Applying Machine Learning Approaches To Whole-brain White Matter Analysis In Medial Temporal Lobe Epilepsy

Part one:Applying machine learning approaches and DTI-TBSS to whole-brain white matter analysis in medial temporal lobe epilepsyObjectivePrevious studies have demonstrated that the white matter abnormalities is not limited to the temporal lobe white matter, but distributed in whole brain white matter in medial temporal lobe epilepsy (mTLE). However, these findings vary across studies and between regions. This may be due to the small sample sizes and differences in experimental protocols. In addition, the traditional group-level statistical methods do not provide a mechanism for evaluating the discriminative power of the identified white matter regions at the individual level. In recent years, machine learning approaches have been increasing used for brain image analysis because they are capable of extracting additional information and stable patterns from neuroimaging data, and finding significant whole brain neuroimaging-based biomarkers and identifying patients from controls at individual subject levels. Therefore in this study we adopted machine learning approaches along with tract based spatial statistic (TBSS) to investigate the whole brain white matter changes in mTLE. Additionally, we also explored whether the left and right mTLE patients and controls, based on DTI data, could be identified from each other with a promising accuracy by machine learning approaches; whether the most discriminating white matter regions could be found by the approach used, which not only provide useful information for lateralization of the seizure focus, but also can be used as a potential biomarker for the diagnosis and treatment of the mTLE.Materials and Methods1. SubjectsThis study was approved by the Research Ethics Review Board of the Institute of Mental Health of Southern Medical University. Each participant was informed of the details of the project and written informed consent was obtained from all participants in accordance with the standards of the Declaration of Helsinki. Sixty six subjects have participated in present study, including32mTLE patients and34normal adults. We enrolled32consecutive right-handed patients with unilateral HS and mTLE, who were undergoing a presurgical evaluation at Guangdong999Brain hospital. The diagnosis and lateralization of the seizure foci into left mTLE (n=17) and right mTLE (n=15) were determined by a comprehensive evaluation including detailed history, video-EEG telemetry and neuroimaging. An increase in T2fluid-attenuated inverted recovery (FLAIR) signal in the hippocampus was used as the diagnostic criterion for HS, and the site of HS was concordant with the epileptogenic site in all patients. None of the patients had a mass lesion (including tumor, vascular malformation or malformations of cortical development), traumatic brain injury or any psychiatric disorder. Thirty-four age-, gender-and education-matched right-handed healthy control participants were recruited in this study. All controls were healthy and free of any neurological or psychiatric disorders at the time of the study.2. DTI data acquisition All participants were scanned with a1.5T Philips Intera MR scanner. During scanning, foam pads were used to reduce head motion and scanner noise. Diffusion-weighted imaging was obtained using a single-shot echo-planar imaging sequence with the following parameters:TR=11000ms; TE=71.6ms; FOV=230mm x230mm; matrix size=144×144; voxel dimensions=1.6×1.6×2mm; slice thickness=2mm;32non-collinear diffusion directions with a b-value of800s/mm2and one additional volume without diffusion weighting (b=0s/mm2); and73transverse slices without gaps, covering the whole brain.3. DTI data preprocessing and imaging analysisImages obtained in DICOM format were first converted to ANALYZE. Then, the four-dimensional diffusion tensor images were aligned to the first volume with McFlirt (FSL tool) to eliminate head motion error. Afterwards, the aligned diffusion tensor images were corrected for distortions caused by eddy currents by using affine registration in Eddy Current Correction (FSL tool). After completing these preprocesses, the resulting images were brain extracted using the FSL Brain Extraction Tool (BET) and a diffusion tensor model was fitted at each voxel using DTIFit (FMRIB Software Library’s Diffusion Toolbox) to generate images of FA.Using TBSS, voxel-wise cross-subject comparisons were made between FA profiles of mTLE and control subjects to identify discrete regions of white matter abnormalities. First, a target image was determined by aligning every subject’s FA image to each other image, in order to determine the most representative subject. This target image was then normalized to MNI152standard space using an affine transformation. All other subjects were then aligned first to the target image and then to1×1×1mm MNI152space with nonlinear registration implemented with FNIRT (FSL tool). This process created a mean FA skeleton that represents the centers of all tracts common to the group. All individual subjects’aligned FA data (TLE and control subjects) was projected onto the FA skeleton, and the resulting data were used for voxel-wise classification.4. ClassificationFirst, the FA skeleton images were concatenated to a feature vector and combined as a row in a large feature matrix. As the FA skeleton account for just a small part of the whole image, we extracted the FA skeleton matrix from the large feature matrix, leaving the non-zero features. However, the remaining non-zero feature dimensionality was still too high for direct classification and the discriminating features were buried by unuseful features due to registration errors and image noise. Reducing the dimensionality