian kernels [35,67,78], upon certain limitsof smoothing sizes. All these findings suggest that both, the frequency range and the spatialscale are important to determine the neuronal patterns.

Similar questions have arisen in other communities, for instance, in the image processingand computer vision domains, it has been described a first order relation between the differentscales of an image and the amount of relevant information within each scale. This is knownas the linear scale-space relation, which has been approximated by a simple diffusion equation[55]. Overall, if a dyadic interscale relationship is defined, such relation is completely fulfilled.We hypothesize that the use of a linear multiscale framework would enhance neuronal activityof interest and would decrease the effect caused by head motion and other artifacts. Weexpect that when increasing the scales, the contribution of motion artifacts will be decreasedmeanwhile the main neuronal activity information will be preserved. From a multiscale pointof view, relevant information is robust to the analysis at different scales, i.e., importantinformation should be coherently present at several image scales [55,56].

A construction of multi-scale analysis for any given D-dimensional signal f : <D → < iscarried out by defining the first scale as the original signal and the sequence of scales consistingof series of dyadic decompositions of the first scale [55,56]:

L(·; 0) = f(·) (2.1)

and the representations at coarser scales t > 0

L(·; t) = g(·; t) ∗ f(·) (2.2)

20

are obtained by convolving with gaussian kernels

g(x; t) =1

(2πt)D/2e−x

T x/2t (2.3)

of increasing variance t. Explicitely,

L(x; t) =∫

ξ∈<D

g(ξ; t)f(x− ξ)dξ (2.4)

For fMRI data, the dimension signal D = 3 for each volume and the gaussian function isexpressed as in [67,87]

g(x;σ2) = (2πσ2)−3/2 exp[−‖x‖2 /(2σ2)] (2.5)

where σ is the standard deviation, x is a vector of coordinates and w = σ√

8 loge 2 is theFWHM.

The convolution of two gaussians is gaussian [55,56], but with a FWHM of size

w =√w2

1 + w22 (2.6)

where w1 and w2 are the FWHM of original gaussians [67,87]. The gaussian kernels operationsare commutable with linear operations [87].

Studies have shown that the fMRI optimum filter size is around 8−12mm when analyzinggroups of subjects. For experiments with a small number of subjects or noise contaminatedfMRI data, the recommended filter sizes are around 10− 14mm [67].

This paper proposes the construction of a multiscale analysis to obtain a robust DMNdetection, using a set of gaussian smoothing kernels. The proposal is formulated under cer-tain restrictions: (1) The standard fMRI preprocessing has a smoothing step with a gaussiankernel of 8mm; (2) the resulting gaussian kernel used in fMRI data (considering the kernel ofpreprocessing and of each scale) should be around 8− 12mm, because it is well known this isthe optimum range for fMRI kernel sizes and (3) we suggest to implement the multiscale rep-resentation on raw fMRI data because this is advantageous for the registration preprocessingstep [67].

The multiscale representation was constructed as a dyadic structure, using kernel sizes withincremental values of 2nmm. The multiresolution analysis did not include any downsamplingprocess. Here a three scale representation was used, considering kernel sizes of 2, 4 and 8mm.These kernel sizes along with the preprocessing kernel of 8 mm, according to Eq. 2.6, willhave a final size of 8.25mm, 8.94mm and 11.31mm at each scale, respectively. Those kernelsare in between the recommended range of 8− 12mm [67].

The proposed MA is shown in Fig. 2.2: First, raw fMRI data is spatially smoothed byapplying a gaussian kernel FWHM of 2, 4 and 8mm (S2, S4 and S8, respectively) to obtaina multiscale descriptor of fMRI. Second, each fMRI scale is then preprocessed with the SPmethod (consider that another preprocessing method could also be used) and decomposed in30 ICs with ICA (as explained in sect. 2.2.2). Neuronal ICs are classified [25] and those withthe highest GoF value are selected as the DMN IC at that scale. If there were not a neuronalcomponent or the highest GoF value was less than zero at any scale, the DMN at that scale

21

Preprocessing

ICA

DMN

Data Processing

fMRI scales Multiscale

DMN

Σ

α1

α3

α2

Figure 2.2: The proposed MA. For each scale a smoothing gaussian kernel is applied to theraw fMRI data with sizes of 2, 4, and 8mm. Each fMRI scale is processed independently andthe last DMN is a weighted mean among the contribution of all scales.

will be null. With this restriction we are just aimed to look for scales that preserved neuronalDMN information.

The final DMN of the MA gathers together the individual DMN scale contribution. Wepropose to use normalized α-weights for each scale, based on the GoF values. Better DMNconnectivity is represented by a higher GoF value. To preserve the DMN strength of eachscale, αi weights are normalized with the GoFi at each scale divided by the sum of all GoF,as shown in Eq. 2.7.

αi =GoFi∑iGoFi

(2.7)

where i represents one scale (S2, S4 or S8), GoFi is the GoF value at that scale and αi isthe normalized weight. The final DMN has the DMN contribution of each scale with thenormalized αi weights. The last DMN is measured separately for spatial (Eq. 2.8) andtemporal (Eq. 2.9) contributions of each scale.

sDMNm =∑

i

αi · sDMNi (2.8)

tDMNm =∑

i

αi · tDMNi (2.9)

where sDMNi is the spatial and tDMNi the temporal DMN components of scale i; sDMNm

and tDMNm corresponds to the final spatio-temporal component of the DMN.

2.2.4 Evaluation

With the aim of illustrating the differences of motion magnitude, we calculated the displace-ment (δ, Eq. 2.10) and speed (σ, Eq. 2.11) motion indices [81] per motion curve. Thetranslation parameters (TraX, TraY , TraZ), were expressed in mm and rotation parame-ters (RotX, RotY , RotZ) in degrees. The displacement corresponds to the value of six motionparameters (three translation and three rotation) compared to the reference volume used in

22

the registration process. Speed motion made reference to the derivatives of the six registra-tion parameters; this registered the rapid motion between volumes. The mean and maximumdisplacement and speed were considered as the motion indices of each motion curve.

δ =⟨√

TraX2 + TraY 2 + TraZ2 +RotX2 +RotY 2 +RotZ2⟩, (2.10)

σ =⟨√

δTRTraX2 + δTRTraY 2 + δTRTraZ2 + δTRRotX2 + δTRRotY 2 + δTRRotZ2⟩

(2.11)

DMN statistical spatial maps

The DMN statistical spatial maps [33] for the group of control subjects were corrected at wholebrain level with False Discovery Rate (FDR) < 0.001. These spatial maps were comparedamong scales (S2, S4 and S8) and between SP, AR and MA methods. For all methods, thepeak t-values of each DMN cluster were selected. Each DMN cluster was composed of themain regions of the DMN [14, 15, 43, 74], such as the Posterior Cinculate Cortex/precuneus(PCC/prec), Medial Prefrontal Cortex (MPFC) and both sides of the Inferior Parietal Lobes(IPL).

The MCS subjects were analyzed in two different groups. For the first group, the MCSsubjects, that revealed a recognizable individual spatial z-map of the DMN, were selectedby visual inspection. Those selected subjects were grouped into a DMN spatial map. Smallvolume corrections were performed with a sphere of 15mm, with Family Wise Error (FWE)< 0.05 at the peak voxels of each DMN cluster. All methods were compared with the t-valuesof each DMN cluster (PCC/prec, MPFC and IPL regions). The second group was composedof all MCS subjects. The spatial maps were corrected at brain level with FDR < 0.05 andthe t-values were compared.

DMN measurements

With the aim of comparing numerically the DMN detection with SP, AR and MA, the GoFapproach [40] was used to indicate how well the spatial maps fitted the spatial templateof the DMN [85]. The GoF is the mean average z-score voxels falling within the templateminus the mean average z-score voxels falling outside of the template. The z-score voxels arenormalized by the standard deviation of each spatial map [40, 85]. The GoF illustrates thematch between the IC spatial component with the DMN spatial mask. A higher GoF valuemeant better DMN connectivity. The DMN spatial mask represented the spatial neuronalregions that should be connected. Note that for all the methods, only the neuronal ICs wereselected, based on the fingerprint measurements [24,25].

2.3 Results

2.3.1 Motion indices

The displacement and speed motion indices were calculated for control, a single EMCS caseand MCS subjects. Head motion of control subjects showed lower motion w.r.t. the totalamplitude (displacement) and w.r.t. the variance between scans (speed), when comparing

23

with MCS subjects (Fig. 2.3). Comparing control and MCS motion indices with a t-test ofunequal variance, the mean displacement was not significant (p = 0.07) while the mean speedwas significantly different (p = 0.02).

0 2 4 6 8 100

20

40

60

80

100

Mean Displacement

Max

Dis

plac

emen

t

Measurements of Displacement

ControlMCSMCSexit

0 1 2 3 4 5 60

2

4

6

8

10

12

14

16

18

Mean Displacement

Max

Dis

plac

emen

t

Measurements of Displacement

ControlMCS

//

//

EMCS

(a) Displacement

0 0.2 0.4 0.6 0.8 1 1.2 1.40

20

40

60

80

100Measurements of Speed

Mean Speed

Max

Spe

ed

ControlMCSMCSExit

0 0.2 0.4 0.6 0.8 1 1.2 1.40

2

4

6

8

10

12

14Measurements of Speed

Mean Speed

Max

Spe

ed

ControlMCSEMCS

//

EMCS

(b) Speed

CONTROL MCS0

2

4

6

8

10

Mea

n4of

4Dis

plac

emen

t

Mean4of4Displacement4compared4with4CONT4and4MCS4subjects

(c) Mean displacement

CONTROL MCS0

0.2

0.4

0.6

0.8

1

1.2

Mea

nuof

uSpe

edMeanuofuSpeeducompareduwithuCONTuanduMCSusubjects

(d) Mean speed

Figure 2.3: Motion indices of control, EMCS and MCS subjects. Graphs of mean vs. maxi-mum values are displayed for (a) displacement and (b) speed. Each dot represents the motionindices of one subject. The mean values comparing the motion indices of control and MCSare illustrated for (c) displacement and (d) speed. The MCS subjects had larger displacementand speed variations, when comparing with control subjects. Likewise, the motion indicesof the EMCS subject were around the displacement and speed means, when comparing withthe MCS group, indicating that the motion curve of the EMCS is representative for the MCSsubjects.

In conclusion, MCS had larger motion than control subjects and the inter-scan motionwas more important. The EMCS subject (Fig. 2.3a,b) had a higher displacement and speedthan controls but his motion was within the variance range of the MCS, being a representativemotion curve for the MCS subjects.

24

2.3.2 Multiscale compared with other processing methods

Evaluation using a single Exit MCS case

An EMCS subject was selected to compare the SP, AR and MA methods (See Fig. 2.4). Thissubject was chosen because his motion curve was representative of the MCS subjects, that isto say, most parameters were within the range illustrated in Fig. 2.3. This case is particularlycharacterized by a stronger DMN connectivity, when compared to the MCS subjects.

SP AR MA x = 3 y =

- 50 z = 22

GoF =0.17 GoF =0.21 GoF =0.21

z-values

Figure 2.4: The DMN spatial z-maps of an EMCS subject were compared with the threemethods: SP, AR and MA. Here the SP did not detect a consistent DMN, i.e., it onlypreserved the MPFC with very minor contributions of the PCC/prec and the left IPL. TheAR and MA methods had a stronger DMN with similar GoF values. However, compared toAR, MA did detect the MPFC region.

