Machine Learning of DTI Structural Brain Connectomes for Lateralization ofTemporal Lobe Epilepsy

Kouhei KAMIYA1*, Shiori AMEMIYA1, Yuichi SUZUKI2, Naoto KUNII3,

Kensuke KAWAI4, Harushi MORI1, Akira KUNIMATSU1, Nobuhito SAITO3,

Shigeki AOKI5, and Kuni OHTOMO1

1Department of Radiology, The University of Tokyo

7–3–1 Hongo Bunkyo, 113–8655 Tokyo, Japan2Department of Radiological Technology, The University of Tokyo Hospital

3Department of Neurosurgery, The University of Tokyo4Department of Neurosurgery, NTT Medical Center Tokyo

5Department of Radiology, Juntendo University School of Medicine

(Received March 12, 2015; Accepted June 10, 2015; published online September 4, 2015)

Background and Purpose: We analyzed the ability of a machine learning approach thatuses diffusion tensor imaging (DTI) structural connectomes to determine lateralization ofepileptogenicity in temporal lobe epilepsy (TLE).Materials and Methods: We analyzed diffusion tensor and 3-dimensional (3D) T1-

weighted images of 44 patients with TLE (right, 15, left, 29; mean age, 33.0 « 11.6 years)and 14 age-matched controls. We constructed a whole brain structural connectome for eachsubject, calculated graph theoretical network measures, and used a support vector machine(SVM) for classification among 3 groups (right TLE versus controls, left TLE versus con-trols, and right TLE versus left TLE) following a feature reduction process with sparselinear regression.Results: In left TLE, we found a significant decrease in local efficiency and the cluster-

ing coefficient in several brain regions, including the left posterior cingulate gyrus, leftcuneus, and both hippocampi. In right TLE, the right hippocampus showed reduced nodaldegree, clustering coefficient, and local efficiency. With use of the leave-one-out cross-validation strategy, the SVM classifier achieved accuracy of 75.9 to 89.7% for right TLEversus controls, 74.4 to 86.0% for left TLE versus controls, and 72.7 to 86.4% for left TLEversus right TLE.Conclusion: Machine learning of graph theoretical measures from the DTI structural

connectome may give support to lateralization of the TLE focus. The present good dis-crimination between left and right TLE suggests that, with further refinement, the classifiershould improve presurgical diagnostic confidence.

Keywords: brain connectome, diffusion tensor imaging, epilepsy, graph theory, machinelearning

Introduction

Temporal lobe epilepsy (TLE) is the most fre-quent type of refractory focal epilepsy. Patientswith concordant findings from electroencephalog-raphy (EEG), seizure semiology, neuropsychologi-

cal assessment, and magnetic resonance (MR)imaging—such as atrophy and fluid-attenuated in-version recovery (FLAIR) hyperintensity of thehippocampus ipsilateral to the side of seizure on-set—do extremely well with resection of the mesialtemporal structures.1 Therefore, establishing thelaterality of the epileptogenic focus with as muchcertainty as possible is an important task in the pre-operative evaluation of patients with TLE. Howev-

*Corresponding author, Phone: +81-3-5800-8666, E-mail:kkamiya-tky@umin.ac.jp

Magn Reson Med Sci, Vol. 15, No. 1, pp. 121–129, 2016©2015 Japanese Society for Magnetic Resonance in Medicine

MAJOR PAPER

doi:10.2463/mrms.2015-0027

121

er, standard MR imaging protocols may fail toshow an identifiable hippocampal abnormality ormay provide only subtle findings that remain in-conclusive. To increase the certainty of lateraliza-tion of the epileptogenic focus and obviate the needfor invasive intracranial electrode placement, priorstudies have investigated the utility of quantitativeor automated MR image analyses, including voxel-based morphometry,2,3 diffusion tensor imaging(DTI),3,4 and functional MR imaging (fMRI).5

