Magn Reson Med Sci © 2015 Japanese Society for Magnetic Resonance in Medicine E-pub ahead of print by J-STAGE doi:10.2463/mrms.2015-0027
MAJOR PAPER
Machine Learning of DTI Structural Brain Connectomes for Lateralization of Temporal Lobe Epilepsy Kouhei KAMIYA1*, Shiori AMEMIYA1, Yuichi SUZUKI2, Naoto KUNII3, Kensuke KAWAI4, Harushi MORI1, Akira KUNIMATSU1, Nobuhito SAITO3, Shigeki AOKI5, and Kuni OHTOMO1 1
Department of Radiology, The University of Tokyo 7–3–1 Hongo Bunkyo, 113–8655 Tokyo, Japan 2 Department of Radiological Technology, The University of Tokyo Hospital 3 Department of Neurosurgery, The University of Tokyo 4 Department of Neurosurgery, NTT Medical Center Tokyo 5 Department 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 that uses diffusion tensor imaging (DTI) structural connectomes to determine lateralization of epileptogenicity in temporal lobe epilepsy (TLE). Materials and Methods: We analyzed diffusion tensor and 3-dimensional (3D) T 1 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 each subject, calculated graph theoretical network measures, and used a support vector machine (SVM) for classification among 3 groups (right TLE versus controls, left TLE versus controls, and right TLE versus left TLE) following a feature reduction process with sparse linear regression. Results: In left TLE, we found a significant decrease in local efficiency and the clustering coefficient in several brain regions, including the left posterior cingulate gyrus, left cuneus, and both hippocampi. In right TLE, the right hippocampus showed reduced nodal degree, clustering coefficient, and local efficiency. With use of the leave-one-out crossvalidation strategy, the SVM classifier achieved accuracy of 75.9 to 89.7% for right TLE versus controls, 74.4 to 86.0% for left TLE versus controls, and 72.7 to 86.4% for left TLE versus 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 discrimination between left and right TLE suggests that, with further refinement, the classifier should improve presurgical diagnostic confidence. Keywords: brain connectome, diffusion tensor imaging, epilepsy, graph theory, machine learning
Introduction Temporal lobe epilepsy (TLE) is the most frequent type of refractory focal epilepsy. Patients with concordant findings from electroencephalography (EEG), seizure semiology, neuropsychologi*Corresponding author, Phone: +81-3-5800-8666, E-mail:
[email protected]
cal assessment, and magnetic resonance (MR) imaging—such as atrophy and fluid-attenuated inversion recovery (FLAIR) hyperintensity of the hippocampus ipsilateral to the side of seizure onset—do extremely well with resection of the mesial temporal structures. 1 Therefore, establishing the laterality of the epileptogenic focus with as much certainty as possible is an important task in the preoperative evaluation of patients with TLE. Howev-
K. Kamiya et al.
er, standard MR imaging protocols may fail to show an identifiable hippocampal abnormality or may provide only subtle findings that remain inconclusive. To increase the certainty of lateralization of the epileptogenic focus and obviate the need for invasive intracranial electrode placement, prior studies have investigated the utility of quantitative or automated MR image analyses, including voxelbased morphometry, 2,3 diffusion tensor imaging (DTI), 3,4 and functional MR imaging (fMRI). 5 DTI can approximate the white matter architecture by describing the directionality and magnitude of water diffusion. In TLE, the decrease in fractional anisotropy (FA) tends to be maximal at the epileptic zone and subtle at a distance, 6 and decreased FA in the extra-temporal regions as well as within the ipsilateral temporal lobe suggests that the network is altered. 7 Recently, in light of several studies relating the presence of extratemporal abnormalities to poor postsurgical outcome, emphasis has shifted to consider epilepsy as a disorder of a widespread brain network. 8 Graph theoretical analysis of brain connectomes has attracted much interest as a method of network analysis suitable for epilepsy research. 8,9 In graph theory terms, a brain network, or connectome, consists of the set of neural elements (nodes) and their interconnections (edges). 10 Nodes usually represent brain regions, whereas edges represent (structural or functional) connections. A set of parameters that characterize specific topological properties of the network can be obtained from graph theoretical analyses of a brain connectome. Several studies have analyzed DTI-based structural connectomes in TLE; the majority have reported altered connectivity to be most prominent within the ipsilateral temporal lobe. 11–15 In recent years, several studies have investigated the performance of machine learning algorithms, such as that of the support vector machine (SVM), for automatic localization of epileptogenic foci using MR voxel-based morphometry (VBM) 2,3 and fMRI. 5 Because graph theory metrics use a subset Table 1.
