Research Article BrainK for Structural Image Processing ...downloads.hindawi.com/journals/cin/2016/1349851.pdf · Research Article BrainK for Structural Image Processing: Creating

Research ArticleBrainK for Structural Image Processing Creating ElectricalModels of the Human Head

Kai Li1 Xenophon Papademetris12 and Don M Tucker1

1Electrical Geodesics Inc 500 E 4th Avenue Eugene OR 97401 USA2Yale University New Haven CT 06520 USA

Correspondence should be addressed to Don M Tucker dtuckeregicom

Received 29 May 2015 Revised 8 September 2015 Accepted 17 March 2016

Academic Editor Christophe Grova

Copyright copy 2016 Kai Li et al This is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

BrainK is a set of automated procedures for characterizing the tissues of the human head from MRI CT and photogrammetryimages The tissue segmentation and cortical surface extraction support the primary goal of modeling the propagation of electricalcurrents through head tissues with a finite difference model (FDM) or finite element model (FEM) created from the BrainKgeometries The electrical head model is necessary for accurate source localization of dense array electroencephalographic (dEEG)measures from head surface electrodes It is also necessary for accurate targeting of cerebral structures with transcranial currentinjection from those surface electrodes BrainK must achieve five major tasks image segmentation registration of the MRI CTand sensor photogrammetry images cortical surface reconstruction dipole tessellation of the cortical surface and Talairachtransformation We describe the approach to each task and we compare the accuracies for the key tasks of tissue segmentationand cortical surface extraction in relation to existing research tools (FreeSurfer FSL SPM and BrainVisa) BrainK achieves goodaccuracy with minimal or no user intervention it deals well with poor quality MR images and tissue abnormalities and it providesimproved computational efficiency over existing research packages

1 Introduction

With the simple recording of the electroencephalogram(EEG) the brainrsquos electrical activity can be measured withmillisecond temporal resolution at the head surface Densearray EEG (dEEG) systems now allow up to 256 channels tobe applied quickly with full coverage of the head assessingthe fields from the basal as well as superior cortical surface[1 2] The cortex with its laminar neural organizationand with locally synchronous activity stemming from itscolumnar organization is the primary generator of the farfields measured by head surface EEG [3] Cortical sourcescan be modeled as point dipoles and their contributionto surface activity can be reconstructed through electricalsource analysis The ambiguity of the inverse estimation inelectrical source analysis can be minimized if the preciselocations and orientations of the cortical sources are wellspecified To a first approximation the source dipoles can beassumed to be oriented perpendicular to the cortical surface

consistent with the orientation of the pyramidal neurons andcortical columns

To model cortical sources with these properties accuratecortical surface extraction is a key challenge Furthermorethe volume conduction of the electrical potentials fromthe cortex to the head surface must be specified throughcharacterizing the conductivity of each tissue compartmentThe skull is the primary resistive medium in the head and itmust be modeled preferably with bone density values fromCT Because the electrical boundary effects of the volumeconduction are affected by discontinuities in current pathsfor example caused by holes in the skull (optical canalsand foramen magnum) spherical shell or boundary elementmodels provide only approximate electrical propagation fromcortex to the surface and more detailed (FDM or FEM)volumetric models are needed Finally the position of theelectrodes must be specified accurately for example withgeodesic photogrammetry [4]

Hindawi Publishing CorporationComputational Intelligence and NeuroscienceVolume 2016 Article ID 1349851 25 pageshttpdxdoiorg10115520161349851

2 Computational Intelligence and Neuroscience

Figure 1 BrainK GUI

By describing the positions of electrodes and cortical sur-face targets an accurate electrical head model also supportsdense array approaches to transcranial current injectionThe electrical head model can be validated with boundedElectrical Impedance Tomography (bEIT) in which currentinjection and potential recovery are analyzed within thebounds of the tissue geometry fromMRI to test whether theelectrical conductivity of head tissues estimated by the modelis accurate in predicting the recovered potentials [5 6] Withan accurate electrical head model it is possible to modulatebrain activity noninvasively using dense array transcranialDirect Current Stimulation (tDCS) or transcranial Alternat-ing Current Stimulation (tACS) in which patterns of sourceand sink electrodes are computationally optimized to targetspecific cerebral sites [7 8]

In the present report we review the methods imple-mented in BrainK In addition we report validation studiesfor accuracy and efficiency of both tissue segmentation andcortical surface extraction Klauschen et al [9] evaluatedautomatic tools for head tissue segmentation from MRimages These included FSL [10] FreeSurfer [11 12] andSPM [13] Although BrainK is designed as an automatedtool visualization and editing capability is provided to allowadaptation to unique image properties such as the presenceof lesions or tissue anomalies

2 Overview of BrainK

Figure 1 shows the graphical user interface for BrainKincluding the multiple steps of image processing Althoughmaximal accuracy for EEG source localization or transcranialneuromodulation requires the full complement of MRI CTand sensor photogrammetry it is important to optimizethe results possible with the imaging data available for aparticular subject or patient Figure 2 shows the architectureof BrainK that incorporates specific workflows designed toadapt to the available data for the person Although it is

possible to assume standard sensor positions for BrainK(when sensor placement is referenced to skull fiducialsas with the Geodesic Sensor Net) a more accurate EEGsource localization workflow begins with sensor positionsfrom photogrammetry such as what is provided by theGeodesic Photogrammetry System (GPS) If no imaging datais available for the person an Atlas head model constructedfromadatabase ofMRIs andCTs for the appropriate ages [14]is then warped with nonlinear registration to fit the sensorpositions (and thus the personrsquos head shape) producing a con-formal Atlas head model If only a structural MRI (typically avolumetric T1) is available for the person the sensor positionsare registered to the MRI a database skull (from CT) isregistered to the MRI and tissue segmentation and corticalsurface extraction are conducted to create the individualhead model with Atlas skull The skull compartment includeseither the original CT Hounsfield units for estimating bonedensity (and thus conductivity) voxel by voxel or a single skullsegment depending on the FDM computational model thatwill use BrainKrsquos head model output If both the CT andMRIare available for the person the MRI is registered to the CTwhich has more accurate dimensions than the MRI to createthe individual head model with individual CT skull

For all workflows when the T1 MRI is available forthe person BrainK segmentation identifies the white matter(WM) and the gray matter (GM) It partitions these into twohemispheres and it differentiates cerebellum from cerebrumIn addition an entire headmask and the two eyeballs are sep-arated as well in the segmentation component The eyeballsmay be important for the electrical head model because ofthe large far fields generated by their cornea-retinal potentials(that must be separated from the brain signal in the EEG)For all workflow scenarios a generic spherical sensor cloudis warped onto the head contour of the subject For thescenarios of Atlas-to-MRI MRI-to-CT and CT-to-MRI anadditional GPS-to-head registration procedure is conductedto register the specific GPS sensor cloud from one individualGeodesic Sensor Net application onto the head contour Forthe scenarios of Atlas-to-MRI-to-GPS and Atlas-to-GPS theindividual GPS data has already played a role in the skullregistration and has been aligned with the head contour Inall scenarios the skull registration is conducted such thatthe resulting head segmentation includes the following tissuetypes WM GM CSF bone flesh and eyeball in which theWM and the GM are further partitioned into two cerebralhemispheres and the cerebellum

3 Methods in BrainK

31 Image Segmentation TheMRI segmentation implementsa cascade of automatic segmentation procedures in order toidentify and separate the head tissues required for electricalhead modeling including the scalp skull cerebrospinal fluid(CSF) brain (gray andwhitematter) and the eyeballs Finallycortical surface extraction is performed to allow characteri-zation of the normal of the cortical surface (allowing dipolesto be fit perpendicular to the cortex) The cascade mainlyconsists of two types of procedures voxel classification ona region of interest (ROI) and the extraction of certain

Computational Intelligence and Neuroscience 3

Top-level workflow

Scenario of Atlas-to-MRI

Scenario of MRI-to-CT

Scenario of CT-to-MRI

Scenario of Atlas-to-MRI-to-GPS

Scenario of Atlas-to-GPS

Segmentationand registration

MRIsegmentation

CTsegmentation

CTsegmentation

MRIsegmentation

Atlas-to-GPSregistration

Cortical surfacereconstruction

Atlas-to-MRIregistration

MRI-to-CTregistration

CT-to-MRIregistration

Atlas-to-MRI-to-GPSregistration

Generic sensorregistration

Dipolecreation

GPS-to-headregistration



Talairachtransformation





Figure 2 BrainK architecture

R1

R2

W1

R3 = W1 + G1

W2Hm

E1 + E2

Foregroundsegmentation

Brain tissueclassification

WM extraction

GM extraction


WM extraction

WM partition

GM extraction

Scalp extraction

Eyeball extraction

W2 + G2

Figure 3 MRI segmentation workflow

anatomical structures The classification procedure aims atlabeling all voxels in the ROI into different tissue types Theextraction procedure performs morphological operations onthe voxel classification so that a certain anatomical structuresuch as the WM the GM or the scalp is separated fromerroneous (false positive) segmentation results

Given the MRI data the segmentation cascade as shownin Figure 3 first takes the entire MR image space as the firstROI 1198771 and classifies all the voxels into two types fore-ground and background The following foreground extrac-tion procedure does simple morphological operations tofurther improve the foreground segmentation resulting in thesecond ROI 1198772 in which all voxels are classified into threetissue types WM GM and CSF as shown in Figure 4(b)The major operation performed in this procedure is BrainKrsquosnovel relative thresholding (RT) technique (Section 311) [15]Given the voxel classification on 1198772 the WM extractionprocedure then produces an initial WM volume 1198821 whichis used as the basis to identify the initial GM volume 1198661by the GM extraction procedure Next the union of 1198661 and1198821 is taken as the third ROI 1198773 in which the WMGMclassification is refined with a different RT scheme Thenew WM volume is then processed by the WM extraction

procedure and aWMpartition procedure ending upwith thesecond WM volume1198822

This WM partition procedure not only separates the twocerebral hemispheres from each other but also makes anoptimal cut between the cerebral WM and the cerebellarWMwith the well-knownmaximumflow algorithm [16]Theresult is specification of the following tissue types in1198822 twotypes of cerebral WM for the two hemispheres respectivelyand the cerebellarWM1198822 then forms the basis for extractingthe new GM volume 1198662 which includes the two types ofcerebral GM respectively for the two hemispheres and thecerebellar GM as shown in Figure 4(c) The brain tissuesegmentation (1198822 plus 1198662) is then taken as a reference datafor the scalp segmentation resulting in the head mask 119867119898The scalp segmentation then conducts a binary thresholdingprocedure on 1198771

To achieve the above segmentation steps BrainK imple-ments a novel morphological image analysis (SMIA) tech-nique (Section 313) and a cell complex based morphometricimage analysis (CCMIA) method (Section 314) These playkey roles in the WM extraction topology correction GMextraction and scalp extraction Together with the relativethresholding RT method the SMIA and CCMIA steps form


(a) (b) (c)

Figure 4 MRI segmentation (a) MR image (b) RT result on the foreground (c) MRI segmentation result

the core techniques used in BrainK and will be describedin the following subsections in detail The eyeball extractionfrom the head mask is a special procedure using a prioriknowledge about the eyeballs

Although segmentation of the skull from the MRI is animportant challenge the characterization of bone propertiesis poor in the typical T1 MRI sequence compared to that pro-vided by CTThe Hounsfield units provided by the CT image(measuring X-ray attenuation) are directly proportional tobone density and thus provide important information on therelative tissue properties of trabecular skull For example theCT Hounsfield units correlate 119903 = 083 with water content intrabecular skull [17]

The CT image segmentation procedure first performsthe simple yet robust thresholding [18] so that the entireimage is classified into three voxel types bone flesh andbackground The original Hounsfield units are then retainedas bone density and conductivity estimates Morphologyoperations follow to smooth the initial segmentation Finallythe maximum flow algorithm is applied to separate thebrain volume from the rest of the soft tissue so that the cutbetween them is minimized CT segmentation results in avoxel classification as well as a brain volume wrapped up bythe cranium and eventually there are four tissue types in thesegmentation bone flesh brain and background