DTI can approximate the white matter architec-ture by describing the directionality and magnitudeof water diffusion. In TLE, the decrease in fraction-al anisotropy (FA) tends to be maximal at the epi-leptic zone and subtle at a distance,6 and decreasedFA in the extra-temporal regions as well as withinthe ipsilateral temporal lobe suggests that the net-work is altered.7 Recently, in light of several stud-ies relating the presence of extratemporal abnor-malities to poor postsurgical outcome, emphasishas shifted to consider epilepsy as a disorder of awidespread brain network.8 Graph theoretical anal-ysis of brain connectomes has attracted much inter-est as a method of network analysis suitable for ep-ilepsy research.8,9 In graph theory terms, a brainnetwork, or connectome, consists of the set of neu-ral elements (nodes) and their interconnections(edges).10 Nodes usually represent brain regions,whereas edges represent (structural or functional)connections. A set of parameters that characterizespecific topological properties of the network canbe obtained from graph theoretical analyses of abrain connectome. Several studies have analyzedDTI-based structural connectomes in TLE; the ma-jority have reported altered connectivity to be mostprominent within the ipsilateral temporal lobe.11–15

In recent years, several studies have investigatedthe performance of machine learning algorithms,such as that of the support vector machine (SVM),for automatic localization of epileptogenic foci us-ing MR voxel-based morphometry (VBM)2,3 andfMRI.5 Because graph theory metrics use a subset

of numeric parameters to summarize the character-istic properties of huge and complex brain net-works, they are mathematically good candidates fora machine learning approach to identify the multi-variate feature combinations that best predict anoutcome of interest. A combination of machinelearning and connectomic measures has been usedto distinguish healthy individuals from patientswith disorders including autism,16 schizophrenia,17

and Alzheimer’s disease.18

In this context, we examined the performance ofa machine learning approach, used in combinationwith DTI-derived structural connectomes, to deter-mine lateralization of the epileptogenic focus of theTLE.

Materials and Methods

SubjectsWe retrospectively reviewed radiological records

of a single institution from 1 January 2007 to 31October 2014 and identified 44 patients with TLE(29 left, 15 right; 21 men, 23 women; mean age,33.0 « 11.6 years). Table 1 summarizes the charac-teristics of the patients. Patients were included ifthey had received a clinical diagnosis of TLE andbeen referred for neuroimaging (including DTI) aspart of an evaluation for surgical indication. For thepurpose of this study, we excluded patients with amass lesion or destructive changes, such as contu-sion or infarction. In addition to evaluating clinicalhistory, seizure semiology and MR imaging, weconducted a comprehensive review of long dura-tion video-EEG recordings (31 of 44 patients), in-tracranial EEG recordings (14 of 44 patients), posi-tron emission tomography (PET) (39 of 44 pa-tients), and single-photon emission computed to-mography (SPECT) (36 of 44 patients) for diagno-sis and lateralization of the seizure focus to the leftor right temporal lobe. We recruited 14 age- andsex-matched healthy controls (6 men, 8 women;mean age, 31.3 « 8.1 years), who were required to

Table 1. Clinical characteristics of patients with right and left temporal lobe epilepsy (TLE)

Right TLE Left TLE P value*

Age (years) 30.6 « 10.4 34.7 « 12.1 0.37

Gender (male/female) 8/7 13/16 0.59

Disease duration (years) 12.6 « 7.7 18.4 « 10.0 0.07

Magnetic resonanceimaging (MRI) lesion

9 hippocampal sclerosis 22 hippocampal sclerosis 0.49

5 MRI negative 5 MRI negative

1 amygdala enlargement 2 amygdala enlargement

*Between-group differences were tested using Student’s t-test for numeric features and chi-squared test for categoricalfeatures (significance level, P < 0.05).

122 K. Kamiya et al.

Magnetic Resonance in Medical Sciences

have no neurological or psychological symptoms,history of neurologic diseases, or apparent abnor-malities observed on conventional MR images.Our institutional review board approved the

study and waived the requirement for informedconsent from patients for the retrospective analy-ses. Written informed consent was obtained fromall the healthy volunteers.