The DMN spatial z-maps of the EMCS subject are illustrated in Fig. 2.4, for the threemethods. The SP method did not properly detect the DMN, it only preserved the MPFCwith very minor contributions of the PCC/prec and the left IPL. In contrast, the AR andMA methods had a similar detection of the DMN, in terms of the GoF values (0.21 for bothcases). However, the AR method did not include the MPFC. In comparison, the MA methodfound the DMN main regions (PCC/Prec, MPFC and right IPL).

DMN statistical spatial maps

The results of the DMN statistical spatial maps for the groups of control subjects, selectedMCS and all MCS subjects are presented in Fig. 2.5. For every group, the peak t-values ofthe main DMN clusters were compared using the three methods.

For the group of control subjects, spatial maps were corrected with FDR p < 0.001 andthe peak t-values of the main DMN clusters are compared in Table 2.1. For every method, theDMN was consistently detected and the four main DMN clusters were preserved, as shown inthe upper row of Fig. 2.5. Comparing between methods, the MA had balanced results for allthe clusters of the DMN, showing the highest t-value in the PCC/prec and near t-values, whencomparing to the maximum peaks of the MPFC and both IPLs regions. The AR method alsopreserved the DMN, but not at the level reached by the MA method which, as mentionedbefore, obtained a highest peak level at the precuneus structure. The AR had the highestpeaks in both IPLs and detected the PCC/prec, but it had the lowest t-value in the MPFC,

25

SP AR MAxr=

r3rryr=r-r50rrzr=

r22CO

NT

RO

LS

ELr

MC

S

FDRrpr<r0.001

Uncorrecedrpr<r0.01

nr=r36

ALL

rMC

S

Uncorrectedrpr<r0.05

nr=r7 nr=r7

nr=r36 nr=r36

t-value

nr=r7

t-value

nr=r14 nr=r14 nr=r14

t-value

A

B

C

Figure 2.5: DMN Statistical spatial maps of (A) control (FDR, p < 0.001), (B) selected MCS(uncorrected with p < 0.05, for illustration purposes) and (C) all MCS subjects (uncorrected,p < 0.01, for illustration purposes) were compared for SP, AR and MA methods. The controlsubjects (A) had a consistent DMN for the three methods, but the MA a higher t-value in thePCC/Prec. The selected MCS subjects (B) preserved the DMN in the MA and AR methods,although the MA had significant results in all the DMN regions. The group of all subjects(C) the DMN was preserved for the MA and AR methods, but those were more perturbed bynoisy components and the DMN regions were smaller and splitted.

although not so different of what was obtained with the other methods. The SP method hada less consistent DMN representation. Although it had the highest t-value in the MPFC, itpresented the lowest results for the other DMN regions.

For the selected MCS cases, seven subjects were chosen by visual inspection. For thisgroup, a small volume correction was applied with a sphere of radius 15mm, centered at thepeak level of each DMN cluster and being significant with FWE p < 0.05. The results ofthe peak t-values are shown in Table 2.2. For illustration purposes, the spatial maps were

26

Table 2.1: Peak t-values of the control subjects group for the main DMN clusters, comparedwith the SP, AR and MA methods. Results had FDR correction with p < 0.001.

DMN Region Method (x,y,z) BA t-value p-valueSP (-9 -49 31) 31 14.77 < 0, 001

PCC/prec AR (-6 -52 28) 31 17.02 < 0, 001MA (-9 -52 28) 31 17.63 < 0, 001SP (-3 50 1) 10 10.67 < 0, 001

MPFC AR (-6 47 -2) 32 10.42 < 0, 001MA (-6 47 -5) 10 10.54 < 0, 001SP (-51 -64 28) 39 10.74 < 0, 001

IPL Left AR (-48 -61 25) 39 12.60 < 0, 001MA (-48 -61 28) 39 12.39 < 0, 001SP (51 -55 22) 39 9.03 < 0, 001

IPL Right AR (51 -58 31) 39 10.83 < 0, 001MA (51 -58 31) 39 10.47 < 0, 001

BA: Brodmann area

Table 2.2: Peak t-values for the seven selected MCS subjects group for the main clusters of theDMN. The SP, AR and MA methods are compared. Results had a small volume correctionand these were significant with FWE p < 0.05

DMN Region Method (x,y,z) BA t-value p-valueSP ( 3 -64 58) 7 4,69 0, 032

PCC/prec AR ( 0 -64 40) 7 4,10 0, 080*MA ( -9 -58 46) 7 4,44 0, 047SP ( 9 59 13) 10 5,04 0, 018

MPFC AR (12 59 10) 10 5,14 0, 015MA ( 9 59 16) 10 4,58 0, 038SP (-39 -49 46) 40 8,39 < 0, 001

IPL Left AR (-36 -49 46) 40 6,25 0, 002MA (-39 -49 46) 40 6,94 < 0, 001SP ( 45 -58 46) 40 3,95 0, 101*

IPL Right AR ( 54 -58 31) 39 5,81 0, 004MA ( 42 -58 46) 40 4,96 0, 021

*Not significant because p > 0.05; BA: Brodmann area

uncorrected with p < 0.05 (Fig. 2.5 middle row). These MCS subjects had a lower statisticalresults than the control subjects, because these patients normally present a weaker neuronalconnectivity and the quantity of MCS subjects was lower.

For these subjects, the MA and AR methods preserved the DMN main clusters in thespatial maps, but these still had the presence of noise artifacts. Considering the peak t-value results, the MA was the only method wherein all the DMN regions were statisticallysignificant, and it kept the balance among the connectivity of all the regions, i.e. the peakt-values for this method were close to the maximum peak values of each DMN region (except

27

for the IPL left region, where the highest peak value was 8, 39, but that could be caused byartifacts). The AR method had comparable results with the MA method. This reported thehighest peaks of the MPFC and the IPL right, but the PCC/prec region was not significant,although this region had a consistent DMN connectivity in the spatial map. The SP methodhad a noisy map of the DMN, and although it had the highest peak for the PCC/prec and theleft IPL, the IPL right region was not significant and all the other results seem to be biasedby artifacts around the cortex.

Table 2.3: Peak t-values of all the MCS subjects group at the main clusters of the DMN. SP,AR and MA methods are compared. Results were corrected with FDR, significant at p < 0.05

DMN Region Method (x,y,z) BA t-value p-valueSP ( 3 -64 58) 7 4.70 0.118*

PCC/prec AR ( 3 -76 40) 7 4.05 0.032MA (-3 -34 46) 7 4.05 0.040SP ( 48 38 43) 8 4.09 0.118*

MPFC AR (-30 29 52) 8 4.83 0.009MA ( 36 23 52) 8 5.72 0.014SP (-54 -46 46) 40 4.88 0.118*

IPL Left AR (-33 -76 34) 40 4.10 0.030MA (-63 -40 25) 40 4.14 0.035SP ( 45 -61 43) 40 4.41 0.118*

IPL Right AR ( 48 -58 43) 40 7.21 < 0, 001MA ( 48 -58 43) 40 5.18 0.014

*Not significant because p > 0.05; BA: Brodmann area

For the group of all the MCS subjects, ICs were discarded among the 18 MCS subjectsgroup. The SP did not find any neuronal IC for two subjects (see Sect. 2.2.2), and the MAmethod did not find any neuronal DMN-IC in three subjects (see Sect. 2.2.3). Among theremaining cases, the 14 common subjects were chosen for the group of all MCS subjects. Withthis selection, the results of the three methods could be comparable. However, the discardedICs cases did not have a considerable detection of the DMN (evaluated by visual inspection).For these subjects, peak t-values had FDR correction at whole brain level, significant atp < 0.05 (results in Table 2.3). Spatial statistical maps were uncorrected with p < 0.01 forillustration purposes (as shown in Fig. 2.5 bottom row).

For this group, the DMN spatial maps of the AR and MA methods preserved the DMNregions, but these had small sizes and some regions were split. For both methods, all theregions were significant and the peak values were comparable in the PCC/prec and IPL left.The MPFC and IPL right values of the MA, although they were significant and had thehighest peak values, these were a bit more perturbed by artifacts around the cortex than theAR method. The SP method did not show a representative DMN spatial map. It showed partof the PCC/prec but higher up, and this spatial map was completely perturbed by artifacts.None region was statistically significant.

28

DMN measurements

The GoF results for control and MCS subjects were compared for the three methods (Fig.2.6). For control subjects, the AR and MA methods showed better GoF results than theSP. The mean values of the GoF were similar in both methods, but MA had more precisionin the results (as it had less standard deviation). The MCS subjects had, in general, lowerGoF values compared to control subjects, due to lower neuronal activity and motion artifacts.Comparing methods, the AR and MA showed higher GoF values than SP. The mean GoFwas similar for both methods, but in this case, the AR method had lower variation than theMA.

SP AR MA SP AR MA

0

0.1

0.2

0.3

0.4

Control MCS

GoF4of4subjects4vs.4methods

GoF

4val

ue

Figure 2.6: GoF value of control and MCS subjects with the SP, AR and MA methods. Themean between AR and MA for all subjects was similar. The MA method had less variationin control subjects meanwhile AR had it for MCS subjects.

2.3.3 Multiscale results

DMN statistical comparison

The peak values of the main DMN clusters for the group of control subjects are shown inTable 2.4, for the scales S2, S4 and S8 and the final MA method. The peak values werecorrected at the whole brain level with FDR p < 0.001. The main spatial regions of the DMN(PCC/prec, MPFC and IPLs) were significant for all scales and these had different patternswhen increasing scales: the peak t-values of PCC/prec increased, having its maximum in scaleS8; the MPFC had a decrease of the peak t-values, where its maximum was in scale S2; andboth IPLs had its maximum in the medium scale, S4. This showed that the scales enhancedifferently each DMN region, depending of their size and their spatial frequency patterns.The final MA results showed that the scale-space representation in combination with theα-weights, preserved a balance of the contribution of each scale for the final DMN detection.

29

Table 2.4: Peak t-values of the control subjects group for the main DMN clusters, comparedbetween scales S2, S4, S8 and the MA method. Results had FDR correction with p < 0.001.

DMN Region Method (x,y,z) BA t-value p-valueS2 ( 3 -58 25) 31 16.42 < 0, 001

PCC/prec S4 ( 3 -58 25) 31 16.72 < 0, 001S8 (-9 -49 31) 31 18.15 < 0, 001

MA (-9 -52 28) 31 17.63 < 0, 001S2 (-9 47 -5) 10 10.28 < 0, 001

MPFC S4 (-6 50 -5) 10 9.47 < 0, 001S8 ( 9 56 -2) 10 8.58 < 0, 001

MA (-6 47 -5) 10 10.54 < 0, 001S2 (-48 -64 28) 39 13.16 < 0, 001

IPL Left S4 (-48 -64 28) 39 13.58 < 0, 001S8 (-48 -61 25) 39 12.64 < 0, 001

MA (-48 -61 28) 39 12.39 < 0, 001S2 (51 -61 31) 39 11.31 < 0, 001

IPL Right S4 (51 -55 22) 39 12.04 < 0, 001S8 (51 -52 22) 39 10.35 < 0, 001

MA (51 -58 31) 39 10.47 < 0, 001BA: Brodmann area

DMN measurements

The GoF values were compared between subjects and scales, as shown in Fig. 2.7. The resultsof the α-weights are presented in Table 2.5. For control subjects, the GoF values had averagesaround 0.25 for all scales. Scale S4 was slightly greater than other scales, meaning that thisscale had a stronger contribution of the DMN. The same result is reflected in the α-weightsfor control subjects, were the mean and standard deviation were almost similar among allscales, but a bit higher for S4.