311 Relative Thresholding (RT) A well-known and promi-nent image artifact that challengesMR image segmentation isthe intensity inhomogeneity (IIH) due to spatial distortionsin the radio frequency (RF) gain in the RF coil [19] Thepresence of IIH leads to a shading effect over the imageand to significant overlap between histograms of differenttissues The result is that intensity based methods such asthresholding and clustering are typically unreliable in braintissue segmentation IIH correction can be performed priorto image segmentation but the procedure may eliminate notonly the image artifacts but also image signals that provideimportant information When the correction is performedtogether with the segmentation additional degrees of free-dom are introduced into the problem formulation makingthe optimization procedure more susceptible to local optima

In this paper we present a new segmentationmethod referredto as relative thresholding (RT) which uses two global relativethresholds to compare with the local intensity contraststhereby bringing global and local information to segment theROI into WM GM and CSF

The nature of RT makes it robust against the influenceof IIT without introducing an explicit IIH correction Liketraditional thresholding methods RT enables exhaustivesearching without possibilities of being trapped into localoptima It also incorporates various a priori knowledgesuch that it is robust to intersubject variability and to theconvolution of cortical structures Finally RT conducts braintissue segmentation without the need of a prior step of brainvolume extraction

In spite of intersubject variability complexity of corticalstructures and variability on MR imaging sequences wecan make the following a priori structural geometrical andradiological observations (1) Skull CSF GM andWM forma layered structure from outside to inside (2) the averageintensities of skull CSF GM and WM in local regions are inascending order in T1-weightedMRI (3) the cortex thicknessis nearly uniform and is a very small value compared tothe size of brains in addition to the structural modelingdescribed above we also formulate an image model byincorporating a multiplicative low-frequency bias field 119887119894 andan additive noise field 120588119894 as follows

119910119894 = 1198871198941199101015840119894 + 120588119894 (1)

where 119910119894 is the observed intensity at the 119894th voxel and 1199101015840119894 is the

unobservable true intensity without the influences of IIH andnoise

The structural modeling and the image modeling enabletwo procedures both of which compare the intensity of asubject voxel119909119894with that of a reference voxel119909119895 on its gradientpath which is a set of ordered voxels emanating from 119909119894 andfollowing the gradient direction at each voxel in the pathThefirst procedure GM-WM differentiation scans all voxels inthe ROI which are initialized asWM and labels them as GMif there is such a reference voxel 119909119895 that its distance from 119909119894 isless than a distance threshold 119889119892119908 and the ratio between their


smoothed intensities 119911119894119911119895 is less than a relative threshold 119905119892119908the second procedure CSF-GM differentiation scans eachGM voxel 119909119894 and relabels it as CSF if there is such aWMvoxel119909119896 on its gradient path whose distance from 119909119894 is less thananother distance threshold 119889119888119908 (119889119888119908 gt 119889119892119908) and the ratiobetween their smoothed intensities 119911119894119911119896 is less than anotherrelative threshold 119905119888119908 (119905119888119908 lt 119905119892119908)

In order to find the optimal relative thresholds BrainKexhaustively tries each pair among the Cartesian productof the candidate GMWM relative threshold set and thecandidate CSFWM relative threshold set and chooses thepair that minimizes an objective function ℎ = 119908119908ℎ119908 +119908119892ℎ119892 + 119908119888ℎ119888 In this equation ℎ119908 is a metric measuring thehomogeneity of theWM object and formulated as the sum ofsmoothed intensity differences of any pair of adjacent voxelslabeled asWM ℎ119892 and ℎ119888 are formulated in the same spirit forGM andCSF respectively119908119908119908119892 and119908119888 are three weightingcoefficients

The second RT scheme takes the optimal relative thresh-olds and differs from the first scheme on the question of howto determine the reference intensity compared to the intensityof the subject voxel The ROI 1198773 is first reset and then aninitial set1198820 of WM voxels are determined by thresholding1198773 with a traditional threshold that maximizes the objectivefunction 119891 = 119891119890 minus 119891119903 where 119891119890 is the sum of intensitydifferences of all 119898 pairs of adjacent voxels labeled as WMand non-WMand119891119903 is the sumof differences of the119898 pairs ofadjacent voxels labeled asWM and with the greatest intensitydifferences Aminimal cardinality of1198820 is set to improve therobustness The remaining region of 1198773 is then segmented asa procedure of iteratively dilating 1198820 by using the optimalGMWM relative threshold to compare the subject voxelrsquosintensity to a reference intensity computed by considering theWM voxels near the subject voxel

312 GM and WM Extraction The GM and WM extractionprocedures are responsible for extracting the GM and theWM structure from the raw voxel classification in theROI We describe GM extraction first assuming that WMextraction is already done The GM extraction utilizes thefollowing a priori knowledge (1) the thickness of cortex isnearly uniform (2) GM wraps around WM such that twotissues form a layered structure (3) the average gray level ofGM is lower than that of WM in any local region Given theextracted WM volume1198822 the GM volume is first created bya gradient flow process in which any voxel originally labeledas GM in the ROI is taken as the true GM if it can reachany WM voxels by following the gradient path emanatingfrom itself within a given distance threshold Each step inthe gradient path is determined by looking at the gradientdirection for each voxel This initial GM volume is thenfurther processed by traditional morphological operations toimprove the segmentation ending up with the GM volume1198662

WM extraction is a more challenging process than GMextraction because it is responsible for extraction of WMitself and it is also responsible for the brain volume extrac-tionThe following a priori knowledge forms the basis ofWM

extraction the WM is highly connected but it has no topo-logical defects Topological defects or false positives presentnarrow bottlenecks The extraction mainly consists of twosteps (1)WM localization in which the center of one of thehemispheres is determined (2) separation of WM from falsepositives The localization of the WM center is essentiallyfinding the WM voxel with the highest connectivity whichmeans that some quantitative measurement of connectivityshould be taken for all WM voxels Since topological defectssuch as a tunnel inside the WM can greatly influencethe geometric measurement of the connectivity the WMshould be first topology-corrected before the connectivitymeasurement Because the cortex surface (ribbon) must bedefined in relation to the correct topology of the WMWM topology correction is also essential for valid corticalsurface reconstruction BrainK introduces a skeletonizationbased morphological image analysis (SMIA) method and acell complex based morphometric image analysis (CCMIA)frameworkThe following sections describe how these meth-ods play central roles in WM extraction through achievingaccurate connectivity measurement and topology correction

313 Skeletonization Based Morphological Image Analysis(SMIA) BrainKrsquos SMIA framework consists of a surfaceskeletonization procedure and a curve skeletonization proce-dure both of which are based on its extensive topologicalpoint classification Given a binary image we say its fore-ground 119865 has a handle whenever there is a closed path in 119865that cannot be deformed through connected deformations in119865 to a single point A handle in 119865 is referred to as a tunnel inits complement 119865 A point in 119865 is simple if it can be added toor removed from 119865 without changing the topology of both119865 and 119865 A simple point[20] is a central concept in digitaltopology and is the basis of our definition of what we call athick-simple point A point 119875 is thick-simple with respect to 119865if it is simple with respect to119865 and the removal of119875 and any ofits neighbors from 119865 does not increase the number of tunnelsand number of connected components in 119865 A point is thin-simple if it is simple but not thick-simple and can be classifiedas thick-surface and thick-curve points Thick-simple pointscan also be further classified as surface-edge curve-end andother point types

There are two steps for BrainKrsquos surface skeletonizationbased on the above point classification thick-surface skele-tonization and thin-surface skeletonizationThe former resultsin a discrete surface with thickness of at most two voxelsand the latter results in a final thin surface skeleton of one-voxel thickness The thick-surface skeletonization iterativelyremoves boundary points of the object that are thick-simpleat the moment of removal The thin-surface skeletonizationiteratively removes boundary points of the thick-surfaceskeleton that are thick-surface points thick-surface-edgepoints or thick curve-end points at the moment of removalThe curve skeletonization procedure iteratively removes theboundary points of the object that are thick-curve points orthick-simple but not curve-end points at the moment ofremoval As a side product BrainKrsquos surface skeletonizationprocedure gives each skeleton point a depth metric as the


distance from the point to the boundary whereas the curveskeleton procedure gives each skeleton point a widenessmetric

If the curve skeletonization procedure is conducted ata certain scale then the points that become the thin-curvepoints can be checked if they are on a handle of the objectRemoval of such handle points then corrects the topologydefects by cutting the handle BrainK supports a multiscaletopology correction [21] method on the WM object so thatthe WM extraction is made more robust and then thetopology-corrected cortex can be generated based on theWM

314 Cell Complex Based Morphometric Image Analysis(CCMIA) CCMIA is motivated by the goal of representingthe true connectivity of an objectrsquos skeleton considering theconvolution of such objects as the cerebral WM whereneither the depth nor the wideness of the structure is goodenough for this need It is essentially a series of transforma-tions on a space called the cell complex [22] A cell complex isa topological space composed of points segments polygonspolyhedrons and the generation to polytope in any dimen-sion Given a set of voxels 119883 in the 3D binary image we canconstruct a 3-dimensional cell complex that is 3-complexby creating a point for each voxel an edge for every twoconnected points a triangle for every three interconnectededges and a tetrahedron for every four interconnected trian-gles CCMIA proceeds as transforming the 3-complex to a 2-complex with points segments and polygons only and thenthe 2-complex to a 1-complex with points and segments onlyThe 3-2 transformation iteratively removes those boundarytetrahedrons by dropping one face 119891 while a connectivitymetric and a depth metric are accumulated on each 1198911015840 ofthe remaining three faces for the removed tetrahedron asdepth(1198911015840) = depth(119891) + 119889(119891 1198911015840) while 119889 represents thedistance between the centers of the two faces In a similarspirit the 2-complex can be transformed to 1-complex andthe remaining edges can be set with two accumulatedmetricsas wideness(1198901015840) = max(widness(1198901015840)wideness(119890) + 119889(119890 1198901015840))and connectivity(1198901015840) = connectivity(1198901015840) + connectivity(119890) +119889(119890 1198901015840)

After the raw WM object is topologically corrected itssurface skeleton is processed by CCMIA ending up with the1-complex with points and edges only The point incidentto the edge of the greatest connectivity is identified as theWM center and a wideness threshold is used to break thebottleneck between the true WM and the false positivesFinally the WM surface skeleton is dilated to restore thevolumetric object

32 Registration Given the varying data available flexibleand accurate registration is an essential BrainK functionWhereas the MR image accurately differentiates soft tissuesthe CT image accurately represents bone When individualCT or evenMRI is not available BrainK provides registrationof the sensor photogrammetry data to a digital head tissueAtlas of the appropriate age and gender BrainK integratesboth soft tissues and bones from multimodal images and

from existing digital Atlases by various image registrationtechniques

Image registration is a problem of finding optimal geo-metric transformations between images so that each pointof one image can be mapped to a corresponding point ofanother imageThere are generally rigid transformations andnonrigid ones Rigid transformation involves only translationand rotation Affine transformation is a typical nonrigidtransformation and allows scaling and shearing Anotherform of nonrigid transformation is an elastic transformationallowing local deformation based on models from elasticitytheory BrainK implements rigid and affine transformationsaswell as a landmark-based elastic transformation using thin-plate spline theory [23] For different registration purposeseither one or more of them is used so that the simple trans-formation serves as initial state for the more sophisticatedtransformation

Both the rigid transformation and the affine transforma-tion can be represented by a simple 3 times 4 matrix The keyissue in applying such transformations is to optimize thesimilarity measure between the source and the target imagewhich in BrainK is represented in terms of landmarks For thethin-plate spline transformation the displacement field at anypoint is computed by interpolation from the vectors definedby a set of source landmarks and a set of target landmarks Itcan be seen that landmark extraction plays a central role inall three transformations BrainK is optimized for extractinglandmarks based on the preceding image segmentation andon a priori anatomical knowledge All registration proceduresin BrainK undergo three steps (1) landmarks extraction (2)transformation coefficients determination and (3) transfor-mation execution Note that three transformations all uselandmarks but in different ways The landmarks in rigid andaffine transformation are used for constructing the similaritymeasurement while there are source landmarks and targetlandmarks in thin-plate spine transformations and they aremapped to build the displacement field