MR imagingWe acquired MR imaging using a 3-tesla MR

imaging system (Signa HDx, GE Medical Systems,Waukesha, WI, USA). DTI data were obtained us-ing a spin-echo echo-planar sequence with diffu-sion gradients along 13 non-collinear directions(b = 1000 s/mm2) and one volume without diffu-sion weighting (b = 0). The other parameters were:repetition time (TR), 13000 ms; echo time (TE), 62ms; field of vision (FOV), 288 © 288 mm2; voxelsize, 3 © 3 © 3 mm3; 50 axial sections; number ofexcitations (NEX), one; and acquisition time, 195 s.Three-dimensional sagittal T1-weighted images

were acquired using an inversion recovery spoiledgradient recalled echo (IR-SPGR) sequence (TR,5.9 ms; TE, 2.3 ms; flip angle, 15°; inversion time[TI], 450 ms; matrix, 256 © 256; FOV, 280 © 280mm2; NEX, 0.5; scan time, 159 s).

Network construction and calculation of graph theo-ry metricsWe processed data of 3-dimensional (3D) T1-

weighted imaging (T1WI) and DTI from each par-ticipant using Connectome Mapper pipeline soft-ware.19 First, we used affine registration in the ed-dy_correct tool implemented in software from theOxford Centre for Functional MRI of the Brain(FSL, FMRIB Software Library, http://www.fmrib.ox.ac.uk/fsl/) to correct each diffusion-weightedimage for distortions caused by head motion andeddy currents. We used FreeSurfer software (Ver-sion 5; http://surfer.nmr.mgh.harvard.edu) to par-cellate the cortical surface, segment gray and whitematter, and define 83 regions of interest (ROIs) (41regions in each hemisphere and one correspondingto the brainstem) with the Desikan-KillianyAtlas.20 The regions were then transformed intoeach subject’s DTI space using boundary-based lin-ear registration (bbregister). All processed imageswere inspected visually for any artifacts, segmen-tation, or registration errors. Diffusion tensor re-construction and whole brain deterministic tractog-raphy were performed with Diffusion Toolkit soft-ware (http://www.trackvis.org/dtk) on the basis ofthe fiber assignment by continuous tracking (FACT)algorithm (threshold angle, 60°). Finally, for each

subject, a connectivity adjacency matrix “A” with83 © 83 entries was generated, with Aij corre-sponding to the weighted connectivity betweenstructures i and j (Fig. 1). Because we observedthat the vast majority of the regional pairs were as-signed zero and represented sparse networks, weapplied no thresholding to the weighted connectiv-ity matrices here. We calculated graph measures foreach individual connectivity matrix using the BrainConnectivity Toolbox (https://sites.google.com/site/bctnet/). We used the degree (k), clustering co-efficient (C), local efficiency (E), and betweennesscentrality (b) to describe the nodal properties of thebrain network on the basis of the results of priorstudies.5,12–15 Table 2 provides definitions of themetrics.21

Comparison with controlsPrior to SVM learning, we performed a nonpara-

metric permutation test to assess between-groupdifferences in each of the nodal parameters com-pared with the controls. We analyzed the left TLEand right TLE groups separately and applied thefalse discovery rate (FDR) to control for multiplecomparisons.

Support vector machine classificationThe SVM is a supervised classification tool that

can automatically learn a classification hyperplanein a feature space by optimizing margin-based cri-teria. We used an SVM with a radial basis functionkernel to solve the classification problem (rightTLE versus controls, left TLE versus controls, andright TLE versus left TLE) using each graph de-scriptor. The gold standard for lateralization of TLEwas a comprehensive review of seizure semiology,EEG, PET, and SPECT. Because our data containedtoo many features compared with the number ofsubjects, we selected features before SVM learningto avoid over-fitting the training data. We used theDantzig selector,22 a type of sparse linear regres-sion, to extract the key features. The generalizationability of the classifier was estimated using a leave-one-out cross-validation (LOOCV) strategy, withreceiver operating characteristic (ROC) curve andthe area under the ROC curve (AUC). All statisticalanalyses, including SVM computing, were per-formed with R version 3.1.1 (The R Foundation forStatistical Computing, Vienna, Austria, http://cran.r-project.org/).