of numeric parameters to summarize the characteristic properties of huge and complex brain networks, they are mathematically good candidates for a machine learning approach to identify the multivariate feature combinations that best predict an outcome of interest. A combination of machine learning and connectomic measures has been used to distinguish healthy individuals from patients with disorders including autism, 16 schizophrenia, 17 and Alzheimer’s disease. 18 In this context, we examined the performance of a machine learning approach, used in combination with DTI-derived structural connectomes, to determine lateralization of the epileptogenic focus of the TLE.
Materials and Methods Subjects We retrospectively reviewed radiological records of a single institution from 1 January 2007 to 31 October 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 characteristics of the patients. Patients were included if they had received a clinical diagnosis of TLE and been referred for neuroimaging (including DTI) as part of an evaluation for surgical indication. For the purpose of this study, we excluded patients with a mass lesion or destructive changes, such as contusion or infarction. In addition to evaluating clinical history, seizure semiology and MR imaging, we conducted a comprehensive review of long duration video-EEG recordings (31 of 44 patients), intracranial EEG recordings (14 of 44 patients), positron emission tomography (PET) (39 of 44 patients), and single-photon emission computed tomography (SPECT) (36 of 44 patients) for diagnosis and lateralization of the seizure focus to the left or right temporal lobe. We recruited 14 age- and sex-matched healthy controls (6 men, 8 women; mean age, 31.3 « 8.1 years), who were required to
Clinical characteristics of patients with right and left temporal lobe epilepsy (TLE)
Age (years) Gender (male/female) Disease duration (years) Magnetic resonance imaging (MRI) lesion
Right TLE
Left TLE
P value*
30.6 « 10.4 8/7 12.6 « 7.7 9 hippocampal sclerosis 5 MRI negative 1 amygdala enlargement
34.7 « 12.1 13/16 18.4 « 10.0 22 hippocampal sclerosis 5 MRI negative 2 amygdala enlargement
0.37 0.59 0.07 0.49
*Between-group differences were tested using Student’s t-test for numeric features and chi-squared test for categorical features (significance level, P < 0.05). E-pub ahead of print
Magnetic Resonance in Medical Sciences
DTI Connectome for Lateralization of TLE
have no neurological or psychological symptoms, history of neurologic diseases, or apparent abnormalities observed on conventional MR images. Our institutional review board approved the study and waived the requirement for informed consent from patients for the retrospective analyses. Written informed consent was obtained from all the healthy volunteers. MR imaging We acquired MR imaging using a 3-tesla MR imaging system (Signa HDx, GE Medical Systems, Waukesha, WI, USA). DTI data were obtained using a spin-echo echo-planar sequence with diffusion gradients along 13 non-collinear directions (b = 1000 s/mm 2 ) and one volume without diffusion weighting (b = 0). The other parameters were: repetition time (TR), 13000 ms; echo time (TE), 62 ms; field of vision (FOV), 288 © 288 mm 2 ; voxel size, 3 © 3 © 3 mm 3 ; 50 axial sections; number of excitations (NEX), one; and acquisition time, 195 s. Three-dimensional sagittal T 1 -weighted images were acquired using an inversion recovery spoiled gradient 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 © 280 mm 2 ; NEX, 0.5; scan time, 159 s). Network construction and calculation of graph theory metrics We processed data of 3-dimensional (3D) T1 weighted imaging (T 1 WI) and DTI from each participant using Connectome Mapper pipeline software. 19 First, we used affine registration in the eddy_correct tool implemented in software from the Oxford Centre for Functional MRI of the Brain (FSL, FMRIB Software Library, http://www.fmrib. ox.ac.uk/fsl/) to correct each diffusion-weighted image for distortions caused by head motion and eddy currents. We used FreeSurfer software (Version 5; http://surfer.nmr.mgh.harvard.edu) to parcellate the cortical surface, segment gray and white matter, and define 83 regions of interest (ROIs) (41 regions in each hemisphere and one corresponding to the brainstem) with the Desikan-Killiany Atlas. 20 The regions were then transformed into each subject’s DTI space using boundary-based linear registration (bbregister). All processed images were inspected visually for any artifacts, segmentation, or registration errors. Diffusion tensor reconstruction and whole brain deterministic tractography were performed with Diffusion Toolkit software (http://www.trackvis.org/dtk) on the basis of the fiber assignment by continuous tracking (FACT) algorithm (threshold angle, 60°). Finally, for each Magnetic Resonance in Medical Sciences