For the MCS subjects, the DMN did not have a strong signal as compared with controls,where the mean GoF values varied between 0.1 and 0.15 among all scales. The results of thesesubjects presented a larger variability among scales. The mean GoF values increased withscales, while it decreased for the mean α values. Although, the variation was increasing withscales for both results. The scale S2 had a more stable contribution of the DMN, followedby S4, and scale S8 had the highest variation. This results that for each subject, the scalesrepresent differently the neuronal connectivity. The quantity of cases where there was notany contribution of the DMN-IC (see Sect. 2.2.3), was of three, three and six cases of scalesS2, S4 and S8, respectively. This null contribution is presented because in some cases, therewere not any neuronal ICs or the maximum GoF value of the neuronal ICs was lower thanzero.

30

S2 S4 S8 S2 S4 S8

0

0.1

0.2

0.3

0.4

GoF

Val

ue

GoF Value vs. Smoothing scales

Control MCS

Figure 2.7: GoF value of control and MCS subjects with the scales S2, S4 and S8.

Table 2.5: Measurements of the α-weights in each scale for control and MCS subjects. Thepercentage is the ratio of the standard deviation w.r.t. the mean value.

α-weights %(std/mean) CONTROL MCSS2 0.33± 0.04 (11%) 0.40± 0.16 (40%)S4 0.34± 0.04 (11%) 0.34± 0.14 (42%)S8 0.33± 0.03 (10%) 0.26± 0.17 (65%)

2.4 Discussion

This document presents a novel multiscale method to improve the detection of large-scaleneuronal networks. By using fMRI perturbed data with head motion, the method has demon-strated a remarkable robustness, comparable to a state-of-the art methodology which basicallyapplies regression of motion curves, removes the outliers and interpolates or “scrubb” the re-sultant data [21,63,70,71]. Two main points are proposed in this work: a multiscale represen-tation, generated by a gaussian smoothing of the data with kernels of different sizes [55, 56];and a multi-scale fusion of this information by linearly combining the different scales usingsome α-weights, computed from a normalized version of the GoF values of each of the re-quired scales. With the multiscale decomposition, we are enhancing the neuronal informationby representing it at different levels of details and the relevant information of each scale isselected by using the GoF as the metrics. The final DMN is then a weighted average of thecontribution of each scale information. This multiscale representation resulted in a strongerDMN detection, as compared with the standard methods using only one single scale. Fur-thermore, this method showed promising results in the MCS subjects. In these subjects, themultiscale representation had comparable results with other methods that aimed to reducemotion artifacts.

A large variety of methods deal with the problem of head motion during fMRI datapreprocessing. Some of these proposals remove movement-related components that introducescertain level of noise (regression of motion parameters) [36] and reject the volumes with a

31

certain level of noise [70]. There is still controversy about which motion regression parametersshould be used, some use only the six motion parameters [36,77] others include their derivatives[70], or motion peaks [53], or the variance of the six motion parameters [26], among others.Additionally, it has been shown that using 36 parameters, including the spike regression, themotion effects remarkably decrease [76]. However, the use of too many regression parametersshould decrease the power of the neuronal signal [53] and some additional effects could beinserted in the fMRI data.

The second type of methods reject noisy volumes, by “scrubbing” those volumes [70]or replacing them with the nearest non-rejected scans [21, 63]. This step, although reducesthe most noisy scans, it loses neuronal information and introduces discontinuities to thefMRI data. Furthermore, these methods do not reduce all motion effects, because as noticedby [70, 71], there should be a minimum number of remaining volumes to detect the neuronalconnectivity. This problem is more relevant in DoC patients because they move more thanhealthy subjects [2,43] (see Fig. 2.3). In this case, a higher quantity of scans will be rejectedand several motion artifacts will still be present in the data. In synthesis, these proposalshave different issues that might reduce the neuronal signal when reducing motion effects, theylose neuronal information by discarding outliers volumes, and they consider only one singlescale of detail for a whole set of subjects (normally only one fixed gaussian kernel is used inthe preprocessing step), without considering their differences of neuronal connectivity.

In comparison with the actual proposals, the MA increases the neuronal signal by repre-senting the fMRI data in different levels of detail. As shown in Table 2.4, the scales enhancedifferently each DMN cluster. Additionally, each scale contributes differently to the wholeDMN, depending on the type of subjects (see Fig. 2.7 and Table 2.5). Here, we preserved themain contribution of each scale with the α-weights. These α values give automatic criteriato determine which scales have a stronger representation of the neuronal connectivity, givinga higher flexibility for the detection of the neuronal components. Moreover, the multiscalerepresentation is a method that is automatically adjusted to the type of subjects and theirneuronal connectivity patterns.

The MA showed promising results, comparing with the basic preprocessing baseline, theSP method [40]; and the preprocessing that included motion regression and interpolationof noisy volumes, the AR [62, 63]. Fig. 2.4 shows the detection of the DMN in an EMCSsubject, which had a large head motion, compared to control subjects (Fig. 2.3). In thiscase, the MA detected a more consistent DMN than the AR (as it preserved the MPFC),and both methods surpassed the results of the SP. These results were replicated in the groupof control and MCS subjects, as shown in Figs. 2.5,2.6 and Tables 2.4,2.2,2.3. As expected,when reducing motion artifacts, the neuronal connectivity was better detected with the AR.However, with the multiscale representation, the neuronal connectivity was enhanced and thesignal strengthen out or comparable to the AR. These results show that at representing thedata with different levels of detail, the neuronal detection will have higher signal strength,compared with the two baseline methods used in the present investigation.

Although the multiscale methodology presented in this work surpassed the basic prepro-cessing pipeline, Fig. 2.5C shows that the AR and MA still had the presence of artifactsmixed with the DMN connectivity. Further tests are needed to know at which stage themultiscale decomposition should be applied (in the raw fMRI data, at the end of the datapreprocessing or in a middle step). Additionally, the GoF matching-template could be im-proved or changed by other metrics. In Fig. 2.4 the GoF value was equal for the MA and

32

the AR, although the MA highlighted a more representative DMN (it preserved 3 out of 4main DMN clusters) than the AR (it preserved only 2 main DMN clusters). Additionally,when the maximum GoF value of the neuronal ICs in one scale is negative, the DMN at thatscale is null. However, some scales preserved an unrepresentative DMN because the GoF hada low but positive value. A more restrictive threshold should lead to discover regions with amore precise biological meaning, although, its definition is a difficult task because the MCSand UWS/VS patients have a low neuronal connectivity, resulting in low GoF values. Finally,expressing the α-weights with an exponential or logarithmic GoF relation would associate astronger weight to the scales that had a major DMN contribution.

The results indicate that the multiscale representation can be used for further neuronalconnectivity and large-scale neuronal networks analysis. This representation enhances theneuronal components, reduces the effects due to motion artifacts and is adaptable to the typeof subjects and their neuronal patterns.

33

Appendix A

Presented conferences

Partial products of this work were presented in two international conferences and one confer-ence’s paper was submitted and accepted:

• Poster presentation [5]:TITLE: An index for large head motion on resting state fMRI.AUTHORS: Katherine Baquero, Francisco Gomez, Audrey Vanhaudenhuyse, AthenaDemertzi, Luaba Tshibanda, Olivia Gosseries, Quentin Noirhomme, Steven Laureys,Eduardo Romero, Andrea Soddu.CONFERENCE: Human Brain Mapping 2012, June 10-14 2012 in Beijing, China.Poster No. 616.https://ww4.aievolution.com/hbm1201/index.cfm?do=abs.viewAbs&abs=6630

• Poster presentation [4]:TITLE: A Multi-Scale Analysis to set the Default Mode Network in noisy fMRI data.AUTHORS: Katherine Baquero, Francisco Gomez, Andrea Soddu, Audrey Vanhauden-huyse, Athena Demertzi, Jean-Flory Tshibanda, Olivia Gosseries, Quentin Noirhomme,Steven Laureys and Eduardo Romero.CONFERENCE: BC11: Conference Abstract: BC11 : Computational Neuroscience &Neurotechnology Bernstein Conference & Neurex Annual Meeting 2011., October 4-6,2011 in Freiburg, Germany.Poster No. T 45.http://www.frontiersin.org/10.3389/conf.fncom.2011.53.00067/event_abstract

• Accepted conference’s paper:TITLE: A multiscale method for a robust detection of the Default Mode Network.AUTHORS: Katherine Baquero, Francisco Gomez, Christian Cifuentes, Pieter Gulden-mund, Athena Demertzi, Audrey Vanhaudenhuyse, Olivia Gosseries, Jean-Flory Tshibanda,Quentin Noirhomme, Steven Laureys, Andrea Soddu, Eduardo Romero.CONFERENCE: IX International Seminar on Medical Information Processing and Anal-ysis (SIPAIM), November 11-14 2013 in Mexico City, Mexico.

34

https://ww4.aievolution.com/hbm1201/index.cfm?do=abs.viewAbs&abs=6630

http://www.frontiersin.org/10.3389/conf.fncom.2011.53.00067/event_abstract

��

��

��

��

�� !��

"#�$��%� �� &��'�$ ��(��)�� '��

��

�� * ��)��+��, �$��&�� #��-�#�� *��.��-�#�� -��'�� +�-��#�� )��- ��, �$��/��0*��0*��#*��

��

/1)2�� *��+��$��#�� #��/�/� �� #�� $��3456�&$��#�� #� ��#�++#�� *�� /�� /�#��*�� /��3456�'��#�� $� �+�$��#��+��+�� /��#$#��/��*�� /��3��5�� #�� /�� #�� 7��$#�� #� ��#�� 3�56�2 ��++��#$#��/��*�� +#��/1)2�$�#��8�� +�� #$#�� #�� /�� 6�2 ��8��9��8�+��+�� $#�� /��+�� #��*�� *�� 8��9��:)'&�;6�!�� #$#��//��/�+��+�)'&��<��/��#�� 7��8 �� 8��+�� $6

��

!�+��+��+�� <��##�� *6��6�2 ��##��<=��*�� /�� *�/��+��#��*�� /��+�� <��3�56�!�� /�� ++#��#�� 7��#�� *�� +�� 8�� +�� #��+�� 6�"��*� �#�� +�� +�+�� *>��*�� #�� *�� *6��*�� /��//�� *��)+��3�56�"��*� �#�� +�� +�� ?��+��+��#��+� �� *�/��2-��3@56�'+��+��#�� /�� /��)'&��/�� +�� /��$�� #��+��#�:�*��/��#�� *��98 ��9��+��# ��+�;�� +��#�� /�� :� =#�*��#�� +��#� ��+��+�8��/��/�$�/�� 3��>?A?6??B��>?6??BA?6?��>?6?�A?6?@��>?6?@A?6�� @>?6�A?6�@�C�5�� 8 ��/��+�8��/�� #�3��5�� = �� #�/�� 3��@5;�3�56� �� #�� /�� 8��9�� +�� *�� 8 ��/��$��/��#�� 7�� /�� /��+�� 6�)'&�� #�� *��//� ��6�!��+�� ++��++#��&�� #�� #��+��/��//� ��6�2 ��8��8��/�#�� /��)'&�� *��//� �� =�� #�$��/��*�:��;�3�56�� ##��D�� 2 �<�:D2;��+�� /�� #��+��6

��

OHBM https://ww4.aievolution.com/hbm1201/index.cfm?do=abs.viewAbs&sub...