Results

Tables 3 and 4 list the brain regions with signifi-cant differences in nodal properties of patients with

123DTI Connectome for Lateralization of TLE

Vol. 15 No. 1, 2016

Table 2. Definitions of the network measures used (We adopted weighted and undirected definitions and used thestreamline counts as weights.)

Measure Definitions Notes

Degree Degree of node i, ki ¼X

j2Nwij Number of links connected to a node

Clustering coefficient Clustering coefficient of node i,

Ci ¼X

j;h2NðwijwihwjhÞ1=3kiðki � 1Þ

Ci = 0 for ki < 2

The fraction of a node’s neighborsthat are neighbors of each other.

Local efficiency Local efficiency of node i,

Ei ¼X

j;h2N;j 6¼iðwijwih½dwjhðNiÞ��1Þ1=3kiðki � 1Þ

Networks with short path lengths areconsidered more efficient.

Betweenness centrality Betweenness centrality of node i,

bi ¼ 1

ðn� 1Þðn� 2ÞX

h;j2Nh6¼j;h 6¼i;j 6¼i

�hjðiÞ

�hj

“Importance” of that node in the network.

N is the set of all nodes in the network, and n is the number of nodes. wij is the connection weight of the link betweennodes i and j (i; j 2 N). djh (Ni) is the length of the shortest path between j and h, which contains only neighbors of i. µhjis the number of shortest paths between h and j, and µhj

(i) is the number of shortest paths between h and j that passthrough i.

Fig. 1. Flowchart of brain network construction. (1) Parcellation. Eighty-threeregions of interest (ROIs) with the Desikan-Killiany Atlas were defined usingFreeSurfer. (2) Registration. Individual T1-weighted images were registered tothe corresponding non-diffusion-weighted (b = 0) images using boundary-basedlinear registration. The ROIs were registered to the diffusion-weighted imagesusing the same transformation. (3) Tractography. Whole brain tractography wasreconstructed using deterministic tractography. (4) Network construction. The reg-istered ROIs and whole brain tractography were combined to construct the struc-tural brain network. The resultant connectivity adjacency matrix “A” had 83 © 83entries (where Aij corresponds to the connectivity between structures i and j).

124 K. Kamiya et al.

Magnetic Resonance in Medical Sciences

TLE compared with controls. Given that none ofthe differences survived FDR correction for theright TLE, we applied a less stringent significancethreshold of P < 0.005 (uncorrected). In patientswith left TLE, we found 6 brain regions with sig-nificant reductions in local efficiency and the clus-tering coefficient, namely the right paracentral gy-rus, right pericalcarine gyrus, left posterior cingu-late gyrus, left cuneus, and both hippocampi. Wealso observed a significant decrease of local effi-ciency in the left isthmus of the cingulate. In pa-tients with right TLE, the right hippocampusshowed reductions in the nodal degree, clusteringcoefficient, and local efficiency. In addition, pa-tients with left TLE and those with right TLEshowed nonsignificant reductions in the clusteringcoefficient and local efficiency in the temporopar-ietal lobes, including the default mode network(DMN) (Fig. 2). Compared with right TLE, left

TLE appeared to be associated with more extensivealteration of graph measures, including in the con-tralateral hemisphere.The nodal parameters identified as discriminat-

ing factors for SVM were predominantly distribut-ed in the limbic or DMN areas (Table 5). When weapplied the LOOCV strategy using a comprehen-sive review of seizure semiology, EEG, PET, andSPECT as the gold standard for lateralization, theSVM classifier achieved accuracy of 75.9 to 89.7%for right TLE versus controls, 74.4 to 86.0% for leftTLE versus controls, and 72.7 to 86.4% for leftTLE versus right TLE (Table 6). The ROC curvesalso demonstrated the efficiency of the classifierwith moderate to high classification accuracy; theAUC was 0.79 to 0.97 for right TLE versus con-trols, 0.84 to 0.91 for left TLE versus controls, and0.82 to 0.91 for left TLE versus right TLE) (Fig. 3).