subject, a connectivity adjacency matrix “A” with 83 © 83 entries was generated, with A ij corresponding to the weighted connectivity between structures i and j (Fig. 1). Because we observed that the vast majority of the regional pairs were assigned zero and represented sparse networks, we applied no thresholding to the weighted connectivity matrices here. We calculated graph measures for each individual connectivity matrix using the Brain Connectivity Toolbox (https://sites.google.com/ site/bctnet/). We used the degree (k), clustering coefficient (C), local efficiency (E), and betweenness centrality (b) to describe the nodal properties of the brain network on the basis of the results of prior studies. 5,12–15 Table 2 provides definitions of the metrics. 21 Comparison with controls Prior to SVM learning, we performed a nonparametric permutation test to assess between-group differences in each of the nodal parameters compared with the controls. We analyzed the left TLE and right TLE groups separately and applied the false discovery rate (FDR) to control for multiple comparisons. Support vector machine classification The SVM is a supervised classification tool that can automatically learn a classification hyperplane in a feature space by optimizing margin-based criteria. We used an SVM with a radial basis function kernel to solve the classification problem (right TLE versus controls, left TLE versus controls, and right TLE versus left TLE) using each graph descriptor. The gold standard for lateralization of TLE was a comprehensive review of seizure semiology, EEG, PET, and SPECT. Because our data contained too many features compared with the number of subjects, we selected features before SVM learning to avoid over-fitting the training data. We used the Dantzig selector, 22 a type of sparse linear regression, to extract the key features. The generalization ability of the classifier was estimated using a leaveone-out cross-validation (LOOCV) strategy, with receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC). All statistical analyses, including SVM computing, were performed with R version 3.1.1 (The R Foundation for Statistical Computing, Vienna, Austria, http://cran. r-project.org/).
Results Tables 3 and 4 list the brain regions with significant differences in nodal properties of patients with E-pub ahead of print
K. Kamiya et al.
Fig. 1. Flowchart of brain network construction. (1) Parcellation. Eighty-three regions of interest (ROIs) with the Desikan-Killiany Atlas were defined using FreeSurfer. (2) Registration. Individual T1-weighted images were registered to the corresponding non-diffusion-weighted (b = 0) images using boundary-based linear registration. The ROIs were registered to the diffusion-weighted images using the same transformation. (3) Tractography. Whole brain tractography was reconstructed using deterministic tractography. (4) Network construction. The registered ROIs and whole brain tractography were combined to construct the structural brain network. The resultant connectivity adjacency matrix “A” had 83 © 83 entries (where Aij corresponds to the connectivity between structures i and j). Table 2. Definitions of the network measures used (We adopted weighted and undirected definitions and used the streamline counts as weights.) Measure Degree
Definitions X w Degree of node i, ki ¼ j2N ij
Notes Number of links connected to a node
Clustering coefficient
Clustering coefficient of node i, X ðwij wih wjh Þ1=3 j;h2N Ci ¼ ki ðki 1Þ Ci = 0 for ki < 2
The fraction of a node’s neighbors that are neighbors of each other.
Local efficiency
Local efficiency of node i, X w ðwij wih ½djh ðNi Þ1 Þ1=3 j;h2N;j6¼i Ei ¼ ki ðki 1Þ
Networks with short path lengths are considered more efficient.
Betweenness centrality
Betweenness centrality of node i, X hj ðiÞ 1 bi ¼ ðn 1Þðn 2Þ h;j2N hj
“Importance” of that node in the network.
h6¼j;h6¼i;j6¼i
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 between nodes 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. µ hj is the number of shortest paths between h and j, and µhj(i) is the number of shortest paths between h and j that pass through i.