1 de 3 23/01/2012 11:46 p.m.

@��#�� 7�� /�� :�"-;�+�� 8��$�#��+�/�� /��+��+�� <6��"-�+�� #��/�� #�� *#��//��/1)2�� *��#��3��56�)�� *=��"��8�� !�1)�� :$�<#��E��6?�F�6?�F��6�@��G��<��E�4��F�4��F��G��+�� E��G��??�� ;6�!��/��#��1��&�8��9:�1&;�8��#��)'&��/�� 3�56�'#�� /��)'&��8��+�/�� *��+��#/�#�� *�� =�/=��:��;��++��3@56�H��9��/�$�� #��8��*��#��:��I?6�B�� +#�� #�� ?6@��;�3�56��+�� $��8��<��/��'D1�+��*�� +��6��#��/��B@�:��F�@;� 8�� 8��* ��6�!+�/�� /��+��+�� <�8�� #��/�� *��#�8�� *�+�� 6�!�� )"-�� *� *��D2��#��8��?6J�6��*6��6��##��++#�� /��D2��#��/��#�8�� *�� $�6�� $��##�� 8��++��#�8�D2�$�#�� $��+�� *�$�#��6

��

C��8�+��+�� $#�+�� <�� /��//�� #$#�� #��*�� )'&�6�!�� <�+��$�� $�� #�� 7��$#��/1)2�� *�� #� ��#�� *�6

��

1�� -�� D�+�� *

��

��

3�5�-��H6-6�:�??B;��K' ��$�� , �/�� *�1�� /��L��&�� -��+��'�� $�#6��JB��++6��?J=��B6

3�5��1�� 6�:�??�;��K-#��/�� /�/1)2�� + � ��+� �� *�2-=/� *�+�� ++��$�� #��/��L��&��2��*��$�#6�� 6��++6��=�J�6

3�5�� )6D6H6�:�?��;��K!��#��+��>��*� *��#�� #�+�� * �� L�D�� )��* �� 6��?6�?�4M76+��6�?��6?�6?�J6

3�5��16�6�:�??�;��K�/��#�=�� 8��9��$�� *��#��N��/��#��*� *>�($�� /��/� �� #�1)2L��D�� *��/��&�� #��/�'�� $�#6��?�� 6��++6��4��=�4��6

3@5�C�$�� 6��:�JJJ;��K�� /�<�=+�� #*��/�� + � ��+� �� #��L�2(((�!�� &��#�&�8��9��$�#6��?��++6�4�4�=�4��6

345�1��9��D6�:�??J;��K1�� '�/�8��/��/1)2�� #��/��-#� ��#�'� 7��L��C�� 1�++� *6

3�5�'��6�:�?��;��K2� ��/�� *��/��#�=��+� �� +��#�2-�� #��/�+�� 8��/�� L��C�� 1�++� *��>��?6�??�M �6��J

OHBM https://ww4.aievolution.com/hbm1201/index.cfm?do=abs.viewAbs&sub...

2 de 3 23/01/2012 11:46 p.m.

9/3/13 Frontiers | A Multi-Scale Analysis to set the Default Mode Network in noisy fMRI data

www.frontiersin.org/10.3389/conf.fncom.2011.53.00067/event_abstract 1/2

A Multi-Scale Analysis to set the Default Mode Network in noisy fMRIdataKatherine Baquero1*, Francisco Gómez2, Andrea Soddu2, Audrey Vanhaudenhuyse2, Athena Demertzi2, Jean-Flory

Tshibanda3, Olivia Gosseries2, Quentin Noirhomme2, Steven Laureys2 and Eduardo Romero1*

1 National University of Colombia, Biomedical Engineering, Colombia

2 University of Liege, Cyclotron Research Centre and Neurology Department, Belgium

3 University of Liege, CHU Sart Tilman Hospital, Belgium

The Default Mode Network (DMN) is presently defined by those brain zones involved in maintaining a baseline brain activation. This

network is usually revealed using Independent Component Analysis upon the fMRI data. However, a number of factors can easily

perturb the acquired data, in particular a large head motion. Yet this problem has been partially overcome by registering the acquired

brain volume, registration is still very limited in case of large head movements, a frequent scenario in patients with disorders of

consciousness.

This article presents a multiscale analysis of the fMRI data which improves the robustness with which the DMN is detected in subjects

that move the head during the acquisition process. Initially, the method obtains multiple scales by filtering the original volumes out

with a sequence of gaussian filters. Each fMRI scale is preprocessed by registering, normalizing, smoothing and coregistering, as

described elsewhere, using the SPM software. These preprocessed data are then decomposed into the spatial-temporal components

with the FastICA algorithm. The DMN is then selected with the Goodness of Fit approach (GoF), for each of the different scales.

Finally, the components of the DMN at different scales are summed up, ruling out the original volume.

The approach was validated by perturbing the acquired sequence of five healthy subjects with the six rigid-body movement series

obtained from 15 subjects (5 Control, 3 Minimally Conscious State, 2 Locked in Syndrome and 5 Vegetative State), with head motion

from 1mm up to 15 mm.

In summary, the proposed method was applied to 18 disturbed data that lost the DMN, from which 11 fMRI data recovered the DMN.

The mean of the Goodness of fit of the DMN component increased from 0.1 to 0.4. This multiscale analysis improves the robustness of

the DMN detection in case of large head movements.

Keywords: Computational Neurosciences, Default Mode Network, fMRI, High head motions, multiscale analysis

Conference: BC11 : Computational Neuroscience & Neurotechnology Bernstein Conference & Neurex Annual Meeting 2011, Freiburg, Germany, 4 Oct - 6 Oct, 2011.

Presentation Type: Abstract

Topic: neurons, networks and dynamical systems (please use "neurons, networks and dynamical systems" as keywords)

Citation: Baquero K, Gómez F, Soddu A, Vanhaudenhuyse A, Demertzi A, Tshibanda J, Gosseries O, Noirhomme Q, Laureys S and Romero E (2011). A Multi-Scale Analysis to

set the Default Mode Network in noisy fMRI data. Front. Comput. Neurosci. Conference Abstract: BC11 : Computational Neuroscience & Neurotechnology Bernstein

Conference & Neurex Annual Meeting 2011. doi: 10.3389/conf.fncom.2011.53.00067

Received: 23 Aug 2011; Published Online: 04 Oct 2011.

* Correspondence:

M iss. Katherine Baquero, National University of Colombia, Biomedical Engineering, Bogotá, Bogota DC, 11001000, Colombia, [email protected]

Prof. Eduardo Romero, National University of Colombia, Biomedical Engineering, Bogotá, Bogota DC, 11001000, Colombia, [email protected]

© 2007 - 2013 Frontiers Media S.A. All Rights Reserved

EVENT ABSTRACT Back to Event

Search 1

Katherine Baquero My Journal Submit

Journals A-Z By Subjects BooksImagesBlogsNewsVideosPeopleJobsEvents

A multiscale method for a robust detection of the DefaultMode Network

Katherine Baqueroa, Francisco Gomezb, Christian Cifuentesa, Pieter Guldenmundc, AthenaDemertzic, Audrey Vanhaudenhuysec, Olivia Gosseriesc, Jean-Flory Tshibandad, Quentin

Noirhommec, Steven Laureysc, Andrea Soddue, Eduardo Romeroa

a Computer Imaging and Medical Applications Laboratory CIM@LAB, Universidad Nacionalde Colombia, Bogota, Colombia

b Computer Science Department, Universidad Central de Colombia, Bogota, Colombiac Cyclotron Research Centre and Neurology Department, University of Liege and Centre

Hospitalier Universitaire du Sart-Tilman, Liege, Belgiumd Neurology Department, CHU Sart Tilman Hospital, University of Liege, Belgium

e The Brain and Mind Institute, Department of Physics & Astronomy, Western University,London Ontario, Canada

ABSTRACT

The Default Mode Network (DMN) is a resting state network widely used for the analysis and diagnosis of mentaldisorders. It is normally detected in resting-state fMRI data, but for its detection in data corrupted by motionartifacts or low neuronal activity, the use of a robust analysis method is mandatory. In fMRI it has been shownthat the signal-to-noise ratio and the detection sensitivity of neuronal regions is increased with different smooth-ing kernels sizes. Here we propose to use a multiscale decomposition based on a linear scale-space representationfor the detection of the DMN. Three main points are proposed in this methodology: first, the use of fMRI dataat different smoothing scale-spaces; second, detection of independent neuronal components of the DMN at eachscale by using standard preprocessing methods and independent component analysis decomposition at scale-level;and third, a weighted contribution of each scale by the goodness of fit measurement. This method was appliedto a group of control subjects and was compared with a standard preprocessing baseline. The detection of theDMN was improved at the single subject level and at the group level. Based on these results, we suggest to usethis methodology to enhance the detection of the DMN in data perturbed with artifacts or in data of subjectswith low neuronal activity. Furthermore, the multiscale method could be extended for the detection of otherresting state neuronal networks.

Keywords: Multiscale representation, fMRI, Default Mode Network, resting state neuronal networks, linearscale-space

1. INTRODUCTION

The brain is a complex network composed of distant anatomical components that are functionally connected.1

This connectivity is defined as the correlation of the temporal activity among different anatomical components.2,3

The components that are connected define some subnetworks of the brain, defined as the resting state or large-scale functional-anatomical networks.4 One of the most important is the DMN.5

The DMN is a large-scale network that activates during the so-called resting state, and deactivates whenexternally-oriented tasks are performed.6,7 In the resting state, the subject is not presented with (salient) externalstimuli and is not thinking about something specific.3 Instead, the brain is in a state of mindwandering, whereinternal awareness is more prominent than awareness of the external world.8 Early neuroimaging experiments

Further author information: (Send correspondence to Dr. Eduardo Romero)Eduardo Romero: e-mail: [email protected], Telephone: +57 (1) 3 16 54 91

found DMN regions to become more connected in periods when the subject was not involved in active tasks,6

and this activity was therefore used as a baseline of brain activity to discriminate brain activity associated withexecution of the active task.9

Nowadays, the resting state is used to measure spontaneous activity of the brain.10 The DMN connectivityhas been associated with internal awareness and has been shown to be disrupted in mental disorders,11 suchas Alzheimer’s disease,12,13 schizophrenia,14 and disorders of consciousness (DOC).15–17 Assessment of DMNconnectivity could therefore be useful for diagnostic and research purposes.8,11,15

The most common methods used to measure this type of neuronal connectivity are seed-based methods7 andindependent component analysis (ICA).3,18 The drawback of seed-based methods is that the connectivity mapsare dependent of voxels selected as region of interests (ROI),1 while ICA has consistently reported the activationof the DMN during rest.1,19

It has been shown that smoothing fMRI data improves the signal-to-noise ratio (SNR)20,21 and increases thesensitivity to signals of specific sizes and shapes depending on filter size.20 The neuronal activity measurementmay vary depending on the size of smoothing kernel used as part of the preprocessing step.20,22 Ball, et al.(2012) presented differences of neuronal activation at different kernel sizes, varying around 4 − 16 mm,22 andthey show that the use of smoothing kernels with varying kernel sizes should give more information about theactivation of neuronal components and that it would decrease spatial noise.22 Lindeberg et. al (1999)23 proposedthe use of linear scale-space concepts to analyze brain activation patterns by using gaussian smoothing kernels.However, scale-space representation is rarely used in current functional brain imaging studies.22

In this paper, we introduce a novel method for enhancing fMRI resting state network detectability applied tocontrol subjects. We propose a ’multiscale’ analysis based on ICA.12,24 Three main points are proposed in thismethodology. First, the use of fMRI data at different smoothing scale-spaces; second, detection of independentneuronal components of the DMN at each scale by using standard preprocessing methods and ICA decompositionat scale-level; and third, a weighted contribution of each scale by goodness of fit (GoF) measurement.12

2. MATERIALS AND METHODS

2.1 Subjects and acquisition

A group of 36 healthy volunteers without a neurological or psychiatric history were selected (mean age = 43.93,SD = 17.05; 17 women). For all subjects, resting-state fMRI-BOLD data were acquired on a 3T MR scanner(Trio Tim, Siemens, Germany), with a gradient echo-planar sequence using axial slice orientation (32 slices;voxel size = 3.0× 3.0× 3.75 mm3; matrix size = 64× 64× 32; repetition time = 2, 000 ms, echo time = 30 ms,flip angle = 78; field of view = 192 mm). A protocol of 300 scans lasting 600 s was performed, and the threeinitial volumes were discarded to avoid T1 saturation effects. A T1 magnetization prepared rapid gradient echosequence was acquired in the same session for coregistration with functional data on each subject.