Discussion

In this study, use of an SVM and graph theorymeasures demonstrated 72.7 to 86.4% classifica-tion accuracy for left TLE versus right TLE, withan AUC of 0.82 to 0.91. These results were com-parable to those of previous studies that used ma-chine learning approaches for volumetry, DTI, andfMRI.2–5 Although our results for classification ac-curacy do not exceed those of previous reports thatapplied machine learning to DTI using voxel-basedapproaches3 or a fractional anisotropy (FA) skele-ton generated by tract-based spatial statistics(TBSS),4 the graph-based approach has considera-ble strengths as follows. First, it is directed to therecent trend to consider diseases of the brain as net-work disorders.23–25 Second, the number of featurecharacteristics becomes too large in the voxel-based or skeleton-based methods, making featureselection difficult and resulting in expensive com-putational cost. In contrast, the graph theory sum-marizes the network properties with a set of rela-tively few numerical metrics. Third, we can com-pare networks across various modalities, such asEEG, volumetry, surface-based morphometry, DTI,and fMRI, using graph metrics. With regard to clas-sification accuracy, considering that we used DTIdata with only 13 motion-probing gradient (MPG)directions, the use of more dedicated DTI schemes(e.g., Q-ball and diffusion spectrum imaging) andprobabilistic tractography is likely to improve clas-sification accuracy.26

Comparisons of patients with TLE with normalcontrols revealed alterations of graph descriptorsdistributed predominantly in the ipsilateral tempo-roparietal lobe, including areas of the DMN. Our

Table 3. Brain regions with significant differences innodal properties between patients with left temporal lobeepilepsy (TLE) and controls (P < 0.005, uncorrected)

Brain regionClusteringcoefficient

Localefficiency

P value P value

Right paracentral gyrus 0.0003* 0.0008*

Right pericalcarine gyrus 0.003* 0.001*

Right hippocampus 0.001* 0.0003*

Left posterior cingulate 0.001* 0.001*

Left cuneus 0.0006* 0.0007*

Left hippocampus 0.0004* 0.0002*

Left isthmus cingulate — 0.002*

*Survived false discovery rate (FDR) correction for mul-tiple comparisons (q < 0.05).

Table 4. Brain regions with significant differences innodal properties between patients with right temporallobe epilepsy (TLE) and controls (P < 0.005, uncorrect-ed)

Brain regionDegree

Clusteringcoefficient

Localefficiency

P value P value P value

Right lateral occipital — 0.001 —

Right pallidum 0.002 — —

Right accumbens area — — 0.001

Right hippocampus 0.004 0.003 0.001

Left frontal pole 0.003 — —

Left cuneus — — 0.003

125DTI Connectome for Lateralization of TLE

Vol. 15 No. 1, 2016

results are in accord with the reported network al-teration in TLE demonstrated by fMRI,27,28 sup-porting the validity of graph analyses using a DTIstructural connectome. They are also in line withthose of previous reports utilizing graph analysesto assess the DTI structural connectome.11–15 Wedid not observe the reported paradoxical increasein clustering coefficient or local efficiency,12,13 andthere has been some variation in the reported alter-ation of the clustering coefficient in patients withTLE.28 One possible explanation for this discrep-

ancy is that the clustering coefficient depends onthe stage of disease; indeed, it has been reported toincrease during most of the sclerotic process anddecrease in the final stages of disease.29 As for dif-ferences between left and right TLE, most studieshave reported a stronger impact of left TLE on net-work function as assessed by DTI7,15 and resting-state fMRI,27 as in this study. Although the reasonfor the preferential vulnerability of the left hemi-sphere is unclear, one plausible explanation is thatit is related to interference by the disease with for-

Fig. 2. Cortical surface representations showing alterations of the clusteringcoefficient and local efficiency in patients with left and right temporal lobe epi-lepsy (TLE) compared with controls. For visualization purposes, values of the t-static from group comparisons are demonstrated in color, where warm colors (redto yellow) indicate increases in TLE and cool colors (blue) indicate decreases.