of feature space can not only speed up the computation but also improve classification performance. Two-sample t-test (TSTT), simple and effective, was adopted in this study to select the most discriminating features.In the machine learning approaches, feature selection is always followed by feature reduction. As an unsupervised nonlinear dimensionality reduction algorithm, local linear embedding (LLE) can obtain a low dimensional embedding while preserving the intrinsic structure of the data due to its nonlinear nature, geometric intuition, and computational feasibility. Here, LLE was used in this study to reduce the dimensionality of feature space to a more manageable level.Finally, in the classification part, we chose support vector machines (SVMs) as our classification algorithms because they are resilient to overfitting, allow the extraction of feature weights and increasingly used in neuroimaging studies.Due to the limited sample size, we adopted a leave-one-out cross-validation (LOOCV) strategy to estimate the generalization rate of the SVM classifier in this paper. Suppose there are N subjects in total. In each fold of LOOCV, N-1subjects were selected to train the SVM classifier and the remaining one subject was left to test the classifier. In each fold of LOOCV, we first adopted TSTT to select the most significantly different D features for the N-1training subjects. Then LLE was performed to reduce the feature space dimensionality from D to d. The results were used to train the SVM classifier and the remaining one subject was employed to evaluate the classifier performance by comparing classification results with the ground truth class labels. As there are N samples, the classifier was trained and tested for N times in LOOCV strategy. The performance of a classifier was quantified using Sensitivity (SS), Specificity (SC) and Generalization Rate (GR) based on the results of LOOCV. The SS indicated the proportion of patients classified correctly, and the SC represented the proportion of controls that were classified correctly. The overall proportion of samples classified correctly was represented by GR.Taking each subject’s predicting score of SVM as a threshold, the receiver operating characteristics (ROC) curve was yielded to further estimate the performance of our classifier. Furthermore, permutation tests were applied using the generalization rate as the statistic to assess the statistical significance level of the observed classification accuracy. The class labels of the training data were first randomly permuted, and then the cross-validation was carried out for each set of label-permuted data. The entire permutation process was repeated10,000times.ResultsThe classification accuracy between left mTLE and control was94.1%. We found out that the FA of those discriminative features in left mTLE were decreased compared to controls. The reduced FA voxels were distributed in the left cingulum hippocampal part, thalamus involving anterior thalamic radiation, genu of corpus callosum, and temporal white matter involving inferior longitudinal fasciculus.The classification accuracy between right mTLE and control was91.8%. The FA of the most discriminative features in right mTLE were decreased compared to controls. The reduced FA voxels were distributed in the right fornix, thalamus involving anterior thalamic radiation, splenium of corpus callosum, and temporal white matter involving inferior fronto-occipital fasciculus and uncinate fasciculus.The classification accuracy between left mTLE and right mTLE was90.6%. Compared with right mTLE, left mTLE had decreased FA in the left temporal white matter involving uncinate fasciculus, while right mTLE had decreased FA in the right frontal white matter involving superior longitudinal fasciculus and inferior fronto-occipital fasciculus, right temporal white matter involving inferior longitudinal fasciculus and uncinate fasciculus, and right posterior corona radiata.ConclusionThis study demonstrates that the machine learning approaches can, based on TBSS, identify left mTLE patients, right mTLE patients and healthy controls from each other with a promising accuracy. Compared with controls, FA value of the most discriminating voxels was decreased in the ipsilateral limbic system, corpus callosum, and temporal white matter in both patient groups. Compared with right mTLE, left mTLE had decreased FA in the left temporal white matter, while right mTLE had decreased FA in the right frontal and temporal white matter, and right posterior corona radiata. These findings not only provide useful information for lateralization of the seizure focus, but also can be used as a potential biomarker for the diagnosis and treatment of the mTLE. Part two:Applying machine learning approaches and DTI to whole-brain white matter network analysis in medial temporal lobe epilepsyObjectivePrevious functional MRI (fMRI) studies have demonstrated that the medial temporal lobe epilepsy (mTLE) is potentially a brain disease with network dysfunction. Neurons within cortical gray matter are presumed to be the generator of epileptic activity, however, as the axons are the transmission pathways of the brain, white matter is an integral part of the epileptic network. In fact, some white matter sub-networks including the limbic system and default mode network have been investigated in mTLE. However, the whole brain white matter network has not been fully characterized in patients with mTLE. Therefore in this study we adopted machine learning approaches and DTI to investigate how mTLE is associated with whole brain white matter network dysfunction. The results may provide new insights into the pathophysiology of mTLE and improve our understanding of cognitive dysfunction in patients with mTLE.Materials