E-pub ahead of print
Magnetic Resonance in Medical Sciences
DTI Connectome for Lateralization of TLE
Table 3. Brain regions with significant differences in nodal properties between patients with left temporal lobe epilepsy (TLE) and controls (P < 0.005, uncorrected) Brain region
Clustering coefficient P value
Local efficiency P value
Right paracentral gyrus Right pericalcarine gyrus Right hippocampus Left posterior cingulate Left cuneus Left hippocampus Left isthmus cingulate
0.0003* 0.003* 0.001* 0.001* 0.0006* 0.0004* —
0.0008* 0.001* 0.0003* 0.001* 0.0007* 0.0002* 0.002*
*Survived false discovery rate (FDR) correction for multiple comparisons (q < 0.05).
Table 4. Brain regions with significant differences in nodal properties between patients with right temporal lobe epilepsy (TLE) and controls (P < 0.005, uncorrected) Brain region Right lateral occipital Right pallidum Right accumbens area Right hippocampus Left frontal pole Left cuneus
P value
Clustering coefficient P value
Local efficiency P value
— 0.002 — 0.004 0.003 —
0.001 — — 0.003 — —
— — 0.001 0.001 — 0.003
Degree
TLE compared with controls. Given that none of the differences survived FDR correction for the right TLE, we applied a less stringent significance threshold of P < 0.005 (uncorrected). In patients with left TLE, we found 6 brain regions with significant reductions in local efficiency and the clustering coefficient, namely the right paracentral gyrus, right pericalcarine gyrus, left posterior cingulate gyrus, left cuneus, and both hippocampi. We also observed a significant decrease of local efficiency in the left isthmus of the cingulate. In patients with right TLE, the right hippocampus showed reductions in the nodal degree, clustering coefficient, and local efficiency. In addition, patients with left TLE and those with right TLE showed nonsignificant reductions in the clustering coefficient and local efficiency in the temporoparietal lobes, including the default mode network (DMN) (Fig. 2). Compared with right TLE, left Magnetic Resonance in Medical Sciences
TLE appeared to be associated with more extensive alteration of graph measures, including in the contralateral hemisphere. The nodal parameters identified as discriminating factors for SVM were predominantly distributed in the limbic or DMN areas (Table 5). When we applied the LOOCV strategy using a comprehensive review of seizure semiology, EEG, PET, and SPECT as the gold standard for lateralization, the SVM classifier achieved accuracy of 75.9 to 89.7% for right TLE versus controls, 74.4 to 86.0% for left TLE versus controls, and 72.7 to 86.4% for left TLE versus right TLE (Table 6). The ROC curves also demonstrated the efficiency of the classifier with moderate to high classification accuracy; the AUC was 0.79 to 0.97 for right TLE versus controls, 0.84 to 0.91 for left TLE versus controls, and 0.82 to 0.91 for left TLE versus right TLE) (Fig. 3).
Discussion In this study, use of an SVM and graph theory measures demonstrated 72.7 to 86.4% classification accuracy for left TLE versus right TLE, with an AUC of 0.82 to 0.91. These results were comparable to those of previous studies that used machine learning approaches for volumetry, DTI, and fMRI. 2–5 Although our results for classification accuracy do not exceed those of previous reports that applied machine learning to DTI using voxel-based approaches 3 or a fractional anisotropy (FA) skeleton generated by tract-based spatial statistics (TBSS), 4 the graph-based approach has considerable strengths as follows. First, it is directed to the recent trend to consider diseases of the brain as network disorders. 23–25 Second, the number of feature characteristics becomes too large in the voxelbased or skeleton-based methods, making feature selection difficult and resulting in expensive computational cost. In contrast, the graph theory summarizes the network properties with a set of relatively few numerical metrics. Third, we can compare networks across various modalities, such as EEG, volumetry, surface-based morphometry, DTI, and fMRI, using graph metrics. With regard to classification accuracy, considering that we used DTI data with only 13 motion-probing gradient (MPG) directions, the use of more dedicated DTI schemes (e.g., Q-ball and diffusion spectrum imaging) and probabilistic tractography is likely to improve classification accuracy. 26 Comparisons of patients with TLE with normal controls revealed alterations of graph descriptors distributed predominantly in the ipsilateral temporoparietal lobe, including areas of the DMN. Our E-pub ahead of print
K. Kamiya et al.
Fig. 2. Cortical surface representations showing alterations of the clustering coefficient and local efficiency in patients with left and right temporal lobe epilepsy (TLE) compared with controls. For visualization purposes, values of the tstatic from group comparisons are demonstrated in color, where warm colors (red to yellow) indicate increases in TLE and cool colors (blue) indicate decreases.