The thin-plate spline theory is based on an analogy to theapproximate shape of thin metal plates deflected by normalforces at discrete points Given 119899 source landmarks 119875119894 =(119909119894 119910119894 119911119894) in the source image and the corresponding targetlandmarks 1198751015840119894 = (119909

1015840119894 1199101015840119894 1199111015840119894 ) in the target image the thin-plate

spline transformation maps any point 119875 = (119909 119910 119911) in thesource image to the point 1198751015840 = (1199091015840 1199101015840 1199111015840) in the target imageas follows

1199091015840 = 1198860 + 1198861119909 + 1198862119910 + 1198863119911 +119899

sum119894=1

1198891198941199032119894 ln 1199032119894

1199101015840 = 1198870 + 1198871119909 + 1198872119910 + 1198873119911 +119899

sum119894=1

1198901198941199032119894 ln 1199032119894

1199111015840 = 1198880 + 1198881119909 + 1198882119910 + 1198883119911 +119899

sum119894=1

1198911198941199032119894 ln 1199032119894

(2)

where 1199032119894 = (119909 minus 119909119894)2 + (119910 minus 119910119894)

2 + (119911 minus 119911119894)2 The coefficients

in the above equations together form a (119899 + 4) times 3 matrix119882 which can be obtained by solving the equation 119871119882 = 119872that forces the matching of 119875119894 and 119875

1015840119894 where119872 is formed by


organizing the coordinates of1198751015840119894 = (1199091015840119894 1199101015840119894 1199111015840119894 ) and 119871 is formed

by composing the coordinates119875119894 = (119909119894 119910119894 119911119894) and the distanceinformation 1199032119894 ln 119903

2119894

When the individual CT as well as the MRI data isavailable for the person the registration component in theMRI-to-CT scenario first segments the CT into the followingtissue types bone brain flesh and background The MRIsegmentation is then aligned with the CT segmentation sothat the brain tissues and the eyeballs recognized in the MRIdata are properly transformed and put into the brain regiondefined by the CT data The result of this transformationis that the entire head segmentation (WM GM CSF boneeyeball and flesh) is now registered with the CT volumewhere the CSF is redefined as the ldquobrainrdquo region recognizedin the CT segmentation but now subtracting the WM andGM that were inserted from the MRI segmentation

When both MRI and CT are available the user couldalso choose the CT-to-MRI registration so that the bonerecognized in the CT is transformed and put into the MRIsegmentation wrapping around the brain and the CSF isinserted between the brain and the cranium However theMRI-to-CT registration is preferred when the CT covers theentire head of the subject because MRI imaging introducesmore geometry distortion than CT However when the CTonly covers the cranium region and the MRI data includesthe face and the jaw the user could consider the CT-to-MRIregistration in order to have adequate face and jaw data forthe electrical head model

When only theMRI data is available for the given subjectBrainK supports the skull Atlas-to-MRI registration inwhichthe bone tissue from an Atlas dataset is warped into the MRIspace and the CSF is also taken as those regions between thecranium and the brain Any region within the headmask thatis not occupied by the GM theWM the CSF the eyeballs orthe bone is set to flesh

When only the MRI and the individual GPS data areavailable and if the MRI is badly warped (dimensionally dis-torted) the user could choose the MRI-to-GPS registrationso that the head region is warped to match the GPS sensorcloud which has good metric properties of the actual headshape The result is transformed segmentation data whichcan be further processed by the Atlas-to-MRI registrationdescribed above

Thus the skull registration can be adapted to fit theavailable image data for the individual When only theindividual GPS data is available for a given subject anAtlas dataset with all required tissue types can be warpedto match the GPS sensor cloud with the so-called Atlas-to-GPS registration producing an individual conformal Atlasthat has the appropriate shape of the volume conductor forelectrical head modeling

321 MRICT Registration After MRI segmentation wealready have the MRI brain volume that is composed ofWM and GM but not CSF The CT brain volume occupiesspaces of WM GM and CSF By applying morphologicalclosing at a proper scale on the union of the MRI WMand GM we obtain the MRI brain mask 119861119898 which hassome inner CSF space filled and the brain contour smoothed

but still misses some CSF voxels mainly wrapping aroundthe superior contour of the brain MRI segmentation alsoproduces the MRI head mask119867119898 and the two eyeball masks1198641 and 1198642 Correspondingly we have the CT brain mask 119861119888and the CT headmask119867119888These data form the input set usedto determine the transformation coefficients of the MRICTregistration

MRICT registration undergoes a rigid transformationfollowed by an affine transformation Different sets of land-marks are extracted for the two transformations The rigidtransformation aims at only providing a good initial statefor the affine transformation and its landmarks are simplythose bordering voxels around 119861119898 The similarity measureto optimize the rigid transformation is the sum of squaresof the distances from all landmarks to the contour of 119861119888The optimal coefficients are found by multiscale exhaustivesearching with acceptable computational efficiency and goodglobal optimization performance

The landmarks for the affine transformation are moresophisticated and consist of two parts the inferior landmarksand the superior landmarks The inferior landmarks arefrom the border voxels of the inferior part of 119861119898 while thesuperior landmarks are from the superior contour of 119867119898The inferior and the superior landmarks are partitioned sothat they ldquovisuallyrdquo wrap up 119861119898 but do not overlap Themotivation of using two sets of landmarks is to align MRIwith CT in terms of both the brain matching and the scalpmatching However the brain volume in the CT may containsignificantly more CSF than that in the MRI at the superiorpart of the volume and we make the observation that theircontours have reliablematching only at the inferior part of thebrain The landmark extraction thus essentially involves thepartitioning of the cephalic space into the inferior part andthe superior part BrainK takes the following three landmarksas reference points to partition the cephalic space the centerof the brainmask 119861119898 and the centers of the two eyeball masks1198641 and 1198642 All three points can be automatically determined

The similarity measure for the affine transformation isformulated as the sumof two termsOne is the sumof squaresof the distances from all inferior landmarks to the contour of119861119888 and the other is the sumof squares of the distances from allsuperior landmarks to the contour of119867119888 A gradient-descentoptimization method is used to obtain the optimal affinecoefficients When all the optimal coefficients of both therigid and the affine transformation are obtained theMRICTregistration can be performed from one direction while theother can be achieved by simply inverting the coefficientmatrix The result of MRI-to-CT registration is shown inFigure 5

322 Skull Atlas-to-MRI Registration Since the Atlas and theindividual subject can have significant geometric differencesthe Atlas-to-MRI registration undergoes an elastic thin-plate spline transformation following the rigid and the affinetransformations The rigid transformation provides a goodinitial state for the affine transformation which in turn pro-vides a good initial state for the elastic local transformationThe result of the Atlas-to-MRI registration is illustrated inFigure 6


(a) (b)

Figure 5 MRI-to-CT registration (a) Transformed MRI (b) Fusion of MRI and CT segmentation

(a) (b)

Figure 6 Skull Atlas-to-MRI registration (a) Transformed MRI (b) Fusion of Atlas skull and MRI segmentation

The rigid transformation and the affine transformationuse the same set of landmarks to formulate the similaritymeasures Let 119861119898 be the MRI brain mask defined as inMRICT registration Let 119861119886 be the Atlas brain mask definedin the same way as 119861119898 The landmarks are then the setof border voxels of 119861119898 and the similarity measure is thesum of squares of distances from all landmarks to thecontour of 119861119886 The coefficients of the rigid transformationare found by multiscale exhaustive searching while those ofthe affine transformation are determined by gradient-descentsearching

The source landmarks used in the thin-plate splinetransformation include several parts The two primary partsare the inferior source landmarks and the superior sourcelandmarks Inferior source landmarks are distributed over theinferior part of the contour of the Atlas brain mask 119861119886 Thesuperior source landmarks are placed over the superior partof the contour of the Atlas head mask119867119886 Their extraction isvery similar to the extraction of those used in the MRICT

registration but they are much more sparsely distributed toreduce the computational overhead

The corresponding target landmarks in the subject headare determined automatically after the rigid and affine trans-formation are performed For each superior source landmarkthe corresponding superior target landmark is obtained asthe intersection of the contour of the subject head mask119867119898 and the line between the superior source landmark andthe subject brain center We have two ways to obtain theinferior target landmarks according to their locations Thefirst method is the same as that obtaining the superior targetlandmarks and works for the posterior part of the inferiorlandmarks that are distributed around the cerebellum For thesecondmethod that works for the anterior part of the inferiorlandmarks we compute a field of distance from the contour ofthe brain mask 119861119898 and each inferior target landmark whichshould be on the contour of 119861119898 is obtained by tracing fromthe corresponding source landmark following the gradient ofthe distance field


(a) (b) (c)

Figure 7 (a) Geodesic Photogrammetry System (GPS) (b) Identification of EEG sensors in the GPS software (c) Registration of EEG sensorswith the MRI head surface in BrainK

In the procedure of the registration the Atlas skullrsquosthickness is adjusted according to an estimated ratio of thesubject skullrsquos thickness to the Atlas skullrsquos thickness Thisis achieved by using another set of source landmarks andanother set of target landmarks For each superior sourcelandmark we determine its ldquomate landmarkrdquo over the innerside of the Atlas skull and on the line between the superiorsource landmark and theAtlas brain center For each superiortarget landmark there is also a ldquomate landmarkrdquo on the linebetween the superior target landmark and the subject braincenter For each pair of the superior source landmark 119878119894 andits mate landmark 1198781015840119894 we have a distance 119889119878119894 for each pairof the superior target landmark 119879119894 and its mate landmark1198791015840119894 we have a distance 119889119879119894 A global ratio is then defined as119889119879119894 119889119878119894 and its optimum is determined as the highest value so

that the majority (such as 90) of the superior target matelandmarks are off the mask 119861119898 For the few superior targetmate landmarks that are within 119861119898 according to the globalratio the global value is decreased locally so that they are offthe mask 119861119898

For each posterior inferior source landmark we deter-mine its mate landmark over the outside of the Atlas skulland on the line between the source landmark and the Atlas

brain center For each posterior inferior target landmark wealso have a mate landmark on the line between the targetlandmark and the subject brain center while its location isdetermined according to the global thickness ratio describedabove In order to further improve alignment performancethere are few other landmarks such as the centers of theeyeballs We have a specific algorithm to automatically detectthe subjectrsquos eyeballs but the details are not described here forthe limit of space When all source landmarks and all targetlandmarks are determined the transformation for any pointin the space can be calculated as shown in (2)

323 Registrations to the EEG Sensor Positions The EEGsensor positions are automatically localized in 3D withthe Geodesic Photogrammetry System (GPS) GPS capturesimages from 11 cameras in fixed positions to allow identifi-cation of the sensors in the images and then computationof the 3D coordinates (Figure 7) The Atlas-to-MRI-to-GPSregistration then consists of two stages the MRI-to-GPSregistration and then the Atlas-to-MRI registration usingthe newly transformed MRI Atlas-to-MRI registration hasbeen described in the above subsection and this subsectiononly deals with the MRI-to-GPS registration There are three


stepsThe first can be an automatic rigid transformation withother alternatives described below The second is an affinetransformation The first two can bring the MRI to a goodalignment state with the GPS data but the GPS points maystill be off the scalp The third step then performs thin-platespline transformation to deform the MRI so that each GPSpoint locates on the scalp The landmarks used in the firsttwo steps are the GPS points themselves and the similaritymeasure is the sum of squares of the distances from thelandmarks to the MRI scalp that is the contour of the headmask119867119898 For the thin-plate spline transformation the targetlandmarks are the GPS points while the source landmarks arethose points on the contour of119867119898 so that each one is on theline connecting the brain center and the corresponding targetlandmark

The first automatic step of the MRI-to-GPS registrationcan be replaced by a transformation with the user providingthree fiducial landmarks on the MRI space the nasion andthe left and right periauricular points (where the jaw hingemeets the skull) BrainK is able to stretch the MRI sothat the user-specified landmarks match the correspondingones in the GPS space This initial state can be directlyfed to the thin-plate spline transformation or it may passthrough an intermediate affine transformation BrainK alsoallows the user to specify only the nasion landmark anddoes rigid transformation in the first step while keeping thecorresponding nasion landmarks matched