and Methods1. SubjectsThis study was approved by the Research Ethics Review Board of the Institute of Mental Health of Southern Medical University. Each participant was informed of the details of the project and written informed consent was obtained from all participants in accordance with the standards of the Declaration of Helsinki. Ninety two subjects have participated in present study, including53mTLE patients and39normal adults. We enrolled43consecutive right-handed patients with unilateral hippocampal sclerosis (HS) and mTLE, who were undergoing a presurgical evaluation at Guangdong999Brain hospital. The diagnosis and lateralization of the seizure foci into left mTLE (n=22) and right mTLE (n=21) were determined by a comprehensive evaluation including detailed history, video-EEG telemetry and neuroimaging. An increase in T2fluid-attenuated inverted recovery (FLAIR) signal in the hippocampus was used as the diagnostic criterion for HS, and the site of HS was concordant with the epileptogenic site in all patients. None of the patients had a mass lesion (including tumor, vascular malformation or malformations of cortical development), traumatic brain injury or any psychiatric disorder. After the MRI acquisition, all patients underwent anterior temporal lobectomy. Following qualitative histopathological analysis, HS was detected in all patients. Thirty-nine age-, gender-and education-matched right-handed healthy control participants were recruited in this study. All controls were healthy and free of any neurological or psychiatric disorders at the time of the study. In addition, we also enrolled10consecutive right-handed patients with mTLE and no visual lesion (left mTLE/right mTLE:5/5). None of the10patients had a mass lesion (including tumor, vascular malformation or malformations of cortical development), traumatic brain injury or any psychiatric disorder.2. Imaging protocolAll participants were scanned with a1.5T Philips Intera MR scanner. During scanning, foam pads were used to reduce head motion and scanner noise. Diffusion-weighted imaging was obtained using a single-shot echo-planar imaging sequence with the following parameters:TR=11000ms; TE=71.6ms; FOV=230mm x230mm; matrix size=144×144; voxel dimensions=1.6×1.6×2mm; slice thickness=2mm;32non-collinear diffusion directions with a b-value of800s/mm2 and one additional volume without diffusion weighting (b=0s/mm2); and73transverse slices without gaps, covering the whole brain. We also acquired high resolution3D brain anatomical images using a T1-weighted MP-RAGE sequence with the following parameters:repetition time (TR)=25ms, echo time (TE)=4.6ms, field of view (FOV)=240mm×240mm, matrix size=256×256, and140contiguous axial slices with slice thickness=1mm.3. DTI data preparationImages obtained in DICOM format were first converted to ANALYZE. Then, the four-dimensional diffusion tensor images were aligned to the first volume with McFlirt (FSL tool) to eliminate head motion error. Afterwards, the aligned diffusion tensor images were corrected for distortions caused by eddy currents by using affine registration in Eddy Current Correction (FSL tool). After completing these preprocesses, the resulting images were brain extracted using the FSL Brain Extraction Tool (BET) and a diffusion tensor model was fitted at each voxel using DTIFit (FMRIB Software Library’s Diffusion Toolbox) to generate images of FA and other parameters.4. Network constructiona. Cortical parcellationOne critical step in the network construction was to parcellate the cortex into regions of interest (ROIs), which were located in an identical topographic position for each participant regardless of the anatomical variance across participants. Here, we adopted an automatic ROI parcellation method to parcellate the cortex into116ROIs, which comprised the nodes in the network. First, we registered the b0images to T1-weighted MP-RAGE images, and then registered the transformed T1-weighted images to the T1ICBM152template in MNI space. Finally, the resulting transformation matrix was inversed, to warp the automated anatomical labelling (AAL) atlas to the diffusion-MRI native space.b. White matter tractographyFor each DTI set, the Gaussian kernel size was set to6for smoothing prior to reconstruction. Then, a diffusion tensor was fitted to each voxel using the linear least-squares fitting method. The fiber assignment by continuous tracking (FACT) algorithm was subsequently adopted for deterministic tractography using TrackVis software (http://www. trackvis.org). For the FACT algorithm, the orientation of the eigenvector with the largest eigenvalue was assumed to correspond to the local orientation of any underlying axonal fiber bundle. The eigenvector with the largest eigenvalue is henceforth referred to as the principal eigenvector. A single seed was placed in the center of each voxel, and a streamline was seeded for the maximum vector of every orientation density function (ODF) at every voxel, extending the streamline along the vector of least curvature in the adjacent voxel; streamlines were terminated at curvatures greater than60°or if FA dropped below a threshold of0.2. Tractographic maps were calculated using Diffusion Toolkit and viewed with TrackVis.c. Network constructionFinally, we combined the output of the cortical parcellation and white matter tractography steps to create the adjacency matrix of brain connectivity. Every ROI in the cortical parcellation became a node in the graph. The ROI associated with the node v is denoted as ROI(v). Connections between two nodes ROI(v) and ROI(u) are defined as an edge e=(v, u). For each edge e we defined its weight w(e) as the fiber number between ROI(v) and ROI(u). Thus, we obtained a symmetric adjacency matrix of116×116for each participant. Removing