Fig. 3. Receiver operating characteristic (ROC) curves of the support vector machine (SVM) classifiers. Red lines, degree; orange, clustering coefficient; blue, local efficiency; and cyan, betweenness centrality. AUC, area under the curve; b, betweenness centrality; C, clustering coefficient; E, local efficiency; k, degree.
results are in accord with the reported network alteration in TLE demonstrated by fMRI, 27,28 supporting the validity of graph analyses using a DTI structural connectome. They are also in line with those of previous reports utilizing graph analyses to assess the DTI structural connectome.11–15 We did not observe the reported paradoxical increase in clustering coefficient or local efficiency, 12,13 and there has been some variation in the reported alteration of the clustering coefficient in patients with TLE. 28 One possible explanation for this discrepE-pub ahead of print
ancy is that the clustering coefficient depends on the stage of disease; indeed, it has been reported to increase during most of the sclerotic process and decrease in the final stages of disease. 29 As for differences between left and right TLE, most studies have reported a stronger impact of left TLE on network function as assessed by DTI 7,15 and restingstate fMRI, 27 as in this study. Although the reason for the preferential vulnerability of the left hemisphere is unclear, one plausible explanation is that it is related to interference by the disease with forMagnetic Resonance in Medical Sciences
DTI Connectome for Lateralization of TLE
Table 5. Nodal parameters identified as discriminators for support vector machine (SVM) classification by sparse linear regression Right temporal lobe epilepsy (TLE) vs controls
Left TLE versus controls
Right TLE versus left TLE
k
Right pallidum Right hippocampus Left frontal pole Left precentral Left rostral anterior cingulate
Right posterior cingulated Right amygdala Left medial orbitofrontal Left pars triangularis Left pars opercularis Left precentral Left entorhinal
Right medial orbitofrontal Right rostral middle frontal Right superior frontal Right caudal anterior cingulate Right fusiform Right parahippocampal 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 accumbens area Right hippocampus Left frontal pole Left cuneus
Right paracentral Right entorhinal Left posterior cingulate Left cuneus Left transverse temporal
Right pars orbitalis Right frontal pole Right caudal middle frontal Right caudal anterior cingulate 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 hippocampus Left frontal pole Left cuneus Left middle temporal
Right paracentral Right pericalcarine Right entorhinal Right hippocampus Left posterior cingulate Left superior parietal
Right pars orbitalis Right frontal pole Right insula Right accumbens area Left isthmus of cingulate Left precuneus Left entorhinal
b
Right supramarginal Right lingual Left frontal pole Left caudal middle frontal Left pericalcarine
Right rostral middle frontal Right lateral occipital Right fusiform Right middle temporal Right amygdala Left caudal middle frontal Left isthmus of cingulate Left cuneus
Right caudal middle frontal Right lingual Right fusiform Right inferior temporal Left pars orbitalis Left isthmus of cingulate Left middle temporal Left transverse temporal
b, betweenness centrality; C, clustering coefficient; E, local efficiency; k, degree.