The Atlas-to-GPS registration is very similar to the MRI-to-GPS registration except that the Atlas typically has greatergeometric differences from the GPS than the subjectrsquos ownMRIThe steps are similar to those for theMRI-to-GPS trans-formation just described For the MRI-to-GPS registrationhowever if they are to be used the fiducial landmarks haveto be manually marked on the MRI whereas the fiducialsare included in the digital Atlas dataset for the Atlas-to-GPS registration In addition one more fiducial landmarkthe vertex is used to accommodate the greater geometricdifference between the Atlas and the subject

The GPS-to-head registration can be seen as the inverseof the registration of MRI-to-GPS but no thin-plate splinetransformation is needed The initial transformation eitherautomatic rigid transformation or the stretching in termsof fiducial landmarks matching can put the GPS in agood initial state If needed an affine transformation canbe performed to further improve the alignment The finaladjustment then moves each GPS point onto the scalp if itis not on it yet with an end result that is similar to finding thecorresponding landmarks in the MRI-to-GPS registration

When GPS is not available a set of generic sensorpositions may be used such as those from the appropriateGeodesic Sensor Net (channel count and size) selected fromthe databaseThis registration has two steps First the genericspherical sensor cloud undergoes a translation transforma-tion so that the center of the sensor point cloud overlaps withthe center of the brain Second in a procedure that matchesthe physical conformation of the geodesic tension structureto the head each sensor is mapped to a point on the scalpby the ray emanating from the brain center to the sensorrsquosposition

33 Cortical Surface Reconstruction and Inflation The well-known marching-cube isosurface algorithm [24] is per-formed on the cerebral cortex volume of each hemisphereto reconstruct the cortical surface mesh In addition BrainKsupports cortical surface inflation by iteratively updating allpoints in the mesh such that the new position of each pointis the weighted sum of its old values with the weighted sumof the centroids of the triangles that are incident to the point

34 Dipole Tessellation The dipole tessellation operates intwo modes If the individualrsquos MRI is available the corticalsurface is extracted and tessellated into patches and adipole is fit to each patch If it is not available then dipoletriples (fitting 119909 119910 119911 components of the unknown dipoleorientation) are distributed evenly through the cortical graymatter

Whereas a trivial algorithm can distribute the triples theeven distribution of the oriented dipoles over the corticalsurface is formulated as a sophisticated graph partition [25]problem Given the cortical mesh from marching cubescomposed of a great number of interconnected tiny trianglesa graph is first constructed so that each vertex corresponds toa triangle and each edge incident to two vertices correspondsto the adjacency between the two corresponding trianglesThe graph partition algorithm then divides the entire graphinto a set of parts with the constraints that (1) the meshis divided into the same number of patches as the desirednumber of oriented dipoles (2) all patches are very compact(the more like to a square or circle the better) and (3) allpatches have similar areas The graph partition library Chaco[26] is used for the implementation

Given the mesh partition performed as above an equiva-lent oriented dipole is then defined for each patch generatedat the center of the patch with its orientation defined bythe mean of all surface normals of the triangles in thepatch weighted by their areas This equivalent dipole thusreflects the orientation of the electrical source if the entirepatch was synchronously active thereby serving as a usefulapproximation for the resolution of cortical activity that is setby the graph partitioning decision The cortical surface andthe oriented dipoles are displayed in Figure 8

35 Talairach Transformation The Talairach transformationcomponent brings all the data the segmentation the corticalsurfaces the sensors and the dipoles into the standardTalairach space [27 28] The transformation is determinedwith three points specified by the user the anterior com-missure (AC) the posterior commissure (PC) and one pointforming the middle plane with AC and PC The proceduremaintains the topology correctness of the cortical surface thatis guaranteed in either the MRI segmentation componentor the Atlas dataset and then maintained through eachregistration operation

4 Evaluation of BrainKrsquos Tissue Segmentationand Cortical Surface Extraction

BrainK provides certain algorithms such as skull bone den-sity (X-ray CT) image registration and cortical surface dipole


(a) (b)

Figure 8 Cortical surface (a) and tessellation of the surface intosim1 cm2 patches (about 1200 per hemisphere) eachwith an oriented equivalentdipole (b) Color is used here only to separate the cortical patches such that every patch has a different color from its neighbors

tessellation which are specifically important to model theelectrical properties of the human head As described belowin Section 8 efficient construction of accurate electricalvolume conduction head models is important to improvethe source analysis of human electrophysiological activitywith dense array technologies Accurate head tissue conduc-tivities and geometries are also important to optimize thetranscranial current delivery to stimulate the brain that is nowpossible with dense array technologies

In addition to these somewhat unique capabilitiesBrainK implements certain tissue segmentation and surfaceextraction algorithms that are useful in medical image pro-cessing generally and that can be evaluated in comparisonwith existing software that has become well known in theneuroimaging research community Examples of the softwareexamined here are FreeSurfer SPM FSL and BrainVisaTo assure accuracy for medical applications validation ofBrainKrsquos unique algorithms against well-known referencesoftware not only is scientifically useful but also forms anintegral part of medical quality verification and validationIn addition medical use requires that the functionality isfast and intuitive such that BrainK must meet standards ofefficiency and ease of use

Our current evaluation work pays particular attention tobrain tissue segmentation due to the fact that the accuracy ofbrain tissue segmentation is critical for a variety of medicaland neurological applications such as source localization andcortical dysplasia detection

41 Comparative Packages Below is a brief description ofthe four brain image analysis tools used for comparativeperformance evaluation Although these tools provide vary-ing functionality in neuroimage analysis they all supportautomatic T1-weighted human brain MR image segmenta-tion Our comparative evaluation is focused on the imagesegmentation task

411 FreeSurfer FreeSurfer [29 30] is a set of tools forreconstruction of cortical surfaces from structural MRI dataand for the overlay of functional data onto the reconstructedsurface FreeSurfer is developed in the Nuclear Magnetic

Resonance (NMR) Center Massachusetts General HospitalThe cortical surface reconstruction pipeline in FreeSurfermainly consists of three steps First a brain mask is extractedwith alignment of the structure MR image to the TalairachAtlas and the bias field is corrected Then the brain volumeis labeled as various cortical or subcortical structures in aprocedure based on both a subject-independent probabilisticAtlas and subject-specific measured values Finally the cor-tical surfaces are constructed from the prior segmentationwhich involves a topology correction procedure

412 SPM SPM (Statistical Parametric Mapping) is a sta-tistical technique for testing hypotheses about functionalimaging data [31] SPM also refers to the software developedby the Wellcome Department of Imaging NeuroscienceUniversity College London to carry out such analysis SPMfeatures structural MRI segmentation as well as a series offunctional neuroimage analysis Structural MRI segmenta-tion in SPM can be characterized as a circular procedurethat involves alternating three processing steps [32] a biascorrection step that corrects the intensity inhomogeneity aregistration step that normalizes the image to standard tissueprobability maps and a segmentation step that classifiesimage voxels into different tissue types As the segmentationresults SPM assigns each image voxel three probabilitieswith respect to three tissue types CSF GM and WM Ourexperiments were conducted with SPM5 which is old withrespect to the latest SPM12 but their implementation isbased on the same algorithm presented in [32] althoughthe newer version makes use of additional tissue classes andmultichannel segmentation and incorporates a more flexibleimage registration component as stated in the release notesof SPM12

413 FSL FSL (the FMRIB Software Library) is a collectionof functional and structural neuroimage analysis tools [33]For structural segmentation FSL applies the Brain ExtractionTool (BET) for segmenting brain from nonbrain regions instructural and functional data and FAST (FMRIBrsquos Auto-mated Segmentation Tool) for bias field correction and brainsegmentation into three tissue types CSF GM and WM


Structural MRI segmentation in FSL consists of two stepsusing BET to extract the brain and using FAST to classifytissue types BET performs skull stripping with a surfacemodel [34] The underlying method of FAST is based onan expectation-maximization algorithm combined with ahidden Markov random field (MRF) model [35] Due to theregularization of the MRF model FAST is supposed to bemore robust to noise than standard finite mixture modelbased methods

414 BrainVisa BrainVisa [36 37] is software developedat Service Hospitalier Frederic Joliot (SHFJ) that encom-passes an image processing factory and is distributed witha toolbox of building blocks dedicated to the segmentationof T1-weighted MR image Structural MRI segmentation inBrainVisa consists of four main steps First the user preparesthe data for segmentation by specifying several key landmarkpoints including the anterior commissure (AC) the posteriorcommissure (PC) an interhemispheric point and a lefthemisphere point A brain mask is then extracted includingonly white matter and gray matter integrating bias fieldcorrection [38] and histogram analysis [39] This is followedby a hemisphere partition and removal of cerebellum withmorphological image analysis [40] Finally cerebral graymatter and white matter are differentiated with histogramanalysis [41]

5 Datasets

The evaluation was performed on three image datasets a setof BrainWeb data with ground truth segmentation a set ofIBSR data with manually guided expert segmentation anda set of real scans of subjects with either mild cognitiveimpairment or Alzheimerrsquos disease

51 BrainWeb The BrainWeb dataset is a group of 8 realisticT1-weighted MR simulated images with ground truth seg-mentation provided by BrainWeb a simulated brain database[42 43] All 8 MR images are simulated on a normalanatomical model The resolution of the images is 1mm3 Inthe ground truth image all voxels in the image are segmentedinto the following tissue types Background CSF GM WMFat MuscleSkin Skin Skull Glial Matter and Connective

A variety of noise levels and levels of intensity inhomo-geneity (ie intensity nonuniformity (INU)) are artificiallyintroduced in the simulated images as listed in Table 1 Asstated in BrainWeb documentation [44] the ldquonoiserdquo in thesimulated images has Rayleigh statistics in the backgroundand Rician statistics in the signal regions The ldquopercentagenoiserdquo number represents the percent ratio of the standarddeviation of the white Gaussian noise versus the signal for areference tissueThe noise reference tissue used in our datasetis white matter The meaning of the intensity inhomogeneitylevel is as follows ldquoFor a 20 level the multiplicative INUfield has a range of values of 090ndash110 over the brain areaFor other INU levels the field is linearly scaled accordingly(eg to a range of 080ndash120 for a 40 level)rdquo According toBrainWeb the INU fields are realistic in that they are slowly

Table 1 Noise levels and IIH levels of the BrainWeb datasets

Dataset 1 2 3 4 5 6 7 8Noise level 3 3 5 5 7 7 9 9IIH level 20 40 20 40 20 40 20 40

varying fields of a complex shape and were estimated fromreal MRI scans

52 IBSR The IBSR dataset is a group of 18 T1-weighted realMR brain datasets Their manually guided expert segmenta-tion is included in the Internet Brain Segmentation Repos-itory (IBSR) supported by the Center for MorphometricAnalysis (CMA) at Massachusetts General Hospital [45]Theslice resolution of all datasets is 15mm and the XY resolutionvaries from 1mm2 to 0837mm2 The MR images have beenldquopositionally normalizedrdquo into the Talairach orientation butall five tools performed on this group of data assumedthat the images were not normalized The MR images werealso processed by the CMA bias field correction routinesbut it is not guaranteed that the intensity inhomogeneityis completely corrected All five tools therefore treated thedatasets as if no bias field correction had been performed

Each MR image was manually segmented into 44 indi-vidual structures including 3rd ventricle 4th ventricle Brain-stem and left and right accumbens area amygdala amygdalaanterior caudate cerebellum cortex cerebellum exteriorcerebellum white matter cerebral cortex cerebral exteriorcerebral white matter hippocampus Inf Lat vent lateralventricle pallidum putamen thalamus proper ventral DCand vessel

The 18 MR images represent various levels of imagequality To organize the evaluation by general image qualitywe divided the images into two subgroups the first 13 MRimages with good quality and 5 more MR images with badquality Note that the ordering of the IBSR datasets is differentfrom the original order A map of the order we used to theoriginal order is ldquo1 2 5 6 7 8 9 10 11 12 14 15 16 17 183 4 13rdquo For example when we refer to the 3rd dataset it isactually the 5th in the original order