the diagonal elements, we selected the upper triangle elements (6670elements) as the classification features.5. Feature selection and classificationDue to noise, low image resolution, registration error and individual differences, the highly discriminating features that account for only a small part of the whole feature matrix are obscured by trivial features. Thus, our initial step was to select the most discriminating features to construct the feature space for later classification. To do this, we applied two-sample t-test (TSTT) to identify the features that were significantly different between groups. These significantly different features could be considered the features with the most discriminating power. As a manifold learning technique, locally linear embedding (LLE) is capable of obtaining a low-dimensional embedding of the data while preserving the intrinsic data structures. Therefore, we adopted LLE to reduce the feature space dimensionality to a more manageable level. Finally, a support vector machine (SVM) was applied for classification.As the number of samples in this study is limited, we adopted a leave-one-out-cross-validation (LOOCV) strategy to estimate the generalization rate of the SVM classifier. The performance of each classifier was quantified using Sensitivity (SS), Specificity (SC) and Generalization Rate (GR) based on the results of the LOOCV. The SS indicates the proportion of patients classified correctly, and the SC represents the proportion of controls that were correctly classified. The overall proportion of samples correctly classified is represented by the GR. We adopted the same strategy (feature extraction, SVM classifier and LOOCV) for identifying left mTLE with HS from controls and identifying right mTLE with HS from controls.To assess the statistical significance of the observed classification accuracy values, we applied permutation tests to evaluate the probability of obtaining generalization rates higher than those obtained using the correct labels by chance. Given the null hypothesis that the observed group differences could occur by chance when the classifier trained on randomly re-labeled data, we randomly assigned labels to each image and repeated the entire cross-validation procedure10,000times. We counted the number of times that the generalization rate for the permuted labels was higher than those obtained using the correct labels. We derived a p value for the classification by dividing this number by10,000.ResultsThe classification results for both left and right TLE indicate that the final correct classification rate of the training data set was100%, using the one hundred and thirty-six most discriminating white-matter connections as features. Using leave-one-out cross-validation, the SVM classifier achieved an accuracy of91.8%(SS=81.8%, SC=97.4%; permutation test,P<0.0001) for left mTLE with HS versus controls and91.7%(SS=81.5%, SC=97.4%; permutation test,p<0.0001) for right mTLE with HS versus controls on the testing data set with a Gaussian radial basis kernel function.Because the training data slightly differed in each LOOCV, the selected features vary in each LOOCV. However, ninety-three and eighty-eight features, termed the consensus features, existed in every LOOCV of classification for left and right mTLE with HS versus controls respectively. These two set of consensus features could be considered the most discriminating features in the classification. All of the consensus connections (consensus features) was decreased in both left and right mTLE patients with HS compared with controls. Though the two set of consensus connections did not match each other well, they were both primarily located in temporal-limbic network, frontal-limbic network, parietal lobe and cerebellum. Classification was also performed using the connections with the temporal lobe masked out. The classifications of left mTLE with HS versus controls and right mTLE with HS versus controls resulted in accuracies of90.2%(permutation test, p<0.0001) and91.7%(permutation test, p<0.0001), respectively. The discriminating connections selected from these two classifications were in accordance with those selected from the two classifications using the whole-brain connections as features respectively. The classification accuracy of10mTLE patients versus39controls turned out to be88%. Besides, we combined mTLE with HS and mTLE with no visual lesion together for classification, which were53patients versus39controls. The classification achieved an accuracy of87%for the53patients versus39controls comparison.ConclusionThe present study characterized the whole-brain white-matter network disturbance which had been long neglected in mTLE with HS using machine learning approaches. The results showed that left and right mTLE patients with HS can be differentiated from healthy controls with a classification accuracy of91.8%and91.7%respectively based on whole-brain white-matter connections, indicating that the most discriminating connections could be viewed as potential biomarkers for mTLE with HS. The most discriminating connections were decreased in both left and right mTLE patients with HS. These two sets of connections were both primarily located in temporal-limbic network, frontal-limbic network, parietal lobe and cerebellum except some connections exhibiting lateralization. In addition, classification accuracies were90.2%and91.7%for left and right mTLE with HS versus controls respectively when using the connections with the temporal lobe masked out as features. This study verified that disruption of the whole brain white matter network was related to mTLE with HS and indicated that white-matter network signatures may provide potential biomarkers for the prediction of mTLE with no visual lesion.