Fig. 3. Receiver operating characteristic (ROC) curves of the support vector machine (SVM) clas-sifiers. Red lines, degree; orange, clustering coefficient; blue, local efficiency; and cyan, betweennesscentrality. AUC, area under the curve; b, betweenness centrality; C, clustering coefficient; E, localefficiency; k, degree.

126 K. Kamiya et al.

Magnetic Resonance in Medical Sciences

Table 5. Nodal parameters identified as discriminators for support vector machine (SVM) classification by sparselinear regression

Right temporal lobe epilepsy(TLE) vs controls

Left TLE versus controls Right TLE versus left TLE

k Right pallidum Right posterior cingulated Right medial orbitofrontal

Right hippocampus Right amygdala Right rostral middle frontal

Left frontal pole Left medial orbitofrontal Right superior frontal

Left precentral Left pars triangularis Right caudal anterior cingulate

Left rostral anterior cingulate Left pars opercularis Right fusiform

Left precentral Right parahippocampal

Left entorhinal Right banks of superior temporal sulcus

Right superior temporal

Right pallidum

Left frontal pole

Left caudal middle frontal

Left superior temporal

C Right lateral occipital Right paracentral Right pars orbitalis

Right accumbens area Right entorhinal Right frontal pole

Right hippocampus Left posterior cingulate Right caudal middle frontal

Left frontal pole Left cuneus Right caudal anterior cingulate

Left cuneus Left transverse temporal Right cuneus

Right fusiform

Right accumbens area

Left posterior cingulate

Left isthmus of cingulate

Left precuneus

Left lingual

Left temporal pole

E Right accumbens Right paracentral Right pars orbitalis

Right hippocampus Right pericalcarine Right frontal pole

Left frontal pole Right entorhinal Right insula

Left cuneus Right hippocampus Right accumbens area

Left middle temporal Left posterior cingulate Left isthmus of cingulate

Left superior parietal Left precuneus

Left entorhinal

b Right supramarginal Right rostral middle frontal Right caudal middle frontal

Right lingual Right lateral occipital Right lingual

Left frontal pole Right fusiform Right fusiform

Left caudal middle frontal Right middle temporal Right inferior temporal

Left pericalcarine Right amygdala Left pars orbitalis

Left caudal middle frontal Left isthmus of cingulate

Left isthmus of cingulate Left middle temporal

Left cuneus Left transverse temporal

b, betweenness centrality; C, clustering coefficient; E, local efficiency; k, degree.

127DTI Connectome for Lateralization of TLE

Vol. 15 No. 1, 2016

mation of the language network.15

It has been suggested that 3 factors will deter-mine the practical utility of a connectome-basedclassifier–accurate and robust prediction, clinicallyinformative outcome that cannot be predicted usingother means, and predictive accuracy superior tothat of other simpler and less expensive meas-ures.10 With regard to the first factor, the ROCanalyses indicated moderate to high accuracy forprediction of TLE laterality. As to the second, themachine learning approach may reduce the inci-dence of hippocampal sclerosis being missed atnonspecialized institutions30,31 and may help pre-dict postoperative seizure control in the future.11

Regarding the third factor, although our study wasnot designed to compare the effectiveness of theconnectome classifier with that of FDG-PET or vid-eo-EEG monitoring, its good discrimination be-tween left and right TLE suggests that, with furtherrefinement, the classifier may improve presurgicaldiagnostic confidence and can contribute to obviat-ing the need for invasive intracranial electrodeplacement.Our study has several limitations. First, the ho-

mogeneity of the patient group was not entirelyguaranteed, considering the controversy aboutwhether TLE without hippocampal sclerosis is adistinct neurobiological entity.32 However, we en-rolled these “MRI-negative” patients because theywould benefit most from a quantitative classifica-tion technique. Second, we cannot rule out a con-founding effect of antiepileptic medications thatmay change diffusion properties. Third, from amethodological viewpoint, our results from DTIconnectomes should be interpreted carefully. Thisis because the streamline counts generated by trac-tography do not quantify “connection strengths”whether probabilistic or deterministic tractography