53 Datasets Reflecting Neuropathology In addition to theBrainWeb and the IBSR datasets which were used for bothquantitative and qualitative evaluation we also tested fivetools on an auxiliary group of 8 real MR images scannedfrom subjects with minor cognitive impairment or thosewith Alzheimerrsquos disease The resolution of these datasetsis 1139 times 1211 times 1211mm3 The source of these datasetsis the Neurobiology Research Unit [46] in the UniversityHospital Rigshospitalet in Denmark No ground truth ormanual segmentation is provided for these datasets the issuefor the evaluation is primarily whether the abnormal brainscan be segmented with apparently reasonable accuracy

6 Quantitative Evaluation

In this section we present a quantitative evaluation onthe segmentation accuracy robustness and computational


efficiency of BrainK in comparison to other four packagesWe use the widely used Dice metric [1 4 46 47] as themeasurement for segmentation accuracy The standard devi-ation of the Dice metric provides a measure of segmentationrobustness Computational efficiency is measured as therequired run time of each package

61 Dice Metric Let TP refer to the number of true positivesFP to false positives and FN to false negatives The Dicemetric is then given by

Dice metric = 2 times TP2 times TP + FP + FN

(3)

TP FP and FN are measured versus the manual segmen-tation of real MR datasets or the ground truth of simulatedimages Note that when the segmentation is given as aprobability between 0 and 1 for each image voxel for eachtissue class (such as in the case of SPM) then TP FP andFN are calculated as the sum of the probabilities instead ofdiscrete counting

For quantitative evaluation using Dice metric we mustdecide the tissue type on which the metric is measuredBrainVisa only performs cerebrum segmentation while FSLand SPM segment the entire brain into CSF GM andWM without extraction of the cerebrum FreeSurfer alsoperforms segmentation on the whole brain but it segmentsthe brain into a greater number of tissue types (includingcerebral white matter and cerebral cortex) Currently BrainKperforms segmentation on the entire image and it is ableto differentiate cerebrum from cerebellum and one cerebralhemisphere from the other In our quantitative evaluation wemust calibrate the segmentation of the five packages within astandard framework so that common tissue types can be usedfor quantitative metric measurements

For the BrainWeb datasets we calibrated the segmenta-tion of five tools to the segmentation of cerebral WM andcerebral GM and measured the Dice metrics with respectto these two tissue types To enable this we manually par-titioned the ground truth whole brain (WM plus GM) at thebrainstem to extract the cerebral WM and the cerebral GMCerebral WM and cerebral GM also need to be extracted forthe SPM and FSL segmentation results We use a procedure(described in the next paragraph) that almost ldquoperfectlyrdquopartitions the segmentation results based on the ground truthpartition For BrainVisa and BrainK no transformation in thecalibration is required For FreeSurfer we just need to relabelall cerebral cortex voxels and all subcortical voxels excludingcerebral WM as cerebral gray matter

Let TP-Cerebrum and TP-Cerebellum respectivelydenote the set of true positives of cerebrum and cerebellumin the segmentation of SPM and FSL Let FP-Brain denotethe entire false positives including those in cerebrum andcerebellum The partition of the brain segmented by FSLand SPM is essentially the partition of FP-Brain into falsepositives in cerebrum and those in cerebellum which isdescribed as follows For each voxel V in FP-Brain if it has ashorter path in FP-Brain to TP-Cerebrum than any paths inFP-Brain from V to TP-Cerebellum then V is taken as a false

positive (of WM or GM) in cerebrum otherwise it is takenas a false positive (of GM or WM) in cerebellum

For the IBSR data we calibrated the segmentation of fivetools to the segmentation of cerebral cortex and cerebralWMand measured the Dice metrics with respect to these twotissue types These quantitative metrics give an evaluationon the accuracy of the cortical surface reconstruction whichdepend on segmentation of cerebral cortex and cerebralWM and are irrelevant to segmentation of subcortical graymatter tissues Since FreeSurfer explicitly labels cerebralcortex and cerebral white matter we do not need to do anytransformation in the calibration The calibration of FSL andSPM first conducts the brain partition to extract the cerebralWM and cerebral GM Given the set of cerebral WM andcerebral GM segmented by FSL SPM BrainVisa or BrainKwe measured the Dice metrics with respect to cerebral cortexand cerebral WM in the way described below

62 Experiment Settings

621 FreeSurfer We tested FreeSurfer on both the BrainWeband the IBSR datasets in a fully automatic mode withoutany user intervention An issue in collecting FreeSurfersegmentation results is the production of the cerebral cortexmask There is a so-called ldquoasegrdquo image and a ldquoribbonrdquoimage and both of these record voxels are labeled as cerebralcortex The ldquoribbonrdquo data is what FreeSurfer suggests touse but it has more false negatives than the ldquoasegrdquo data Incontrast the ldquoasegrdquo image is an intermediate result and ithas more false positives than the ldquoribbonrdquo data We applieda simple morphological closing operation on the union of thecortex ribbon and the subcortical structures so that certaintrue cerebral cortex voxels labeled in ldquoasegrdquo but missed inldquoribbonrdquo are covered This procedure apparently improvedthe performance of the cerebral cortex segmentation TheDice metric for the ldquoasegrdquo image was 07353 for the ldquoribbonrdquoimage was 08009 and for the ldquoclosedrdquo image was 08212

622 FSL In our first batch of experiments with FSL welet FSL automatically extract the brain and performed braintissue classification on both the BrainWeb and the IBSRdatasets However FSL generated poor results on the brainextraction and brain tissue classification on 6 of the IBSRdatasets (dataset 5 to dataset 10) In our second batch ofexperiments therefore we used different parameters in FSLobtained better brain masks for these datasets and thenrepeated the brain tissue classification The brain masksgenerated in the second batch of experiments were still notgood enough We therefore used the brain masks generatedby FreeSurfer for the brain tissue classification in FSL Thisdelivered the best FSL brain segmentation performance onthe 6 difficult IBSR datasets The three batches of exper-iments on FSL show that the brain extraction algorithmof FSL is not robust on the IBSR datasets but the braintissue segmentation itself performed well given good brainmasks The performance of FSL on the 6 IBSR datasets withrespect to the three batches of experiments is shown inTable 2


Table 2 Dice metrics collected for FSL using different brain masks

Brain masks Tissue types Dice metrics on 6 IBSR datasets5 6 7 8 9 10

Default FSLBrain mask

Cerebral cortex 06591 06806 07268 06887 07713 06762Cerebral WM 08845 08928 08891 08335 09095 08792

Customized FSLBrain mask


FreeSurferBrain mask


623 SPM In our first batch of experimental tests with SPMwe used the default parameters and let SPM automaticallyperform brain tissue segmentation on the BrainWeb andthe IBSR datasets In the second batch of experiments wechanged the parameter ldquobias regularizationrdquo from the defaultldquovery light regularizationrdquo to ldquomedium regularizationrdquo andreran SPM on the IBSR datasets SPM is supposed to be usedwith greater bias regularization when it is known a priori thatthere is less intensity inhomogeneity in the image Since theIBSR datasets were processed with bias field correction theuse of ldquomedium regularizationrdquo rather than the default ldquoverylight regularizationrdquo improved the performance of SPM onalmost all IBSR datasets For the cerebral cortex the meanDice metric over the 18 ISBR datasets was 07609 for verylight regularization and 07656 for medium regularizationThe mean Dice metric for white matter was 08572 for verylight regularization and 08707 for medium regularizationWe used the best regularization for each dataset in ourcomparative evaluation

624 BrainVisa We tested BrainVisa on both the BrainWeband the IBSR datasets automatically with the exception thatwe manually specified landmark points including the ACpoint the PC point an interhemispheric point and a lefthemisphere point BrainVisa produced an empty brain maskin the 9th IBSR dataset and was unable to generate brainmasks for the 13th and the 18th datasets In both cases weset the Dice metrics to be 0

625 BrainK BrainK was tested on the BrainWeb and theIBSR datasets fully automatically

63 Comparison Results

631 Segmentation Accuracy on the IBSR Datasets Wecollected the Dice metrics with respect to cerebral cortex(Table 3) and cerebral WM (Table 4) using the five softwarepackages on the IBSR datasets

On average BrainK performed best on cerebral cortexsegmentation on all 18 IBSR datasets good and bad andon the 13 good datasets exclusively In particular BrainKrsquoscerebral cortex performance is consistently better than thefour other packages on the 13 good datasets except for the4th dataset where BrainKrsquos performance is almost identicalto the best and the 12th dataset where BrainKrsquos performance

Table 3 Dice metrics of five tools with respect to cerebral cortex onthe IBSR datasets

IBSR datasets Dice metrics with respect to cerebral cortexBrainVisa SPM5 FreeSurfer FSL BrainK

1 07461 07705 08039 07803 086822 07953 08048 08175 08121 086193 07674 08080 08362 08361 087144 07233 08356 08641 08028 086125 02875 04621 07794 07312 086386 06610 06609 08068 07559 084417 07108 07065 07888 07898 086388 06982 07670 07800 07587 087909 0 07595 08128 08277 0870010 07707 06868 07729 07471 0861111 08688 08396 08702 08833 0863412 08596 08541 08458 08582 0877213 0 08588 08647 08554 0867314 08406 08426 08065 08429 0831515 08441 08439 08487 08381 0845716 08260 08263 08413 08426 0828117 08445 08443 08076 08278 0837918 0 08539 08337 08278 08379Mean 06247 07656 08212 08121 08557Mean on 13good datasets 06068 07365 08187 08030 08656

is close to the best The cerebral cortex segmentation perfor-mance of the five packages on the five bad datasets is similarexcept that BrainVisa generated empty brainmask for the 18thdataset

To provide an estimate of the statistical significance of thecomparison of BrainKwith the other packages we conductedpaired comparison 119905-tests across these 18 datasets BrainKwas more accurate at segmenting cerebral cortex with theISBR datasets than BrainVisa (119901 lt 002) SPM (119901 lt 0006)FreeSurfer (119901 lt 00005) and FSL (119901 lt 00009) If we omittedthe three datasets on which BrainVisa failed the comparisonsagain showed BrainK to be more accurate than BrainVisa(119901 lt 0002) SPM (119901 lt 0009) FreeSurfer (119901 lt 0001) andFSL (119901 lt 0002)


Table 4 Dice metrics of five tools with respect to cerebral WM onthe IBSR datasets

IBSR datasetsDice metrics with respect to cerebral WM

BrainVisa SPM5 FreeSurfer FSL BrainK

1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862

Mean on 13good datasets

06886 08685 08708 08999 08971

On average FSL performed best on cerebral WM seg-mentation on the 18 IBSR datasets and on the 13 gooddatasets However BrainKrsquos performance was very close toFSLrsquos performance in this comparison The performance ofthe five packages on the five bad datasets is similar except thatBrainVisa generated empty brain mask for the 18th datasetand FreeSurfer gave particularly poor performance for the14th dataset

632 Segmentation Robustness on the IBSR Datasets Wecalculated the standard deviations of the Dice metrics onthe IBSR datasets Together with the mean Dice metricsthese values indicate the segmentation robustness of the fivepackages on MR images scanned from different subjectsGreater mean Dice metric and lower standard deviationindicate greater robustness with respect to segmentationaccuracy

Two groups of standard deviations were calculated oneon the total 18 IBSR datasets and one on the 13 good IBSRdatasets BrainK demonstrated the lowest standard deviationof the software packages with respect to cerebral cortexon both the total 18 IBSR datasets and the 13 good IBSRdatasets as shown in Table 5 The lowest mean and standarddeviation Dice metric with respect to cerebral cortex indicatethat BrainK demonstrates the best accuracy and robustnesswith respect to cerebral cortex on the IBSR datasets Forcerebral WM BrainK and FSL performed very similarlywith respect to both the mean and the standard deviationof the Dice metric on both the total 18 IBSR datasets andthe 13 good IBSR datasets BrainK and FSL tied for thebest accuracy and robustness with respect to cerebral WMon the IBSR datasets Considering both cerebral cortex andcerebral WM these results suggest that in general BrainKdemonstrated the best overall performance on the IBSRdatasets