Magnetic Resonance in Medical Sciences
E-pub ahead of print
K. Kamiya et al.
Table 6. Classification accuracies in the leave-one-out cross validation Right temporal lobe epilepsy (TLE) versus controls
Left TLE versus controls
Right TLE versus left TLE
89.7% 86.2%
86.0% 74.4%
86.4% 84.1%
82.8%
86.0%
72.7%
75.9%
74.4%
86.4%
Degree Clustering coefficient Local efficiency Betweenness centrality
mation of the language network. 15 It has been suggested that 3 factors will determine the practical utility of a connectome-based classifier–accurate and robust prediction, clinically informative outcome that cannot be predicted using other means, and predictive accuracy superior to that of other simpler and less expensive measures. 10 With regard to the first factor, the ROC analyses indicated moderate to high accuracy for prediction of TLE laterality. As to the second, the machine learning approach may reduce the incidence of hippocampal sclerosis being missed at nonspecialized institutions 30,31 and may help predict postoperative seizure control in the future. 11 Regarding the third factor, although our study was not designed to compare the effectiveness of the connectome classifier with that of FDG-PET or video-EEG monitoring, its good discrimination between left and right TLE suggests that, with further refinement, the classifier may improve presurgical diagnostic confidence and can contribute to obviating the need for invasive intracranial electrode placement. Our study has several limitations. First, the homogeneity of the patient group was not entirely guaranteed, considering the controversy about whether TLE without hippocampal sclerosis is a distinct neurobiological entity. 32 However, we enrolled these “MRI-negative” patients because they would benefit most from a quantitative classification technique. Second, we cannot rule out a confounding effect of antiepileptic medications that may change diffusion properties. Third, from a methodological viewpoint, our results from DTI connectomes should be interpreted carefully. This is because the streamline counts generated by tractography do not quantify “connection strengths” whether probabilistic or deterministic tractography E-pub ahead of print
is applied, and no alternative has been established to overcome this intrinsic limitation. 10,33 The other methodological concern is the thresholding of the graph. 10 Generally, a weighted connectome needs to undergo a thresholding process because weak and spurious connections may obscure the topology of the true significant networks, but defining a threshold also introduces arbitrariness. We chose to use no thresholding in consideration of the sparsity of the generated connectome (probably due to the small numbers of motion-probing gradients) and expectation that the feature selection process would identify robust discriminating features less dependent on such spurious connections. In conclusion, our results suggest that machine learning of graph theoretical measures of the DTI structural connectome is a promising tool for discriminating left and right TLE. Optimization of the DTI acquisition scheme and machine learning algorithms, including the feature selection method, should further improve classification accuracy. References 1.
2.
3.
4.
5.
6.
7.
8.
de Tisi J, Bell GS, Peacock JL, et al. The long-term outcome of adult epilepsy surgery, patterns of seizure remission, and relapse: a cohort study. Lancet 2011; 378:1388–1395. Akhondi-Asl A, Jafari-Khouzani K, Elisevich K, Soltanian-Zadeh H. Hippocampal volumetry for lateralization of temporal lobe epilepsy: automated versus manual methods. Neuroimage 2011; 54(Suppl 1): S218–S226. Focke NK, Yogarajah M, Symms MR, Gruber O, Paulus W, Duncan JS. Automated MR image classification in temporal lobe epilepsy. Neuroimage 2012; 59:356–362. An J, Fang P, Wang W, Liu Z, Hu D, Qiu S. Decreased white matter integrity in mesial temporal lobe epilepsy: a machine learning approach. Neuroreport 2014; 25:788–794. Chiang S, Levin HS, Haneef Z. Computer-automated focus lateralization of temporal lobe epilepsy using fMRI. J Magn Reson Imaging 2015; 41:1689–1694. Otte WM, van Eijsden P, Sander JW, Duncan JS, Dijkhuizen RM, Braun KP. A meta-analysis of white matter changes in temporal lobe epilepsy as studied with diffusion tensor imaging. Epilepsia 2012; 53: 659–667. Ahmadi ME, Hagler DJ Jr, McDonald CR, et al. Side matters: diffusion tensor imaging tractography in left and right temporal lobe epilepsy. AJNR Am J Neuroradiol 2009; 30:1740–1747. Bernhardt BC, Hong S, Bernasconi A, Bernasconi N. Imaging structural and functional brain networks in temporal lobe epilepsy. Front Hum Neurosci 2013; 7:624. Magnetic Resonance in Medical Sciences