is applied, and no alternative has been establishedto overcome this intrinsic limitation.10,33 The othermethodological concern is the thresholding of thegraph.10 Generally, a weighted connectome needsto undergo a thresholding process because weakand spurious connections may obscure the topolo-gy of the true significant networks, but defining athreshold also introduces arbitrariness. We chose touse no thresholding in consideration of the sparsityof the generated connectome (probably due to thesmall numbers of motion-probing gradients) andexpectation that the feature selection process wouldidentify robust discriminating features less depend-ent on such spurious connections.In conclusion, our results suggest that machine

learning of graph theoretical measures of the DTIstructural connectome is a promising tool for dis-criminating left and right TLE. Optimization of theDTI acquisition scheme and machine learning algo-rithms, including the feature selection method,should further improve classification accuracy.

References

1. de Tisi J, Bell GS, Peacock JL, et al. The long-termoutcome of adult epilepsy surgery, patterns of seizureremission, and relapse: a cohort study. Lancet 2011;378:1388–1395.

2. Akhondi-Asl A, Jafari-Khouzani K, Elisevich K,Soltanian-Zadeh H. Hippocampal volumetry for lat-eralization of temporal lobe epilepsy: automated ver-sus manual methods. Neuroimage 2011; 54(Suppl 1):S218–S226.

3. Focke NK, Yogarajah M, Symms MR, Gruber O,Paulus W, Duncan JS. Automated MR image classi-fication in temporal lobe epilepsy. Neuroimage 2012;59:356–362.

4. An J, Fang P, Wang W, Liu Z, Hu D, Qiu S. Decreas-ed white matter integrity in mesial temporal lobe ep-ilepsy: a machine learning approach. Neuroreport2014; 25:788–794.

5. Chiang S, Levin HS, Haneef Z. Computer-automatedfocus lateralization of temporal lobe epilepsy usingfMRI. J Magn Reson Imaging 2015; 41:1689–1694.

6. Otte WM, van Eijsden P, Sander JW, Duncan JS,Dijkhuizen RM, Braun KP. A meta-analysis of whitematter changes in temporal lobe epilepsy as studiedwith diffusion tensor imaging. Epilepsia 2012; 53:659–667.

7. Ahmadi ME, Hagler DJ Jr, McDonald CR, et al. Sidematters: diffusion tensor imaging tractography in leftand right temporal lobe epilepsy. AJNR Am J Neu-roradiol 2009; 30:1740–1747.

8. Bernhardt BC, Hong S, Bernasconi A, Bernasconi N.Imaging structural and functional brain networks intemporal lobe epilepsy. Front Hum Neurosci 2013;7:624.

Table 6. Classification accuracies in the leave-one-outcross validation

Right temporallobe epilepsy(TLE) versus

controls

Left TLEversuscontrols

Right TLEversusleft TLE

Degree 89.7% 86.0% 86.4%

Clusteringcoefficient

86.2% 74.4% 84.1%

Localefficiency

82.8% 86.0% 72.7%

Betweennesscentrality

75.9% 74.4% 86.4%

128 K. Kamiya et al.

Magnetic Resonance in Medical Sciences

9. Richardson MP. Large scale brain models of epilep-sy: dynamics meets connectomics. J Neurol Neuro-surg Psychiatry 2012; 83:1238–1248.

10. Fornito A, Zalesky A, Breakspear M. Graph analysisof the human connectome: promise, progress, andpitfalls. Neuroimage 2013; 80:426–444.

11. Bonilha L, Helpern JA, Sainju R, et al. Presurgicalconnectome and postsurgical seizure control in tem-poral lobe epilepsy. Neurology 2013; 81:1704–1710.

12. Bonilha L, Nesland T, Martz GU, et al. Medial tem-poral lobe epilepsy is associated with neuronal fibreloss and paradoxical increase in structural connectiv-ity of limbic structures. J Neurol Neurosurg Psychia-try 2012; 83:903–909.

13. DeSalvo MN, Douw L, Tanaka N, Reinsberger C,Stufflebeam SM. Altered structural connectome intemporal lobe epilepsy. Radiology 2014; 270:842–848.