633 Segmentation Robustness with respect to Noise andIIH on the BrainWeb Datasets As described in Section 31the BrainWeb datasets vary in noise levels and intensityinhomogeneity (IIH) levelsThe performance inDicemetricsof the five packages with respect to cerebral GM and cerebralWM is listed in Tables 6 and 7

Among the five packages FreeSurfer demonstrated thelowest performance variation over different noise levels SPMandBrainVisa are similar in showing the highest performancevariations over different noise levels BrainK and FSL havemedium performance variations over different noise levelscompared to the other three packages Although FreeSurferperformed consistently over different noise levels it also gaveresults with the lowest accuracy on average

For each of the four noise levels we also tested thepackages on images with two different IIH levels All fivepackages gave little variation over different IIH levels Theonly exception is for BrainVisa with119873 = 9 and IIH = 40This is due to a poor brain mask

It is worth noting that in real MR scans the intensityinhomogeneity may occur in various and unknown patternsSuch noise could occur together with other difficulties thatmay be not present in the simulated BrainWeb datasetsTherefore we remark that our experiments with the Brain-Web dataset should not be taken to give a thorough andsufficient evaluation on the five packages with respect to theIIH robustness

634 Computational Efficiency The execution times of thefives packages tested on the IBSR datasets and the BrainWebdatasets are listed in Table 8 The experiments were allrun on a single 28Ghz Intel Xeon processor Among thefive packages BrainVisa took the least amount of timebut it also produced the lowest segmentation accuracy androbustness on the IBSR datasets FreeSurfer is well known forlong execution times However these long execution timescover segmentation of more subcortical structures as well asreconstruction of cortical surfaces It should be noted thatmost of the BrainK time was spent for topology correction


Table 5 Standard deviation of Dice metrics of five tools on the IBSR datasets

Sample groups Standard deviations

Datasets Tissue type BrainVisa SPM5 FreeSurfer FSL BrainK

All IBSR datasets Cerebral cortex 03155 01284 00308 00429 00192

Cerebral WM 03351 00505 00473 00230 00225

13 good datasets Cerebral cortex 03047 01414 00345 00476 00087

Cerebral WM 03305 00593 00482 00143 00147

Table 6 Dice metrics of five tools with respect to cerebral GM on the BrainWeb datasets

BrainWeb datasets Dice metrics with respect to cerebral GM

Noise level IIH level BrainVisa SPM5 FreeSurfer FSL BrainK

3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678

Table 7 Dice metrics of five tools with respect to cerebral WM on the BrainWeb datasets

BrainWeb datasets Dice metrics with respect to cerebral WM


3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289

Table 8 Computation times of five tools on the IBSR and the BrainWeb datasets

Datasets Computation times

BrainVisa SPM5 FreeSurfer FSL BrainK BrainK (topology correction)

IBSR 15m 34m 272 h 5m 17m 14m

BrainWeb 16m 20m 245 h 9m 21m 18m


which was not included in the execution times of the otherthree

7 Qualitative Evaluation

In this section we give a qualitative evaluation of the fivepackages based on the results of applying the packagesto the IBSR datasets the BrainWeb datasets and the 8pathological datasets with mild cognitive impairment orAlzheimerrsquos disease We first summarize and compare thesegmentation functionalities of the five packages followedby the discussion of their automaticity Finally we presentvarious segmentation abnormalities that we observed in theexperiments

71 Segmentation Functionalities The following is a summa-rization and comparison on the main segmentation featuresof the five packages

(i) Bias Field Correction Freesurfer SPM FSL andBrainVisa all integrate a bias field correction proce-dure either prior to tissue classification or combinedwith the classification BrainK on the other handdoes not need an explicit bias field correction becausethe relative thresholdingmethod is robust to bias fieldin arbitrary patterns

(ii) Brain Extraction FSL FreeSurfer and BrainVisaprovide separate tools for brain extraction (ie skullstripping) prior to brain tissue classification whereasSPM combines brain extraction with tissue classifi-cation BrainK on the other hand performs brainextraction after tissue classification Note that thebrainmask generated byBrainVisa ismeant to containonly GM andWMwhile the brain mask generated byFSL and FreeSurfer is meant to contain CSF as well asGM andWM

(iii) Tissue Classification FSL and SPM segment the brainvolume into three tissue types CSF GM and WMBrainVisa extracts cerebral WM and cerebral GMBrainK classifies the entire region of interest intoCSF GM and WM extracts the GM and WM andseparates cerebellum from cerebrum BrainVisa andBrainK also provide cerebral hemisphere partitionFreeSurfer segments a whole brain into 37 individualstructures including cerebral cortex cerebral WM aset of subcortical structures brainstem and cerebellarstructures

(iv) Cortical Surface Reconstruction BrainVisa Free-Surfer and BrainK support cortical surface recon-struction but FSL and SPM do not A core require-ment for cortical surface reconstruction is to insurethat the topology of the cortical surface is correct

(v) Multichannel Segmentation It is worth noting thatFSL and SPM12 provide multichannel segmentationsuch that T1- and T2-weighted images can be com-bined together for increased accuracy However T2-weighted images are oftennot available andwhen they

are available T2-weighted images often decrease thecontrast between grey and white matter and hencelead to bad performance in practice

72 Segmentation Automaticity and User Intervention Allfive packages support highly automatic brain segmentationwith little or no user intervention FreeSurfer allows theuser to start the cortical surface reconstruction without anyintervention In case the segmentation is not satisfactoryFreeSurfer supports interactive tools to allow the user tomodify the brain mask and to add control points to improvethe intensity normalization of WM a procedure that isextremely important for the performance of FreeSurferrsquoswhole brain segmentation FreeSurfer also supports inter-active tools for editing the final results generated by theautomatic processing

FSL also allows the user to start the segmentation withoutany intervention In FSL brain extraction and tissue classi-fication are performed respectively by the BET and FASTalgorithms The BET performance substantially influencesthat of FAST If the user is not satisfied with the brainextraction FSL allows the user to select different parametersand rerun BET However our experiments with BET onthe IBSR datasets and the pathological datasets show thatBET cannot guarantee good brain extraction even withuser intervention FAST has custom options for the user toselect whether to use the 119896-means segmentation or a prioriprobability maps for initial segmentation and to guide the 119896-means segmentation with manual intervention

In SPM brain segmentation can also be automaticallystarted with the default parameters and SPM often generatesgood results An important custom parameter of SPM is theone that controls the extent of bias field regularizationWhenany parameter is changed the segmentation procedure has tobe started over from scratch

BrainVisa requires the user to prepare the data by firstspecifying several landmark points including the AC pointthe PC point an interhemispheric point and a left hemi-sphere pointWhen the data is prepared by the user BrainVisaautomatically performs segmentation BrainVisa supportsinteractive tools for the user to edit the segmentation results

BrainK performs segmentation fully automatically butthe user also has the opportunity to tune the performanceby changing key parameters such as the relative thresholdsCompared to the user intervention mechanisms in the otherfour packages user intervention in BrainK is in the form ofglobal parameter selection and has the following advantagesFirst it is straightforward and requires little or no expertise tounderstand the meaning of the parameters and the criterionfor selecting optimal ones Second it is very easy to operate bysliding a value bar Third it is very efficient and the user canobtain the effect of parameter selection in real time for suchparameters as relative thresholds Fourth the parametershave global effect for segmentation and the user does not needto repeat similar operations for different local regions

73 Segmentation Abnormalities In examining the quality ofthe results we first attempted to understandwhy BrainK con-sistently achieved better performance than the other packages


(a) MR image (b) BrainVisa result (c) FreeSurfer result

(d) FSL result (e) SPM result (f) BrainK result

Figure 9 GM-shrinking phenomenon

with respect to cerebral cortex segmentation on the 13 goodIBSR datasets [48] We found that there was a ldquoshrinkingrdquoeffect on the cerebral cortex segmentation for FreeSurferFSL SPM and BrainVisa which gives rise to a significantnumber of false negatives This problem did not occur orwas much milder in BrainK We think the underlying reasonis that BrainK uses a new image modeling mechanism thatcan adapt to wider variations of GM intensities whereasthe statistical methods used in other packages were misledby these intensity variations resulting in missing many GMvoxels that had lower intensities

The contrasting cortex segmentation of the five packagesis shown for a representative IBSR dataset in Figure 9 Pinkstands for correct WM red for false WM negative lightgreen for correct GM very light green for false WM negativeand false GM negative dark green for false GM positiveand gray false for GM negative Note that for SPM lightgreen represents correct GM segmentation darker greenrepresents false GM negative and gray level represents falseGM positive

A common problem in FSL BrainVisa and SPM is thepoor brain extraction Some examples are shown in Figure 10Poor generation of the brain mask was the primary factor in

the very poor performance for SPM and BrainVisa on theIBSR datasets Since we used the considerably better brainmasks generated by FreeSurfer for the FSL analysis the FSLanalysis is not representative FreeSurfer did not encounterpoor skull stripping but an inaccurate brain mask may stillbe generated as shown in Figure 11(b) where some nonbrainvoxels with high intensities are taken as brain tissues BrainKon the other hand does not depend on a brain extractionpreprocessing step and robustly generated clean cerebrummasks as the union of the cerebral white matter and thecerebral gray matter on all tested datasets

We also found some other interesting abnormalities withFreeSurfer as shown in Figure 11 For example in Figure 11(c)FreeSurfer cut off a significant amount of cerebral WM andcortex at the top of the brain In Figure 11(d) FreeSurferwas unable to segment the complete lateral ventricle of thesubject with Alzheimerrsquos disease In Figure 12 FreeSurfergenerated poor GMWM segmentation even for a simulatedimage with excellent quality (noise level is 3 and IIHlevel is 20) while BrainK and FSL generated excellentresults These abnormalities we believe are due to theoverregularization of the a priori probability maps used inFreeSurfer


(a) BrainVisa brain mask (b) FreeSurfer brain mask

(c) FSL brain mask (d) SPM GMmask

Figure 10 Poor brain masks

8 Applications

At this point BrainK has built functional electrical headmodels for more than 200 persons Critical evaluation ofBrainKrsquos performance for generating segmented volumes andcortical surfaces continues in the applications of the ModalImage Pipeline EGIrsquos commercial software that incorporatesthe BrainK algorithms and workflows Validation tests ofphysical tissue boundaries are conducted in the routineworkflow for each head model construction A first step inthe FDM electrical lead field generation [49 50] EGIrsquos headphysics code is to test whether there are physically realisticsegmentation properties in the BrainK output required foraccurate electrical modeling For example there must be atleast one voxel of CSF between the outer gray matter of thecortex and the inner table of the skull

There have been many applications of BrainK in researchand development projects that have tested its capabilitiesand limitations For example although the skull Atlases areorganized by age and gender in a database to minimizethe warping required to fit an individualrsquos MRI tests havebeen conducted with skull CTs and MRIs of divergent shape(Figure 13) with good results

Electrical source analysis with children requires age-appropriate electrical head models Infants particularlyrequire unique models because of their small head size andhighly conductive skulls BrainK has proven effective inadapting to the unique challenges of processing infant MRIsand CTs Figure 14 shows results from a current project tocreate pediatric electrical head models directed by SergeiTurovets at EGI and by Linda Larsen-Prior and Fred Priorat University of Washington St Louis Figure 14(a) showsa normal infantrsquos segmented MRI with that infantrsquos CTregistered to the segmented MRI showing the thin skulland fontanel anterior to the vertex Figure 14(b) shows theextracted cortical surface and Figure 14(c) shows the dipoletessellated cortical surface prepared for electrical sourceanalysis

The FDM electrical head models created by BrainKallow accurate estimation of the delivery of current as wellas its measurement from the dense electrode array WithBrainKrsquos anatomical results providing the geometry for thehead physics code (Poisson equations for electromagneticpropagation in a finite difference model or FDM) lead fieldsare generated in theModal Image Pipeline head physics codedescribing the propagation from cortical sources to head


(a) MR image (b) MR image

(c) FreeSurfer result for MRI in (a) (d) FreeSurfer result for MRI in (b)

(e) BrainK result for MRI in (a) (f) BrainK result for MRI in (b)