2.2 Data Processing

fMRI data were preprocessed using the Statistical Parametric Mapping (SPM8) (http://www.fil.ion.ucl.ac.uk/spm/). The preprocessing consisted of realignment and adjustment for head movement-related effectsusing a least-square minimization without higher-order corrections,25 coregistration of functional onto structuraldata, segmentation of structural data, spatial and functional normalization into standard stereotactic MontrealNeurological Institute (MNI) space, and spatial smoothing with a Gaussian kernel of 8 mm full width at half-maximum (FWHM) to decrease spatial noise.

fMRI data were decomposed into 30 independent components (IC)24,26 by using the Infomax algorithm.27

IC decomposition was performed at individual level using the Group ICA fMRI toolbox (GIFT http://icatb.sourceforge.net/).18 The spatial mean was subtracted per time point and principal component analysis wasused for dimensionality reduction before applying ICA. Each IC was analyzed using the fingerprint28 to discrim-inate neuronal (BOLD) signal from artifactual or other non-neuronal components. Classification of neuronal ICcomponents was performed with a perceptron neural network classifier, as used in.29 Among the neuronal com-ponents, the DMN component was selected by searching for the highest value of the GoF.12 The GoF is defined

Figure 1. The multiscale approach. For each scale, a smoothing gaussian kernel with a FWHM of 2, 4, and 8 mm, isapplied to raw fMRI data. Each scale is processed independently and the DMN is detected at scale level. The resultingDMN is a weighted contribution of all scales.

by12 as the average mean z-score voxels falling within the template minus the average mean z-score voxels fallingoutside of the template. The spatial template of the DMN created for a previous study in DoC patients in17 wasused in this work.

2.3 Multiscale approach

Resting state neuronal networks are produced by the low frequency activity of certain brain anatomical regions,within which groups of neurons are synchronically activated in scattered areas of about 1 cm2.30 It has beenshown that neuronal activity detection differs with the smoothing kernel,20,22,31 where the SNR of fMRI dataincreases when the size of gaussian smoothing kernels increases,20,21,32 up until a certain smoothing size limit.Furthermore, smoothing raw fMRI data improves the registration preprocessing. All these findings suggest thatboth the frequency range and the spatial scale are important for determining the neuronal patterns.

It has been described that the linear scale-space (multiscale) representation preserves the main informationamong scales,33 wherein neuronal components are highlighted depending on the smoothing kernel sizes.20 Wesuggest that a multiscale representation based on linear scale-spaces with gaussian kernels would give moreinformation about the neuronal connectivity and would increase the robustness of the neuronal measure.

The definition of a multiscale representation is based on gaussian kernels, which is expressed by20,31 as

g(x;σ2) = (2πσ2)−3/2 exp[−‖x‖2 /(2σ2)] (1)

where w = σ√

8 loge 2 is the FWHM and x is a D-vector of coordinates, in which for fMRI volumes D = 3. Thevarying factor for the multiscale representation is the FWHM kernel size.

This paper proposes the construction of a multiscale analysis to obtain a robust DMN detection, using aset of gaussian smoothing kernels. The multiscale representation was constructed as a dyadic structure,33 usingkernel sizes with incremental values of 2nmm. The multiresolution analysis did not include any downsamplingprocess.

The proposed multiscale approach is shown in Fig. 1: first, raw fMRI data are spatially smoothed by applyingFWHM gaussian kernels of 2, 4 and 8mm (S2, S4 and S8, respectively) to obtain a multiscale descriptor of fMRI.Second, each fMRI scale is then preprocessed and processed with ICA and the DMN at each scale is detected(as explained in section 2.2). The final DMN of the multiscale method considers a weighted contribution of allscales based on the GoF measurement. We propose to use normalized α-weights for each scale, based on the GoFvalues. Higher GoF means better DMN detection. Normalized αi-weights represents the amount of contribution

of each scale, based on GoF values. The αi-weight of one scale is defined as the GoFi value at that scale dividedby the sum of all the GoF, as shown in Eq. 2.

αi =GoFi∑iGoFi

(2)

where i represents one scale (S2, S4 or S8), GoFi is the GoF value at that scale and αi is the normalized weight.The final DMN has the DMN contribution of each scale with the normalized αi weights. The last DMN ismeasured separately for spatial and temporal contributions of each scale.

2.4 Evaluation

We compared our multiscale approach with the standard prepocessing pipeline proposed for the DMN extrac-tion.12 We aim to compare the DMN detection capacity for both methods across all the control subjects. TheGoF approach was used to indicate how well the spatial maps fitted the DMN spatial template.17 We obtainedthe z-scores from the IC components by subtracting the mean and dividing by the standard deviation of thespatial map.17

3. RESULTS

Figure 2 shows two examples of DMNs extracted using the multiscale approach and the baseline method. Themultiscale approach recovered the DMN for both cases, where the baseline had a perturbed DMN detection.Figure 3 details how one DMN is changing through scales in the proposed method. When increasing the scales,the noise was reduced and the DMN spatial pattern was enhanced. Scale S4 had the highest contribution tothe DMN, with the highest GoF and α-weight values. The scale S8 resulted in a more uniform version of theDMN. However, some details of the components are lost, specially in the medial prefrontal cortex and part of leftinferior parietal lobe. The DMN computed with the baseline method is less consistent with the spatial patternstypically described for the DMN.8 In constrast, the one extacted using the multiscale approach resulted in amore recognizable spatial patterns. Interestingly, the DMN computed using the multiscale method preserved themain DMN regions obtained from the different scales reducing their spatial noise.

Baseline Multiscale

GoF = 0.228 GoF = 0.278

z-scores

x = 3, y =

-50, z = 22

(a) Subject 1

GoF = 0.153 GoF = 0.231

Baseline Multiscale

z-scores

x = 3, y =

-50, z = 22

(b) Subject 2

Figure 2. Results of the DMN detection for two subjects. The multiscale method recovered the DMN detection, comparedwith the baseline.

Figure 4 shows the GoF group results across different scales, the baseline and the multiscale approach. TheGoF mean value of all scales was around 0.25. The main differences were observed in the variation. The GoF ofthe scale S4 was higher than other scales, probably meaning that this scale had a stronger contribution of theDMN. The multiscale method had higher GoF mean value with less variability compared with the baseline.

Finally, table 1 details the α-weights obtained for each scale. All the scales had similar contributions to theDMN. However, the scale S4 provided a higher average contribution.

S2 S4 S8

Baseline Multiscale

GoF,,=,0.195 GoF,,=,0.233 GoF,,=,0.177

GoF,,=,0.130 GoF,,=,0.200

α,,=,,0.323 α,,=,0.385 α,,=,0.292

z-scores

x,=,3,,y,=

,-50,,z,=,22

Sca

les

Met

hod

s

Figure 3. DMN spatial z-maps for one subject. The multiscale recovered the DMN detection, as compared with thebaseline. The DMN connectivity was consistent in all scales and noise components were decreased with scales. The GoFvalue increased up until S4. The final DMN preserved the main DMN information of each scale.

S2 S4 S8 Bas. Mult.

0.15

0.2

0.25

0.3

0.35

0.4

GoF

Val

ue

Scales and Methods

GoF Value for scales and methods

Figure 4. Goodness of fit value of healthy subjects. Blue bars represent the results of each scale (S2, S4 and S8), followedby the results of the baseline preprocessing with the green bar and finally the multiscale method with the pink bar. Themultiscale method surpassed the GoF value of the baseline and had a similar values as the scales. The variation of themultiscale was lower than the results obtained with the baseline and with each independent scale.

Table 1. Measurements of the α-weights for each scale

Scales α-weightS2 0.33± 0.04S4 0.34± 0.04S8 0.33± 0.03

4. DISCUSSION

In this paper, we proposed a novel multiscale method to improve the detection of resting state neuronal net-works.19,30 The main idea was to decompose the data into different spatial scales, extracting the network ofinterests in each scale, and then mixing this information into a one representative network. By using a multi-scale representation, we expect to enhance the information of large scale interactions naturally expected for theresting state activity.22,23,33 In addition, we introduced prior information in the mixing process in form of the

α-weights. These weights aim to increase the contribution of informative scales into the final RSN network. Ourresults show that this strategy is helpful to improve quality in the resting state network detection.

Extraction of resting state networks at individual level is a critical step in the resting state analysis for clinicalapplications.17 The most commonly employed strategy consists of two steps:12 1) preprocessing and 2) networkextraction. The preprocessing steps aim to correct possible confounds than can affect the analysis, such as headmotion, anatomical variability, among others. The network extraction step uses a predefined network templateto compare independent components against, searching for the component with the highest degree of spatialsimilarity.12 Importantly, this procedure uses only one fixed scale, assuming that all the spatial interactionscan be captured at only one level of detail. Nevertheless, recent evidence suggests that brain activity hasdifferent levels of interactions depending on the scale.22 In this work, we exploited a multiscale representationto capture these different levels of interaction. Our results do not support the hypothesis that one specific scalehas significantly more information than other scales (Figs. 3,4 and Table 1). However, when the information ofdifferent scales was weighted and mixed, an improvement of information quality was observed as compared tosingle scale results (Fig. 4).

Based on these results, we conclude that the use of a multiscale representation for fMRI results in a morerobust detection of the DMN as compared to the information of only one scale. This methodology could improvethe neuronal detection of data that was perturbed by motion artifacts or low neuronal activity due to mentaldisorders. Additionally, this methodology could be extended to studies of other resting state neuronal networks.

REFERENCES[1] Van Den Heuvel, M. P. and Hulshoff Pol, H. E., “Exploring the brain network: a review on resting-state

fmri functional connectivity,” European Neuropsychopharmacology 20(8), 519–534 (2010).[2] Buckner, R. L., Vincent, J. L., et al., “Unrest at rest: default activity and spontaneous network correlations,”

Neuroimage 37(4), 1091–1096 (2007).[3] Greicius, M. D., Krasnow, B., Reiss, A. L., and Menon, V., “Functional connectivity in the resting brain: A

network analysis of the default mode hypothesis,” Proceedings of the National Academy of Sciences of theUnited States of America 100, 253–258 (January 2003).