DTI Connectome for Lateralization of TLE
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
Richardson MP. Large scale brain models of epilepsy: dynamics meets connectomics. J Neurol Neurosurg Psychiatry 2012; 83:1238–1248. Fornito A, Zalesky A, Breakspear M. Graph analysis of the human connectome: promise, progress, and pitfalls. Neuroimage 2013; 80:426–444. Bonilha L, Helpern JA, Sainju R, et al. Presurgical connectome and postsurgical seizure control in temporal lobe epilepsy. Neurology 2013; 81:1704 –1710. Bonilha L, Nesland T, Martz GU, et al. Medial temporal lobe epilepsy is associated with neuronal fibre loss and paradoxical increase in structural connectivity of limbic structures. J Neurol Neurosurg Psychiatry 2012; 83:903–909. DeSalvo MN, Douw L, Tanaka N, Reinsberger C, Stufflebeam SM. Altered structural connectome in temporal lobe epilepsy. Radiology 2014; 270:842– 848. Xu Y, Qiu S, Wang J, et al. Disrupted topological properties of brain white matter networks in left temporal lobe epilepsy: a diffusion tensor imaging study. Neuroscience 2014; 279:155–167. Besson P, Dinkelacker V, Valabregue R, et al. Structural connectivity differences in left and right temporal lobe epilepsy. Neuroimage 2014; 100:135–144. Anderson JS, Nielsen JA, Froehlich AL, et al. Functional connectivity magnetic resonance imaging classification of autism. Brain 2011; 134:3742–3754. Bassett DS, Nelson BG, Mueller BA, Camchong J, Lim KO. Altered resting state complexity in schizophrenia. Neuroimage 2012; 59:2196–2207. Chen G, Ward BD, Xie C, et al. Classification of Alzheimer disease, mild cognitive impairment, and normal cognitive status with large-scale network analysis based on resting-state functional MR imaging. Radiology 2011; 259:213–221. Daducci A, Gerhard S, Griffa A, et al. The connectome mapper: an open-source processing pipeline to map connectomes with MRI. PLoS One 2012; 7: e48121. Desikan RS, Ségonne F, Fischl B, et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 2006; 31:968–980. Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. Neu-
Magnetic Resonance in Medical Sciences
22.
23. 24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
roimage 2010; 52:1059–1069. Candès E, Tao T. The Dantzig selector: statistical estimation when p is much larger than n. Ann Stat 2007; 35:2313–2351. Carrera E, Tononi G. Diaschisis: past, present, future. Brain 2014; 137:2408–2422. De Benedictis A, Duffau H. Brain hodotopy: from esoteric concept to practical surgical applications. Neurosurgery 2011; 68:1709–1723; discussion 1723. Griffa A, Baumann PS, Thiran JP, Hagmann P. Structural connectomics in brain diseases. Neuroimage 2013; 80:515–526. Buchanan CR, Pernet CR, Gorgolewski KJ, Storkey AJ, Bastin ME. Test-retest reliability of structural brain networks from diffusion MRI. Neuroimage 2014; 86:231–243. Haneef Z, Lenartowicz A, Yeh HJ, Levin HS, Engel J Jr, Stern JM. Functional connectivity of hippocampal networks in temporal lobe epilepsy. Epilepsia 2014; 55:137–145. Haneef Z, Chiang S. Clinical correlates of graph theory findings in temporal lobe epilepsy. Seizure 2014; 23:809–818. Dyhrfjeld-Johnsen J, Santhakumar V, Morgan RJ, Huerta R, Tsimring L, Soltesz I. Topological determinants of epileptogenesis in large-scale structural and functional models of the dentate gyrus derived from experimental data. J Neurophysiol 2007; 97: 1566–1587. McBride MC, Bronstein KS, Bennett B, Erba G, Pilcher W, Berg MJ. Failure of standard magnetic resonance imaging in patients with refractory temporal lobe epilepsy. Arch Neurol 1998; 55:346–348. Von Oertzen J, Urbach H, Jungbluth S, et al. Standard magnetic resonance imaging is inadequate for patients with refractory focal epilepsy. J Neurol Neurosurg Psychiatry 2002; 73:643–647. Berg AT, Berkovic SF, Brodie MJ, et al. Revised terminology and concepts for organization of seizures and epilepsies: report of the ILAE Commission on Classification and Terminology, 2005–2009. Epilepsia 2010; 51:676–685. Jones DK, Knösche TR, Turner R. White matter integrity, fiber count, and other fallacies: the do’s and don’ts of diffusion MRI. Neuroimage 2013; 73:239– 254.
E-pub ahead of print