Figure 11 FreeSurfer abnormalities on the pathological datasets compared to BrainK


(a) BrainWeb MRI N = 3 and IIH = 20 (b) FreeSurfer result (c) FSL result

(d) BrainK result

Figure 12 BrainK segmentation on the BrainWeb MRI compared to FreeSurfer and FSL

(a) (b)

Figure 13 Continued


(c) (d)

Figure 13 Skull Atlas-to-MRI registration with extensive warping of the skull to fit (a) A volunteer in a 256 Net (c) This personrsquos MRI (b)A database Atlas skull with a different shape (d) Warp of the Atlas database skull to the volunteerrsquos MRI

(a) (b) (c)

Figure 14 Electrical headmodel for an infant (a) Registration of skull fromCTwith segmentedMRI (b) Cortical surface extraction overlaidwith MRI (c) Dipole tessellation of cortical surface for electrical source analysis

surface electrodes (for EEG analysis) and from head surfaceelectrodes to cortical sources (for transcranial electricalneuromodulation) The results are visualized on BrainKrsquosanatomy and registered electrode positions in the reciprocityengine developed by ErikAnderson of EGI (Figure 15) In thisillustration cortical sources were chosen in the approximateregion of auditory cortex with the equivalent dipole foreach cortical patch visualized with a gold arrow (seen bestin Figure 15(b)) The lead field computation with the headphysics model then shows the resulting head surface fieldsin the color of both the scalp surface and the electrodes (redis positive blue is negative) Given this generated surfacefield the reciprocity theorem states that impressing currenton those cortical sources can be accomplished by applyingcurrent in the same surface pattern (red for anodal andblue for cathodal) The lead field from the head physicscode is again used to estimate the current density thatis impressed on the cortex by electrical neuromodulationwith this electrode pattern (Figures 15(c) and 15(d)) The

accuracy of the electrical solutions is fully dependent on thedetailed geometry provided by BrainK including not only theaccurate (lt1mm) registration of sensors to the head surfacebut also the specification of the orientation of the corticalpatches provided by BrainKrsquos characterization of the corticalsurface

9 Conclusion

BrainK is a software package that implements certain tissuesegmentation and surface extraction algorithms that areuseful in scientific and medical image processing It alsoapplies certain algorithms such as skull image registrationand cortical surface dipole tessellation which are specificallyimportant to model the electrical properties of the humanhead Accurate electrical head models are proving importantfor improving the source analysis of human electrophysiolog-ical activity with dense array technologies and for optimizingtranscranial current delivery to modulate brain activity


(a) (b)

(c) (d)

Figure 15 ((a) (b)) Using BrainKrsquos anatomy in the reciprocity visualization environment software to show the electrical field at the headsurface (positive potential at red electrodes and negative potential at blue electrodes) created by electrical activity in auditory cortex (dipolesshown as gold arrows seen best in the inflated cortex at (b)) ((c) (d))Through reciprocity theory the field pattern of the cortical sources canbe used to estimate the optimal electrodes (sources red sinks blue) to impress electrical current on the auditory cortex for neuromodulationHere the impressed cortical current density is imaged with a hot metal palette (gold orange and yellow) thresholded at 50 of maximum

Competing Interests

Kai Li and Don M Tucker are employees of EGI a medicaldevice company that is developing medical software from themethods described here EGI and the University of Oregonhold a US patent (no 8478011) on the relative thresholdingmethod Xenophon Papademetris is a paid consultant forEGI

Acknowledgments

The first author would like to thank his PhD advisors AllenMalony and Dr Don Tucker for their long and continuingsupport in the procedure of the research work Dr JasmineSong for her valuable help in revising the paper and DrPhan Luu Dr Sergei Turovets Dr DavidHammond ChelseaMattson Colin Davey Ryan Jayne and KyleMorgan for their

constructive comments advice and requests provided in theprocess of using and testing the BrainK software XenophonPapademetris worked on this project as a paid consultantof Electrical Geodesics Inc The intellectual property of thissoftware is protected by US Pat no 8478011

References

[1] D M Tucker ldquoSpatial sampling of head electrical fields thegeodesic sensor netrdquo Electroencephalography and Clinical Neu-rophysiology vol 87 no 3 pp 154ndash163 1993

[2] J Song D M Tucker T Gilbert et al ldquoMethods for examiningelectrophysiological coherence in epileptic networksrdquo Frontiersin Neurology vol 4 article 55 2013

[3] D L Schomer and F L da Silva Eds Niedermeyerrsquos Elec-troencephalography Basic Principles Clinical Applications andRelated Fields LippincottWilliams ampWilkins 6th edition 2011


[4] G S Russell K J Eriksen P Poolman P Luu andDM TuckerldquoGeodesic photogrammetry for localizing sensor positions indense-array EEGrdquo Clinical Neurophysiology vol 116 no 5 pp1130ndash1140 2005

[5] S Goncalves J C de Munck J P Verbunt R M Heethaarand F H da Silva ldquoIn vivo measurement of the brain andskull resistivities using an EIT-basedmethod and the combinedanalysis of SEFSEP datardquo IEEE Transactions on BiomedicalEngineering vol 50 no 9 pp 1124ndash1128 2003

[6] U Aydin J Vorwerk P Kupper et al ldquoCombining EEGand MEG for the reconstruction of epileptic activity using acalibrated realistic volume conductor modelrdquo PLoS ONE vol9 no 3 Article ID e93154 2014

[7] M Dannhauer R MacLeod S Guler et al ldquoComparison ofpatch electrodes and high density electrode arrays regardingvolume conduction in tdcs using a realistic head modelrdquo inProceedings of the 19th Annual Meeting of the Organization forHuman Brain Mapping (OHBM rsquo13) Seattle Wash USA June2013

[8] S Guler M Dannhauer R MacLeod et al ldquoDense electrodearray current optimization for targeting and directionality oftDCS studies in a realistic head modelrdquo in Proceedings of theOrganization for Human Brain Mapping Meeting (OHBM rsquo13)Seattle Wash USA June 2013

[9] F Klauschen A Goldman V Barra A Meyer-Lindenbergand A Lundervold ldquoEvaluation of automated brain MR imagesegmentation and volumetrymethodsrdquoHuman BrainMappingvol 30 no 4 pp 1310ndash1327 2009

[10] S M Smith M Jenkinson M W Woolrich et al ldquoAdvances infunctional and structural MR image analysis and implementa-tion as FSLrdquo NeuroImage vol 23 pp S208ndashS219 2004

[11] A M Dale B Fischl and M I Sereno ldquoCortical surface-based analysis I Segmentation and surface reconstructionrdquoNeuroimage vol 9 no 2 pp 179ndash194 1999

[12] B Fischl M I Sereno and A M Dale ldquoCortical surface-basedanalysis II inflation flattening and a surface-based coordinatesystemrdquo NeuroImage vol 9 no 2 pp 195ndash207 1999

[13] J Ashburner and K J Friston ldquoUnified segmentationrdquo Neu-roImage vol 26 no 3 pp 839ndash851 2005

[14] S Turovets A Salman AMalony P Poolman C Davey andDM Tucker ldquoAnatomically constrained conductivity estimationof the humanhead tissues in vivo computational procedure andpreliminary experimentsrdquo in Proceedings of the 7th Conferenceon Biomedical Applications of Electrical Impedance TomographySeoul South Korea 2006

[15] K Li A D Malony and D M Tucker ldquoAutomatic brain MRimage segmentation by relative thresholding and morphologi-cal image analysisrdquo in Proceedings of the 1st International Con-ference on Computer Vision Theory and Applications (VISAPPrsquo06) pp 354ndash364 Setubal Portugal 2006

[16] T H Cormen C E Leiserson R L Rivest and C SteinIntroduction to Algorithms MIT Press and McGraw-Hill 2ndedition 2001

[17] J Sierpowska M A Hakulinen J Toyras et al ldquoInterrela-tionships between electrical properties and microstructure ofhuman trabecular bonerdquo Physics in Medicine and Biology vol51 no 20 pp 5289ndash5303 2006

[18] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[19] A P Zijdenbos and B M Dawant ldquoBrain segmentation andwhite matter lesion detection in MR imagesrdquo Critical Reviewsin Biomedical Engineering vol 22 no 5-6 pp 401ndash465 1994

[20] G Bertrand ldquoSimple points topological numbers and geodesicneighborhoods in cubic gridsrdquo Pattern Recognition Letters vol15 no 10 pp 1003ndash1011 1994

[21] K Li A D Malony and D M Tucker ldquoA multiscale morpho-logical approach to topology correction of cortical surfacesrdquoinMedical Imaging and Augmented Reality Third InternationalWorkshop Shanghai China August 17ndash18 2006 Proceedings vol4091 of Lecture Notes in Computer Science pp 52ndash59 SpringerBerlin Germany 2006

[22] J-D Boissonnat and M Yvinec Algorithmic Geometry Cam-bridge University Press 1998

[23] M H Davis A Khotanzad D P Flamig and S E HarmsldquoA physics-based coordinate transformation for 3-D imagematchingrdquo IEEE Transactions on Medical Imaging vol 16 no3 pp 317ndash328 1997

[24] B K Natarajan ldquoOn generating topologically consistent isosur-faces from uniform samplesrdquo The Visual Computer vol 11 no1 pp 52ndash62 1994

[25] B L Chamberlain ldquoGraph partitioning algorithms for dis-tributing workloads of parallel computationsrdquo Tech Rep TR-98-10-03 Department of Computer Science amp EngineeringUniversity of Washington 1998

[26] B Hendrickson and R Leland ldquoThe Chaco users guide version20rdquo Tech Rep SAND 94-2692 Sandia National LaboratoriesAlbuquerque NM USA 1994

[27] J Talairach and P Tournoux Co-Planar Stereotaxic Atlas of theHuman Brain Thieme New York NY USA 1988

[28] J L Lancaster M GWoldorff L M Parsons et al ldquoAutomatedTalairach Atlas labels for functional brain mappingrdquo HumanBrain Mapping vol 10 no 3 pp 120ndash131 2000

[29] B Fischl D Salat E Busa et al ldquoWhole brain segmentationautomated labeling of neuroanatomical structures in the humanbrainrdquo Neuron vol 33 no 3 pp 341ndash355 2002

[30] Freesurfer httpssurfernmrmghharvardedu[31] SPM httpwwwfilionuclacukspm[32] J Ashburner and K J Friston ldquoUnified segmentationrdquo Neu-

roImage vol 26 no 3 pp 839ndash851 2005[33] FSL httpwwwfmriboxacukfsl[34] S M Smith ldquoFast robust automated brain extractionrdquo Human

Brain Mapping vol 17 no 3 pp 143ndash155 2002[35] Y Zhang M Brady and S Smith ldquoSegmentation of brain MR

images through a hidden Markov random field model andthe expectation-maximization algorithmrdquo IEEE Transactionson Medical Imaging vol 20 no 1 pp 45ndash57 2001

[36] Y Cointepas J-F Mangin L Garnero J-B Poline and HBenali ldquoBrainVISA software platform for visualization andanalysis of multi-modality brain datardquo in Proceedings of the 7thHuman Brain Mapping Conference p S98 Brighton UK 2001

[37] BrainVisa httpbrainvisainfo[38] J-F Mangin ldquoEntropy minimization for automatic correction

of intensity nonuniformityrdquo in Proceedings of the IEEE Work-shop on Mathematical Methods in Biomedical Image Analysis(MMBIA rsquo00) pp 162ndash169 Hilton Head Island SC USA June2000

[39] J-F Mangin O Coulon and V Frouin ldquoRobust brain segmen-tation using histogram scale-space analysis and mathematicalmorphologyrdquo in Proceedings of the Medical Image Computing


and Computer-Assisted Interventation (MICCAI rsquo98) W MWells A Colchester and S Delp Eds pp 1230ndash1241 MITBoston Mass USA 1998

[40] J-F Mangin J Regis and V Frouin ldquoShape bottlenecks andconservative flow systemsrdquo in Proceedings of the IEEESIAMWorkshop on Mathematical Methods in Biomedical Image Anal-ysis (MMBIA rsquo96) pp 319ndash328 San Francisco Calif USA 1996