14. Xu Y, Qiu S, Wang J, et al. Disrupted topologicalproperties of brain white matter networks in left tem-poral lobe epilepsy: a diffusion tensor imaging study.Neuroscience 2014; 279:155–167.

15. Besson P, Dinkelacker V, Valabregue R, et al. Struc-tural connectivity differences in left and right tempo-ral lobe epilepsy. Neuroimage 2014; 100:135–144.

16. Anderson JS, Nielsen JA, Froehlich AL, et al. Func-tional connectivity magnetic resonance imaging clas-sification of autism. Brain 2011; 134:3742–3754.

17. Bassett DS, Nelson BG, Mueller BA, Camchong J,Lim KO. Altered resting state complexity in schizo-phrenia. Neuroimage 2012; 59:2196–2207.

18. Chen G, Ward BD, Xie C, et al. Classification ofAlzheimer disease, mild cognitive impairment, andnormal cognitive status with large-scale networkanalysis based on resting-state functional MR imag-ing. Radiology 2011; 259:213–221.

19. Daducci A, Gerhard S, Griffa A, et al. The connec-tome mapper: an open-source processing pipeline tomap connectomes with MRI. PLoS One 2012; 7:e48121.

20. Desikan RS, Ségonne F, Fischl B, et al. An automat-ed labeling system for subdividing the human cere-bral cortex on MRI scans into gyral based regions ofinterest. Neuroimage 2006; 31:968–980.

21. Rubinov M, Sporns O. Complex network measuresof brain connectivity: uses and interpretations. Neu-

roimage 2010; 52:1059–1069.22. Candès E, Tao T. The Dantzig selector: statistical

estimation when p is much larger than n. Ann Stat2007; 35:2313–2351.

23. Carrera E, Tononi G. Diaschisis: past, present, fu-ture. Brain 2014; 137:2408–2422.

24. De Benedictis A, Duffau H. Brain hodotopy: fromesoteric concept to practical surgical applications.Neurosurgery 2011; 68:1709–1723; discussion 1723.

25. Griffa A, Baumann PS, Thiran JP, Hagmann P. Struc-tural connectomics in brain diseases. Neuroimage2013; 80:515–526.

26. Buchanan CR, Pernet CR, Gorgolewski KJ, StorkeyAJ, Bastin ME. Test-retest reliability of structuralbrain networks from diffusion MRI. Neuroimage2014; 86:231–243.

27. Haneef Z, Lenartowicz A, Yeh HJ, Levin HS, EngelJ Jr, Stern JM. Functional connectivity of hippocam-pal networks in temporal lobe epilepsy. Epilepsia2014; 55:137–145.

28. Haneef Z, Chiang S. Clinical correlates of graphtheory findings in temporal lobe epilepsy. Seizure2014; 23:809–818.

29. Dyhrfjeld-Johnsen J, Santhakumar V, Morgan RJ,Huerta R, Tsimring L, Soltesz I. Topological deter-minants of epileptogenesis in large-scale structuraland functional models of the dentate gyrus derivedfrom experimental data. J Neurophysiol 2007; 97:1566–1587.

30. McBride MC, Bronstein KS, Bennett B, Erba G,Pilcher W, Berg MJ. Failure of standard magneticresonance imaging in patients with refractory tempo-ral lobe epilepsy. Arch Neurol 1998; 55:346–348.

31. Von Oertzen J, Urbach H, Jungbluth S, et al. Stand-ard magnetic resonance imaging is inadequate forpatients with refractory focal epilepsy. J Neurol Neu-rosurg Psychiatry 2002; 73:643–647.

32. Berg AT, Berkovic SF, Brodie MJ, et al. Revised ter-minology and concepts for organization of seizuresand epilepsies: report of the ILAE Commission onClassification and Terminology, 2005–2009. Epilep-sia 2010; 51:676–685.

33. Jones DK, Knösche TR, Turner R. White matter in-tegrity, fiber count, and other fallacies: the do’s anddon’ts of diffusion MRI. Neuroimage 2013; 73:239–254.

129DTI Connectome for Lateralization of TLE

Vol. 15 No. 1, 2016