[4] Perlbarg, V. and Marrelec, G., “Contribution of exploratory methods to the investigation of extendedlarge-scale brain networks in functional mri: methodologies, results, and challenges,” Journal of BiomedicalImaging 2008, 1–14 (2008).

[5] Heine, L., Soddu, A., Gomez, F., Vanhaudenhuyse, A., Tshibanda, L., Thonnard, M., Charland-Verville,V., Kirsch, M., Laureys, S., and Demertzi, A., “Resting state networks and consciousness: alterations ofmultiple resting state network connectivity in physiological, pharmacological, and pathological consciousnessstates,” Frontiers in psychology 3 (2012).

[6] Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., and Shulman, G. L., “Adefault mode of brain function,” Proceedings of the National Academy of Sciences of the United States ofAmerica 98, 676–682 (January 2001).

[7] Fox, M. D., Snyder, A. Z., Vincent, J. L., Corbetta, M., Van Essen, D. C., and Raichle, M. E., “Thehuman brain is intrinsically organized into dynamic, anticorrelated functional networks,” Proceedings of theNational Academy of Sciences of the United States of America 102(27), 9673–9678 (2005).

[8] Buckner, R. L., Andrews-Hanna, J. R., and Schacter, D. L., “The brain’s default network,” Annals of theNew York Academy of Sciences 1124(1), 1–38 (2008).

[9] Gusnard, D. A. and Raichle, M. E., “Searching for a baseline: Functional imaging and the resting humanbrain,” Nature Reviews Neuroscience 2(10), 685–694 (2001).

[10] Fox, M. D. and Raichle, M. E., “Spontaneous fluctuations in brain activity observed with functional magneticresonance imaging.,” Nature reviews. Neuroscience 8, 700–711 (September 2007).

[11] Broyd, S. J., Demanuele, C., Debener, S., Helps, S. K., James, C. J., and Sonuga-Barke, E. J., “Default-modebrain dysfunction in mental disorders: a systematic review.,” Neuroscience and biobehavioral reviews 33,279–296 (March 2009).

[12] Greicius, M. D., Srivastava, G., Reiss, A. L., and Menon, V., “Default-mode network activity distin-guishes Alzheimer’s disease from healthy aging: Evidence from functional MRI,” Proceedings of the NationalAcademy of Sciences of the United States of America 101(13), 4637–4642 (2004).

[13] Buckner, R. L., Snyder, A. Z., Shannon, B. J., LaRossa, G., Sachs, R., Fotenos, A. F., Sheline, Y. I.,Klunk, W. E., Mathis, C. A., Morris, J. C., and Mintun, M. A., “Molecular, Structural, and FunctionalCharacterization of Alzheimer’s Disease: Evidence for a Relationship between Default Activity, Amyloid,and Memory,” J. Neurosci. 25(34), 7709–7717 (2005).

[14] Garrity, A. G., Pearlson, G. D., Mckiernan, K., Lloyd, D., Kiehl, K. A., and Calhoun, V. D., “Aberrantdefault mode functional connectivity in schizophrenia,” Am J Psychiatry 164, 450–457 (March 2007).

[15] Guldenmund, P., Vanhaudenhuyse, A., Boly, M., Laureys, S., and Soddu, A., “A default mode of brainfunction in altered states of consciousness.,” Archives italiennes de biologie 150(2-3) (2012).

[16] Boly, M., Tshibanda, L., Vanhaudenhuyse, A., Noirhomme, Q., Schnakers, C., Ledoux, D., Boveroux, P.,Garweg, C., Lambermont, B., Phillips, C., et al., “Functional connectivity in the default network duringresting state is preserved in a vegetative but not in a brain dead patient,” Human brain mapping 30(8),2393–2400 (2009).

[17] Vanhaudenhuyse, A., Noirhomme, Q., Tshibanda, L. J.-F., Bruno, M.-A., Boveroux, P., Schnakers, C.,Soddu, A., Perlbarg, V., Ledoux, D., Brichant, J.-F., et al., “Default network connectivity reflects the levelof consciousness in non-communicative brain-damaged patients,” Brain 133(1), 161–171 (2010).

[18] Calhoun, V. D., Adali, T., Pearlson, G. D., and Pekar, J. J., “A method for making group inferences fromfunctional mri data using independent component analysis,” Human brain mapping 14(3), 140–151 (2001).

[19] Damoiseaux, J., Rombouts, S., Barkhof, F., Scheltens, P., Stam, C., Smith, S. M., and Beckmann, C.,“Consistent resting-state networks across healthy subjects,” Proceedings of the National Academy of Sci-ences 103(37), 13848–13853 (2006).

[20] Mikl, M., Marecek, R., Hlustık, P., Pavlicova, M., Drastich, A., Chlebus, P., Brazdil, M., and Krupa, P.,“Effects of spatial smoothing on fmri group inferences,” Magnetic resonance imaging 26(4), 490–503 (2008).

[21] Friston, K. J., Holmes, A. P., Poline, J., Grasby, P., Williams, S., Frackowiak, R. S., and Turner, R.,“Analysis of fmri time-series revisited,” Neuroimage 2(1), 45–53 (1995).

[22] Ball, T., Breckel, T. P., Mutschler, I., Aertsen, A., Schulze-Bonhage, A., Hennig, J., and Speck, O., “Variabil-ity of fmri-response patterns at different spatial observation scales,” Human brain mapping 33(5), 1155–1171(2012).

[23] Lindeberg, T., Lidberg, P., and Roland, P. E., “Analysis of brain activation patterns using a 3-d scale-spaceprimal sketch,” Human brain mapping 7(3), 166–194 (1999).

[24] Mckeown, M. J., Makeig, S., Brown, G. G., Jung, T.-P., Kindermann, S. S., Kindermann, R. S., Bell, A. J.,and Sejnowski, T. J., “Analysis of fmri data by blind separation into independent spatial components,”Human Brain Mapping 6, 160–188 (1998).

[25] Friston, K., Ashburner, J., Frith, C. D., Poline, J.-B., Heather, J. D., Frackowiak, R. S., et al., “Spatialregistration and normalization of images,” Human brain mapping 3(3), 165–189 (1995).

[26] Soddu, A., Vanhaudenhuyse, A., Bahri, M. A., Bruno, M.-A., Boly, M., Demertzi, A., Tshibanda, J.-F.,Phillips, C., Stanziano, M., Ovadia-Caro, S., et al., “Identifying the default-mode component in spatial icanalyses of patients with disorders of consciousness,” Human brain mapping 33(4), 778–796 (2012).

[27] Bell, A. J. and Sejnowski, T. J., “An information-maximization approach to blind separation and blinddeconvolution,” Neural Computation 7(6), 1129–1159 (1995).

[28] De Martino, F., Gentile, F., Esposito, F., Balsi, M., Di Salle, F., Goebel, R., and Formisano, E., “Clas-sification of fmri independent components using ic-fingerprints and support vector machine classifiers,”Neuroimage 34(1), 177–194 (2007).

[29] Demertzi, A., Gomez, F., Vanhaudenhuyse, A., Tshibanda, L., Thonnard, M., Charland, V., Kirsch, M.,Gosseries, O., Laureys, S., and Soddu, A., “Searching for patient-specific residual functional connectivitywith independent component analysis,” in [Human Brain Mapping ], (2012).

[30] Marrelec, G., Bellec, P., and Benali, H., “Exploring large-scale brain networks in functional mri,” Journalof Physiology-Paris 100(4), 171–181 (2006).

[31] Worsley, K. J., Marrett, S., Neelin, P., and Evans, A., “Searching scale space for activation in pet images,”Human brain mapping 4(1), 74–90 (1996).

[32] Scouten, A., Papademetris, X., and Constable, R., “Spatial resolution, signal-to-noise ratio, and smoothingin multi-subject functional mri studies,” Neuroimage 30(3), 787–793 (2006).

[33] Lindeberg, T., [Scale-space theory in computer vision ], Kluwer Academic Print on Demand (1993).

Bibliography

[1] Andersson, Jesper L. ; Hutton, Chloe ; Ashburner, John ; Turner, Robert ; Friston, Karl: Modelinggeometric deformations in EPI time series. En: Neuroimage 13 (2001), Nr. 5, p. 903–919

[2] Andronache, Adrian ; Rosazza, Cristina ; Sattin, Davide ; Leonardi, Matilde ; D’Incerti, Ludovico; Minati, Ludovico: Impact of functional MRI data preprocessing pipeline on default-mode networkdetectability in patients with disorders of consciousness. En: Frontiers in Neuroinformatics 7 (2013), Nr.16. – ISSN 1662–5196

[3] Ball, Tonio ; Breckel, Thomas P. ; Mutschler, Isabella ; Aertsen, Ad ; Schulze-Bonhage, Andreas; Hennig, Jurgen ; Speck, Oliver: Variability of fMRI-response patterns at different spatial observationscales. En: Human brain mapping 33 (2012), Nr. 5, p. 1155–1171

[4] Baquero, Katherine ; Gomez, Francisco ; Soddu, Andrea ; Vanhaudenhuyse, Audrey ; Demertzi,Athena ; Tshibanda, Jean-Flory ; Gosseries, Olivia ; Noirhomme, Quentin ; Laureys, Steven ;Romero, Eduardo: A Multi-Scale Analysis to set the Default Mode Network in noisy fMRI data. En:Frontiers in Computational Neuroscience (2011), Nr. 67. – ISSN 1662–5188

[5] Baquero, Katherine ; Gomez, Francisco ; Vanhaudenhuyse, Audrey ; Demertzi, Athena ; Tshibanda,Luaba ; Gosseries, Olivia ; Noirhomme, Quentin ; Laureys, Steven ; Romero, Eduardo ; Soddu,Andrea: An index for large head motion on resting state fMRI. En: Annual Meeting of the Organizationfor Human Brain Mapping, 2012

[6] Beauchamp, Michael S.: Detection of eye movements from fMRI data. En: Magnetic resonance inmedicine 49 (2003), Nr. 2, p. 376–380

[7] Bell, Anthony J. ; Sejnowski, Terrence J.: An Information-Maximization Approach to Blind Separationand Blind Deconvolution. En: Neural Computation 7 (1995), Nr. 6, p. 1129–1159

[8] Birn, Rasmus M. ; Bandettini, Peter A. ; Cox, Robert W. ; Jesmanowicz, Andrzej ; Shaker, Reza:Magnetic field changes in the human brain due to swallowing or speaking. En: Magnetic Resonance inMedicine 40 (1998), Nr. 1, p. 55–60

[9] Birn, Rasmus M. ; Smith, Monica A. ; Jones, Tyler B. ; Bandettini, Peter A.: The respirationresponse function: the temporal dynamics of fMRI signal fluctuations related to changes in respiration.En: Neuroimage 40 (2008), Nr. 2, p. 644–654

[10] Bisley, James W. ; Goldberg, Michael E.: Attention, intention, and priority in the parietal lobe. En:Annual review of neuroscience 33 (2010), p. 1

[11] Boly, M. ; Phillips, C. ; Tshibanda, L. ; Vanhaudenhuyse, A. ; Schabus, M. ; Dang-Vu, T. T. ;Moonen, G. ; Hustinx, R. ; Maquet, P. ; Laureys, S.: Intrinsic brain activity in altered states ofconsciousness: how conscious is the default mode of brain function? En: Annals of the New York Academyof Sciences 1129 (2008), p. 119–129. – ISSN 0077–8923