[41] J-F Mangin O Coulon and V Frouin ldquoRobust brain segmen-tation using histogram scale-space analysis and mathematicalmorphologyrdquo in Medical Image Computing and Computer-Assisted InterventationmdashMICCAI rsquo98 W M Wells A Colch-ester and S Delp Eds vol 1496 of Lecture Notes in ComputerScience pp 1230ndash1241 1998

[42] C Cocosco V Kollokian R-S Kwan andA Evans ldquoBrainWebonline interface to a 3D MRI simulated brain databaserdquo Neu-roImage vol 5 no 4 p S425 1997

[43] BrainWeb httpwwwbicmnimcgillcabrainweb[44] CMAThe Center for Morphometric Analysis at Massachusetts

General Hospital httpwwwcmamghharvardeduibsr[45] NRU The Neurobiology Research Unit httpnrudk[46] D W Shattuck S R Sandor-Leahy K A Schaper D A

Rottenberg and RM Leahy ldquoMagnetic resonance image tissueclassification using a partial volume modelrdquo NeuroImage vol13 no 5 pp 856ndash876 2001

[47] K van Leemput F Maes D Vandermeulen and P SuetensldquoAutomated model-based tissue classification of MR images ofthe brainrdquo IEEE Transactions onMedical Imaging vol 18 no 10pp 897ndash908 1999

[48] K Li Neuroanatomical segmentation in MRI exploiting a prioriknowledge [PhD thesis] University of Oregon 2007

[49] A Salman S Turovets A Malony J Eriksen and DM TuckerldquoComputational modeling of human head conductivityrdquo in Pro-ceedings of the 5th International Conference on ComputationalScience (ICCS rsquo05) Atlanta Ga USA May 2005

[50] A Salman S Turovets A Malony et al ldquoNoninvasive con-ductivity extraction for high-resolution eeg source localizationrdquoAdvances in Clinical Neuroscience andRehabilitation vol 26 no1 pp 27ndash28 2005

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



Figure 1 BrainK GUI

By describing the positions of electrodes and cortical sur-face targets an accurate electrical head model also supportsdense array approaches to transcranial current injectionThe electrical head model can be validated with boundedElectrical Impedance Tomography (bEIT) in which currentinjection and potential recovery are analyzed within thebounds of the tissue geometry fromMRI to test whether theelectrical conductivity of head tissues estimated by the modelis accurate in predicting the recovered potentials [5 6] Withan accurate electrical head model it is possible to modulatebrain activity noninvasively using dense array transcranialDirect Current Stimulation (tDCS) or transcranial Alternat-ing Current Stimulation (tACS) in which patterns of sourceand sink electrodes are computationally optimized to targetspecific cerebral sites [7 8]

In the present report we review the methods imple-mented in BrainK In addition we report validation studiesfor accuracy and efficiency of both tissue segmentation andcortical surface extraction Klauschen et al [9] evaluatedautomatic tools for head tissue segmentation from MRimages These included FSL [10] FreeSurfer [11 12] andSPM [13] Although BrainK is designed as an automatedtool visualization and editing capability is provided to allowadaptation to unique image properties such as the presenceof lesions or tissue anomalies

2 Overview of BrainK

Figure 1 shows the graphical user interface for BrainKincluding the multiple steps of image processing Althoughmaximal accuracy for EEG source localization or transcranialneuromodulation requires the full complement of MRI CTand sensor photogrammetry it is important to optimizethe results possible with the imaging data available for aparticular subject or patient Figure 2 shows the architectureof BrainK that incorporates specific workflows designed toadapt to the available data for the person Although it is

possible to assume standard sensor positions for BrainK(when sensor placement is referenced to skull fiducialsas with the Geodesic Sensor Net) a more accurate EEGsource localization workflow begins with sensor positionsfrom photogrammetry such as what is provided by theGeodesic Photogrammetry System (GPS) If no imaging datais available for the person an Atlas head model constructedfromadatabase ofMRIs andCTs for the appropriate ages [14]is then warped with nonlinear registration to fit the sensorpositions (and thus the personrsquos head shape) producing a con-formal Atlas head model If only a structural MRI (typically avolumetric T1) is available for the person the sensor positionsare registered to the MRI a database skull (from CT) isregistered to the MRI and tissue segmentation and corticalsurface extraction are conducted to create the individualhead model with Atlas skull The skull compartment includeseither the original CT Hounsfield units for estimating bonedensity (and thus conductivity) voxel by voxel or a single skullsegment depending on the FDM computational model thatwill use BrainKrsquos head model output If both the CT andMRIare available for the person the MRI is registered to the CTwhich has more accurate dimensions than the MRI to createthe individual head model with individual CT skull

For all workflows when the T1 MRI is available forthe person BrainK segmentation identifies the white matter(WM) and the gray matter (GM) It partitions these into twohemispheres and it differentiates cerebellum from cerebrumIn addition an entire headmask and the two eyeballs are sep-arated as well in the segmentation component The eyeballsmay be important for the electrical head model because ofthe large far fields generated by their cornea-retinal potentials(that must be separated from the brain signal in the EEG)For all workflow scenarios a generic spherical sensor cloudis warped onto the head contour of the subject For thescenarios of Atlas-to-MRI MRI-to-CT and CT-to-MRI anadditional GPS-to-head registration procedure is conductedto register the specific GPS sensor cloud from one individualGeodesic Sensor Net application onto the head contour Forthe scenarios of Atlas-to-MRI-to-GPS and Atlas-to-GPS theindividual GPS data has already played a role in the skullregistration and has been aligned with the head contour Inall scenarios the skull registration is conducted such thatthe resulting head segmentation includes the following tissuetypes WM GM CSF bone flesh and eyeball in which theWM and the GM are further partitioned into two cerebralhemispheres and the cerebellum

3 Methods in BrainK

31 Image Segmentation TheMRI segmentation implementsa cascade of automatic segmentation procedures in order toidentify and separate the head tissues required for electricalhead modeling including the scalp skull cerebrospinal fluid(CSF) brain (gray andwhitematter) and the eyeballs Finallycortical surface extraction is performed to allow characteri-zation of the normal of the cortical surface (allowing dipolesto be fit perpendicular to the cortex) The cascade mainlyconsists of two types of procedures voxel classification ona region of interest (ROI) and the extraction of certain


Top-level workflow







MRIsegmentation

CTsegmentation

CTsegmentation

MRIsegmentation








Dipolecreation










R1

R2

W1

R3 = W1 + G1

W2Hm

E1 + E2



WM extraction

GM extraction


WM extraction

WM partition

GM extraction

Scalp extraction

Eyeball extraction

W2 + G2








(a) (b) (c)









119910119894 = 1198871198941199101015840119894 + 120588119894 (1)

























1199091015840 = 1198860 + 1198861119909 + 1198862119910 + 1198863119911 +119899

sum119894=1

1198891198941199032119894 ln 1199032119894

1199101015840 = 1198870 + 1198871119909 + 1198872119910 + 1198873119911 +119899

sum119894=1

1198901198941199032119894 ln 1199032119894

1199111015840 = 1198880 + 1198881119909 + 1198882119910 + 1198883119911 +119899

sum119894=1

1198911198941199032119894 ln 1199032119894

(2)








2119894













(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Top-level workflow







MRIsegmentation

CTsegmentation

CTsegmentation

MRIsegmentation








Dipolecreation










R1

R2

W1

R3 = W1 + G1

W2Hm

E1 + E2



WM extraction

GM extraction


WM extraction

WM partition

GM extraction

Scalp extraction

Eyeball extraction

W2 + G2








(a) (b) (c)









119910119894 = 1198871198941199101015840119894 + 120588119894 (1)

























1199091015840 = 1198860 + 1198861119909 + 1198862119910 + 1198863119911 +119899

sum119894=1

1198891198941199032119894 ln 1199032119894

1199101015840 = 1198870 + 1198871119909 + 1198872119910 + 1198873119911 +119899

sum119894=1

1198901198941199032119894 ln 1199032119894

1199111015840 = 1198880 + 1198881119909 + 1198882119910 + 1198883119911 +119899

sum119894=1

1198911198941199032119894 ln 1199032119894

(2)








2119894













(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b) (c)









119910119894 = 1198871198941199101015840119894 + 120588119894 (1)

























1199091015840 = 1198860 + 1198861119909 + 1198862119910 + 1198863119911 +119899

sum119894=1

1198891198941199032119894 ln 1199032119894

1199101015840 = 1198870 + 1198871119909 + 1198872119910 + 1198873119911 +119899

sum119894=1

1198901198941199032119894 ln 1199032119894

1199111015840 = 1198880 + 1198881119909 + 1198882119910 + 1198883119911 +119899

sum119894=1

1198911198941199032119894 ln 1199032119894

(2)








2119894













(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in
























1199091015840 = 1198860 + 1198861119909 + 1198862119910 + 1198863119911 +119899

sum119894=1

1198891198941199032119894 ln 1199032119894

1199101015840 = 1198870 + 1198871119909 + 1198872119910 + 1198873119911 +119899

sum119894=1

1198901198941199032119894 ln 1199032119894

1199111015840 = 1198880 + 1198881119909 + 1198882119910 + 1198883119911 +119899

sum119894=1

1198911198941199032119894 ln 1199032119894

(2)








2119894













(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in















1199091015840 = 1198860 + 1198861119909 + 1198862119910 + 1198863119911 +119899

sum119894=1

1198891198941199032119894 ln 1199032119894

1199101015840 = 1198870 + 1198871119909 + 1198872119910 + 1198873119911 +119899

sum119894=1

1198901198941199032119894 ln 1199032119894

1199111015840 = 1198880 + 1198881119909 + 1198882119910 + 1198883119911 +119899

sum119894=1

1198911198941199032119894 ln 1199032119894

(2)








2119894













(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






2119894













(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)


(a) (b)







(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b) (c)





















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in

















(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)













5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






5 Datasets

















(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







(3)


































1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



























1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







1 08652 08927 07964 08971 08789

2 08899 08940 08208 09160 08926

3 08649 08936 08138 09084 08874

4 08596 09013 08489 09168 09070

5 04097 07312 09240 08862 09101

6 07970 07422 09115 08953 08954

7 08255 08734 09147 08912 09001

8 08110 08923 09203 08914 09170

9 0 09059 09179 09146 09224

10 08457 08744 09069 09009 08790

11 08975 08927 08711 09142 08969

12 08858 09014 08099 08988 09006

13 0 08955 08647 08673 08748

14 08613 08678 07824 08632 08622

15 08541 08880 08746 08743 08637

16 08792 08948 08479 08933 08713

17 08619 08551 08592 08585 08463

18 0 08764 08124 08323 08463

Mean 06893 08707 08610 08900 08862


06886 08685 08708 08999 08971














Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








Cerebral WM 03351 00505 00473 00230 00225


Cerebral WM 03305 00593 00482 00143 00147




3 20 09292 09173 08333 09242 09084

40 09247 09189 08342 09268 09086

5 20 09197 08989 08323 09193 08908

40 09201 08998 08323 09193 08858

7 20 08628 08673 08320 09113 08816

40 08740 08713 08312 09127 08827

9 20 08166 08255 08259 08996 08658

40 07836 08301 08264 09019 08678




3 20 09550 09471 08849 09672 09558

40 09599 09533 08889 09664 09593

5 20 09552 09314 08824 09567 09494

40 09534 09315 08863 09581 09476

7 20 09325 08978 08779 09448 09382

40 09311 09008 08796 09467 09370

9 20 08926 08656 08757 09332 09296

40 08748 08701 08740 09354 09289




IBSR 15m 34m 272 h 5m 17m 14m
































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in

































8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in
















8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







8 Applications












(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in










(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





(d) BrainK result


(a) (b)

Figure 13 Continued


(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(c) (d)


(a) (b) (c)




9 Conclusion



(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)

(c) (d)


Competing Interests


Acknowledgments



References


























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in

























































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






















Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in










Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Documents

Research Article BrainK for Structural Image Processing ...downloads.hindawi.com/journals/cin/2016/1349851.pdf · Research Article BrainK for Structural Image Processing: Creating