[12] Boly, Melanie ; Tshibanda, Luaba ; Vanhaudenhuyse, Audrey ; Noirhomme, Quentin ; Ledoux,Didier ; Boveroux, Pierre ; Garweg, Christophe ; Lambermont, Bernard ; Phillips, Andre ; Moonen,Gustave ; Bassetti, C ; Maquet, Pierre ; Laureys, Steven: Functional connectivity in the defaultnetwork during resting state is preserved in a vegetative but not in a brain dead patient. En: HumanBrain Mapping 30 (2009), p. 2393–2400. – ISSN 1065–9471

[13] Born, JD ; Albert, Adelin ; Hans, Pol ; Bonnal, J: Relative prognostic value of best motor responseand brain stem reflexes in patients with severe head injury. En: Neurosurgery 16 (1985), Nr. 5, p. 595–601

46

[14] Broyd, Samantha J. ; Demanuele, Charmaine ; Debener, Stefan ; Helps, Suzannah K. ; James,Christopher J. ; Sonuga-Barke, Edmund J.: Default-mode brain dysfunction in mental disorders: asystematic review. En: Neuroscience and biobehavioral reviews 33 (2009), March, Nr. 3, p. 279–296. –ISSN 0149–7634

[15] Buckner, Randy L. ; Andrews-Hanna, Jessica R. ; Schacter, Daniel L.: The brain’s default network.En: Annals of the New York Academy of Sciences 1124 (2008), Nr. 1, p. 1–38

[16] Buckner, Randy L. ; Snyder, Abraham Z. ; Shannon, Benjamin J. ; LaRossa, Gina ; Sachs, Rimmon ;Fotenos, Anthony F. ; Sheline, Yvette I. ; Klunk, William E. ; Mathis, Chester A. ; Morris, John C.; Mintun, Mark A.: Molecular, Structural, and Functional Characterization of Alzheimer’s Disease:Evidence for a Relationship between Default Activity, Amyloid, and Memory. En: J. Neurosci. 25 (2005),Nr. 34, p. 7709–7717

[17] Buckner, Randy L. ; Vincent, Justin L. [u. a.]: Unrest at rest: default activity and spontaneous networkcorrelations. En: Neuroimage 37 (2007), Nr. 4, p. 1091–1096

[18] Calhoun, V. D. ; Adali, T. ; Pearlson, G. D. ; Pekar, J. J.: A method for making group inferencesfrom functional MRI data using independent component analysis. En: Human brain mapping 14 (2001),Nr. 3, p. 140–151

[19] Calhoun, VD ; Adali, T ; Pearlson, GD ; Pekar, JJ: Spatial and temporal independent componentanalysis of functional MRI data containing a pair of task-related waveforms. En: Human brain mapping13 (2001), Nr. 1, p. 43–53

[20] Calhoun, Vince D. ; Liu, Jingyu ; Adalı, Tulay: A review of group ICA for fMRI data and ICA for jointinference of imaging, genetic, and ERP data. En: NeuroImage 45 (2009), March, Nr. 1, p. S163–S172. –ISSN 10538119

[21] Carp, Joshua: Optimizing the order of operations for movement scrubbing: Comment on Power et al.En: NeuroImage 76 (2013), p. 436–438

[22] Caspers, Svenja ; Schleicher, Axel ; Bacha-Trams, Mareike ; Palomero-Gallagher, Nicola ;Amunts, Katrin ; Zilles, Karl: Organization of the human inferior parietal lobule based on recep-tor architectonics. En: Cerebral Cortex 23 (2013), Nr. 3, p. 615–628

[23] Damoiseaux, JS ; Rombouts, SARB ; Barkhof, F ; Scheltens, P ; Stam, CJ ; Smith, Stephen M.; Beckmann, CF: Consistent resting-state networks across healthy subjects. En: Proceedings of theNational Academy of Sciences 103 (2006), Nr. 37, p. 13848–13853

[24] De Martino, Federico ; Gentile, Francesco ; Esposito, Fabrizio ; Balsi, Marco ; Di Salle, Francesco ;Goebel, Rainer ; Formisano, Elia: Classification of fMRI independent components using IC-fingerprintsand support vector machine classifiers. En: Neuroimage 34 (2007), Nr. 1, p. 177–194

[25] Demertzi, Athena ; Gomez, Francisco ; Vanhaudenhuyse, Audrey ; Tshibanda, Luaba ; Thonnard,Marie ; Charland, Vanessa ; Kirsch, Murielle ; Gosseries, Olivia ; Laureys, Steven ; Soddu, Andrea:Searching for patient-specific residual functional connectivity with independent component analysis. En:Human Brain Mapping, 2012

[26] Diedrichsen, Jorn ; Shadmehr, Reza: Detecting and adjusting for artifacts in fMRI time series data.En: Neuroimage 27 (2005), Nr. 3, p. 624–634

[27] Drobnjak, Ivana: FMRI Simulator: Development and Applications, University of Oxford, Tesis deGrado, 2007

[28] Esposito, Fabrizio ; Formisano, Elia ; Seifritz, Erich ; Goebel, Rainer ; Morrone, Renato ;Tedeschi, Gioacchino ; Di Salle, Francesco: Spatial independent component analysis of functionalMRI time-series: To what extent do results depend on the algorithm used? En: Human Brain Mapping16 (2002), Nr. 3, p. 146–157

[29] Fox, Michael D. ; Raichle, Marcus E.: Spontaneous fluctuations in brain activity observed with func-tional magnetic resonance imaging. En: Nature reviews. Neuroscience 8 (2007), September, Nr. 9, p.700–711. – ISSN 1471–003X

[30] Fox, Michael D. ; Snyder, Abraham Z. ; Vincent, Justin L. ; Corbetta, Maurizio ; Van Essen,David C. ; Raichle, Marcus E.: The human brain is intrinsically organized into dynamic, anticorrelatedfunctional networks. En: Proceedings of the National Academy of Sciences of the United States of America102 (2005), Nr. 27, p. 9673–9678

47

[31] Frank, Lawrence R. ; Buxton, Richard B. ; Wong, Eric C.: Estimation of respiration-induced noisefluctuations from undersampled multislice fMRI data. En: Magnetic Resonance in Medicine 45 (2001),Nr. 4, p. 635–644

[32] Fransson, P.: How default is the default mode of brain function? Further evidence from intrinsic BOLDsignal fluctuations. En: Neuropsychologia 44 (2006), Nr. 14, p. 2836–2845. – ISSN 00283932

[33] Friston, K. J. ; Holmes, A. P. ; Worsley, K. J. ; Poline, J. P. ; Frith, C. D. ; Frackowiak, R.S. J.: Statistical parametric maps in functional imaging: A general linear approach. En: Human BrainMapping 2 (1994), Nr. 4, p. 189–210. – ISSN 1097–0193

[34] Friston, Karl ; Ashburner, John ; Frith, Christopher D. ; Poline, J-B ; Heather, John D. ; Frack-owiak, Richard S. [u. a.]: Spatial registration and normalization of images. En: Human brain mapping3 (1995), Nr. 3, p. 165–189

[35] Friston, Karl J. ; Holmes, Andrew P. ; Poline, JB ; Grasby, PJ ; Williams, SCR ; Frackowiak,Richard S. ; Turner, R: Analysis of fMRI time-series revisited. En: Neuroimage 2 (1995), Nr. 1, p.45–53

[36] Friston, Karl J. ; Williams, Steven ; Howard, Robert ; Frackowiak, Richard S. ; Turner, Robert:Movement-related effects in fMRI time-series. En: Magnetic resonance in medicine 35 (1996), Nr. 3, p.346–355

[37] Garrity, Abigail G. ; Pearlson, Godfrey D. ; Mckiernan, Kristen ; Lloyd, Dan ; Kiehl, Kent A.; Calhoun, Vince D.: Aberrant Default Mode Functional Connectivity in Schizophrenia. En: Am JPsychiatry 164 (2007), March, Nr. 3, p. 450–457. – ISSN 0002–953X

[38] Giacino, Joseph T. ; Kalmar, Kathleen ; Whyte, John: The JFK Coma Recovery Scale-Revised:measurement characteristics and diagnostic utility. En: Archives of physical medicine and rehabilitation85 (2004), Nr. 12, p. 2020–2029

[39] Greicius, Michael D. ; Krasnow, Ben ; Reiss, Allan L. ; Menon, Vinod: Functional connectivity inthe resting brain: A network analysis of the default mode hypothesis. En: Proceedings of the NationalAcademy of Sciences of the United States of America 100 (2003), January, Nr. 1, p. 253–258

[40] Greicius, Michael D. ; Srivastava, Gaurav ; Reiss, Allan L. ; Menon, Vinod: Default-mode networkactivity distinguishes Alzheimer’s disease from healthy aging: Evidence from functional MRI. En: Pro-ceedings of the National Academy of Sciences of the United States of America 101 (2004), Nr. 13, p.4637–4642

[41] Grootoonk, S ; Hutton, C ; Ashburner, J ; Howseman, AM ; Josephs, O ; Rees, G ; Friston, KJ; Turner, R [u. a.]: Characterization and correction of interpolation effects in the realignment of fMRItime series. En: NeuroImage 11 (2000), Nr. 1, p. 49–57

[42] Guldenmund, Pieter ; Stender, Johan ; Heine, Lizette ; Laureys, Steven: Mindsight: Diagnostics inDisorders of Consciousness. En: Critical care research and practice 2012 (2012)

[43] Guldenmund, Pieter ; Vanhaudenhuyse, Audrey ; Boly, Melanie ; Laureys, Steven ; Soddu, Andrea:A default mode of brain function in altered states of consciousness. En: Archives italiennes de biologie150 (2012), Nr. 2-3

[44] Gusnard, D A. ; Akbudak, E ; Shulman, G L. ; Raichle, M E.: Medial prefrontal cortex and self-referential mental activity: relation to a default mode of brain function. En: Proc Natl Acad Sci U S A98 (2001), Nr. 7, p. 4259–64. – ISSN 0027–8424

[45] Gusnard, Debra A. ; Raichle, Marcus E.: Searching for a baseline: Functional imaging and the restinghuman brain. En: Nature Reviews Neuroscience 2 (2001), Nr. 10, p. 685–694

[46] Heine, Lizette ; Soddu, Andrea ; Gomez, Francisco ; Vanhaudenhuyse, Audrey ; Tshibanda, Luaba ;Thonnard, Marie ; Charland-Verville, Vanessa ; Kirsch, Murielle ; Laureys, Steven ; Demertzi,Athena: Resting state networks and consciousness: alterations of multiple resting state network connectiv-ity in physiological, pharmacological, and pathological consciousness States. En: Frontiers in psychology3 (2012)

[47] Hyvarinen, Aapo ; Oja, Erkki: A Fast Fixed-Point Algorithm for Independent Component Analysis.En: Neural Computation 9 (1997), Nr. 7, p. 1483–1492

48

[48] Kruger, Gunnar ; Glover, Gary H.: Physiological noise in oxygenation-sensitive magnetic resonanceimaging. En: Magnetic Resonance in Medicine 46 (2001), Nr. 4, p. 631–637

[49] Laureys, Steven ; Boly, Melanie: The changing spectrum of coma. En: Nature clinical practice.Neurology 4 (2008), Oktober, Nr. 10, p. 544–546. – ISSN 1745–8358

[50] Laureys, Steven ; Boly, Melanie: Unresponsive Wakefulness Syndrome. En: Archives italiennes debiologie 150 (2012), Nr. 2/3, p. 31–35

[51] Laureys, Steven ; Owen, Adrian M. ; Schiff, Nicholas D.: Brain function in coma, vegetative state,and related disorders. En: Lancet Neurol 3 (2004), Nr. 9, p. 537–46. – ISSN 1474–4422

[52] Leech, Robert ; Braga, Rodrigo ; Sharp, David J.: Echoes of the brain within the posterior cingulatecortex. En: The Journal of Neuroscience 32 (2012), Nr. 1, p. 215–222

[53] Lemieux, Louis ; Salek-Haddadi, Afraim ; Lund, Torben E. ; Laufs, Helmut ; Carmichael, David:Modelling large motion events in fMRI studies of patients with epilepsy. En: Magnetic resonance imaging25 (2007), Nr. 6, p. 894–901

[54] Li, Kaiming ; Guo, Lei ; Nie, Jingxin ; Li, Gang ; Liu, Tianming: Review of methods for functionalbrain connectivity detection using fMRI. En: Comput Med Imaging Graph 33 (2009), Nr. 2, p. 131–9. –ISSN 1879–0771

[55] Lindeberg, Tony: Scale-space theory in computer vision. Kluwer Academic Print on Demand, 1993

[56] Lindeberg, Tony ; Lidberg, Par ; Roland, Per E.: Analysis of brain activation patterns using a 3-Dscale-space primal sketch. En: Human brain mapping 7 (1999), Nr. 3, p. 166–194

[57] Luo, Wen-Lin ; Nichols, Thomas E.: Diagnosis and exploration of massively univariate neuroimagingmodels. En: NeuroImage 19 (2003), Nr. 3, p. 1014–1032

[58] Maclaren, Julian ; Herbst, Michael ; Speck, Oliver ; Zaitsev, Maxim: Prospective motion correctionin brain imaging: a review. En: Magnetic Resonance in Medicine (2012)

[59] Mantini, D. ; Perrucci, M. G. ; Del Gratta, C. ; Romani, G. L. ; Corbetta, M.: Electrophysiologicalsignatures of resting state networks in the human brain. En: Proceedings of the National Academy ofSciences 104 (2007), Nr. 32, p. 13170–13175

[60] Marrelec, Guillaume ; Bellec, Pierre ; Benali, Habib: Exploring large-scale brain networks in func-tional MRI. En: Journal of Physiology-Paris 100 (2006), Nr. 4, p. 171–181

[61] Matthews, PM ; Jezzard, P: Functional magnetic resonance imaging. En: Journal of Neurology,Neurosurgery & Psychiatry 75 (2004), Nr. 1, p. 6–12

[62] Mazaika, P ; Whitfield-Gabrieli, S ; Reiss, Allan ; Glover, G: Artifact repair for fMRI data fromhigh motion clinical subjects. En: Annual Meeting of the Organization for Human Brain Mapping, 2007

[63] Mazaika, Paul K. ; Hoeft, Fumiko ; Glover, GH ; Reiss, Allan L.: Methods and Software for fMRIAnalysis of Clinical Subjects. En: NeuroImage 47 (2009), Nr. 1, p. S58

[64] Mazoyer, B. ; Zago, L. ; Mellet, E. ; Bricogne, S. ; Etard, O. ; Houde, O. ; Crivello, F. ; Joliot,M. ; Petit, L. ; Tzourio-Mazoyer, N.: Cortical networks for working memory and executive functionssustain the conscious resting state in man. En: Brain Research Bulletin 54 (2001), Nr. 3, p. 287 – 298. –ISSN 0361–9230

[65] Mckeown, Martin J. ; Makeig, Scott ; Brown, Greg G. ; Jung, Tzyy-Ping ; Kindermann, Sandra S.; Kindermann, Ra S. ; Bell, Anthony J. ; Sejnowski, Terrence J.: Analysis of fMRI Data by BlindSeparation Into Independent Spatial Components. En: Human Brain Mapping 6 (1998), p. 160–188

[66] Menon, Vinod ; Uddin, Lucina Q.: Saliency, switching, attention and control: a network model of insulafunction. En: Brain Structure and Function 214 (2010), Nr. 5-6, p. 655–667

[67] Mikl, Michal ; Marecek, Radek ; Hlustık, Petr ; Pavlicova, Martina ; Drastich, Ales ; Chlebus,Pavel ; Brazdil, Milan ; Krupa, Petr: Effects of spatial smoothing on fMRI group inferences. En:Magnetic resonance imaging 26 (2008), Nr. 4, p. 490–503

[68] Murphy, Kevin ; Birn, Rasmus M. ; Handwerker, Daniel A. ; Jones, Tyler B. ; Bandettini, Pe-ter A.: The impact of global signal regression on resting state correlations: are anti-correlated networksintroduced? En: Neuroimage 44 (2009), Nr. 3, p. 893

49

[69] Perlbarg, V. ; Marrelec, G.: Contribution of exploratory methods to the investigation of extendedlarge-scale brain networks in functional MRI: methodologies, results, and challenges. En: Journal ofBiomedical Imaging 2008 (2008), p. 1–14. – ISSN 1687–4188

[70] Power, Jonathan D. ; Barnes, Kelly A. ; Snyder, Abraham Z. ; Schlaggar, Bradley L. ; Petersen,Steven E.: Spurious but systematic correlations in functional connectivity MRI networks arise fromsubject motion. En: Neuroimage 59 (2012), Nr. 3, p. 2142–2154

[71] Power, Jonathan D. ; Barnes, Kelly A. ; Snyder, Abraham Z. ; Schlaggar, Bradley L. ; Petersen,Steven E.: Steps toward optimizing motion artifact removal in functional connectivity MRI; a reply toCarp. En: Neuroimage (2012)

[72] Raichle, Marcus E.: A brief history of human brain mapping. En: Trends in neurosciences 32 (2009),Nr. 2, p. 118–126

[73] Raichle, Marcus E. ; Gusnard, Debra A.: Intrinsic brain activity sets the stage for expression ofmotivated behavior. En: The Journal of Comparative Neurology 493 (2005), Nr. 1, p. 167–176

[74] Raichle, Marcus E. ; MacLeod, Ann M. ; Snyder, Abraham Z. ; Powers, William J. ; Gusnard,Debra A. ; Shulman, Gordon L.: A default mode of brain function. En: Proceedings of the NationalAcademy of Sciences of the United States of America 98 (2001), January, Nr. 2, p. 676–682. – ISSN0027–8424

[75] Raichle, Marcus E. ; Snyder, Abraham Z.: A default mode of brain function: A brief history of anevolving idea. En: NeuroImage 37 (2007), Nr. 4, p. 1083 – 1090. – ISSN 1053–8119

[76] Satterthwaite, Theodore D. ; Elliott, Mark A. ; Gerraty, Raphael T. ; Ruparel, Kosha ; Loug-head, James ; Calkins, Monica E. ; Eickhoff, Simon B. ; Hakonarson, Hakon ; Gur, Ruben C. ;Gur, Raquel E. [u. a.]: An improved framework for confound regression and filtering for control of motionartifact in the preprocessing of resting-state functional connectivity data. En: Neuroimage 64 (2012), p.240–256

[77] Satterthwaite, Theodore D. ; Wolf, Daniel H. ; Loughead, James ; Ruparel, Kosha ; Elliott,Mark A. ; Hakonarson, Hakon ; Gur, Ruben C. ; Gur, Raquel E.: Impact of in-scanner head motionon multiple measures of functional connectivity: relevance for studies of neurodevelopment in youth. En:Neuroimage 60 (2012), Nr. 1, p. 623–632

[78] Scouten, A ; Papademetris, X ; Constable, RT: Spatial resolution, signal-to-noise ratio, and smoothingin multi-subject functional MRI studies. En: Neuroimage 30 (2006), Nr. 3, p. 787–793

[79] Shulman, Gordon L. ; Fiez, Julie A. ; Corbetta, Maurizio ; Buckner, Randy L. ; Miezin, Francis M.; Raichle, Marcus E. ; Petersen, Steven E.: Common Blood Flow Changes across Visual Tasks: II.Decreases in Cerebral Cortex. En: Journal of Cognitive Neuroscience 9 (1997), October, Nr. 5, p. 648–663

[80] Smith, Stephen M. ; Fox, Peter T. ; Miller, Karla L. ; Glahn, David C. ; Fox, P M. ; Mackay, Clare E.; Filippini, Nicola ; Watkins, Kate E. ; Toro, Roberto ; Laird, Angela R. [u. a.]: Correspondence ofthe brain’s functional architecture during activation and rest. En: Proceedings of the National Academyof Sciences 106 (2009), Nr. 31, p. 13040–13045

[81] Soddu, A. ; Vanhaudenhuyse, A. ; Bahri, M.A. ; Bruno, M. ; Boly, M. ; Demertzi, A. ; Tshibanda,J. ; Phillips, C. ; Stanziano, M. ; Ovadia-Caro, S. ; Nir, Y. ; Maquet, P. ; Papa, M. ; Malach,R. ; Laureys, S. ; Noirhomme, Q.: Identifying the default-mode component in spatial IC analyses ofpatients with disorders of consciousness. En: Hum Brain Mapp (2011)

[82] Talairach, Jean ; Tournoux, Pierre: Co-planar stereotaxic atlas of the human brain. 3-Dimensionalproportional system: an approach to cerebral imaging. (1988)

[83] Van Den Heuvel, Martijn P. ; Hulshoff Pol, Hilleke E.: Exploring the brain network: a review onresting-state fMRI functional connectivity. En: European Neuropsychopharmacology 20 (2010), Nr. 8, p.519–534

[84] Van Dijk, Koene R. ; Sabuncu, Mert R. ; Buckner, Randy L.: The influence of head motion on intrinsicfunctional connectivity MRI. En: Neuroimage 59 (2012), Nr. 1, p. 431–438

[85] Vanhaudenhuyse, Audrey ; Noirhomme, Quentin ; Tshibanda, Luaba J-F ; Bruno, Marie-Aurelie; Boveroux, Pierre ; Schnakers, Caroline ; Soddu, Andrea ; Perlbarg, Vincent ; Ledoux, Didier; Brichant, Jean-Francois [u. a.]: Default network connectivity reflects the level of consciousness innon-communicative brain-damaged patients. En: Brain 133 (2010), Nr. 1, p. 161–171

50

[86] Wilke, Marko: An alternative approach towards assessing and accounting for individual motion in fMRItimeseries. En: Neuroimage 59 (2012), Nr. 3, p. 2062–2072

[87] Worsley, Keith J. ; Marrett, S ; Neelin, P ; Evans, AC: Searching scale space for activation in PETimages. En: Human brain mapping 4 (1996), Nr. 1, p. 74–90

[88] Zaitsev, M ; Dold, Christian ; Sakas, Georgios ; Hennig, Jurgen ; Speck, O: Magnetic resonanceimaging of freely moving objects: prospective real-time motion correction using an external optical motiontracking system. En: Neuroimage 31 (2006), Nr. 3, p. 1038–1050

51

Documents

Universidad Nacional de Colombia Nacional de Colombia Facultad de Medicina Method to detect the Default Mode Network in fMRI data corrupted with head motion Katherine Andrea Baquero