Multiple Features Analysis And Automatic Identification Of Temporal Lobe Epilepsy Based On MR Images

Temporal lobe epilepsy(TLE)is a neurologic disorder caused by abnormal epileptiform discharge,which affects more than 50 million epilepsy patients worldwide.As the etiology and clinical symptoms are complicated,differential diagnosis of the TLE still mainly relies on experienced doctors,and specific clinical indicators remain to be found.Over the past decades,there were widely extensive studies on the genetics,pathology,neuroimaging of the TLE.Nevertheless,the aberrant patterns of cerebral cortical features in the brain of patients with TLE is still unclear,accurate and effective diagnosis of TLE and focus locatelization need to be explored.First,the structural MR images of 117 patients with TLE and 115 healthy volunteers(healthy controls,HC)were collected.Then the FreeSurfer software was applied to perform cortical reconstruction and cortical features calculation.Next,four cortical measures,namely cortical thickness(CTh),cortical surface area(CSA),gray matter volume(GMV),and mean curvature(MCu),were extracted for statistical and discriminative analysis.Two different brain atlases including the Desikan-Killiany and Destrieux,and two feature selection methods incorporating the two sample t-test filtering and support vector machine-recursive feature elimination(SVM-RFE)were explored to distinguish the TLE using the SVM classifier.Furthermore,deep learning technique was introduced to fulfill the classification task.Two convolutional neural networks(CNNs)architectures including CaffeNet and GoogleNet were employed to construct classifiers for discerning the TLE using two training strategies such as training from scratch and transfer learning from pre-trained models.Finally,the performance of the SVM classifier and deep learning technique was compared and discussed.This study revealed that the TLE exhibited extensive cortical thinning,especially brain regions in the left hemisphere,including the bilateral inferior temporal,the bilateral caudal middle frontal,the left lateral orbital frontal,the left precunes and the right pericalcarine.For the CSA,the left precentral,the left parsopercularis,the bilateral caudal middle frontal and the right superior frontal were increased,while GMV atrophy was detected in the left fusiform,the left precunes,the left superior frontal and the right inferior temporal of the TLE.The results exhibited that SVM classifier using the cortical features generated by the Destrieux atlas achieved most prominent performance with 100% accuracy,while the classifier using the Desikan-Killiany atlas obtained only 65.09% accuracy.Moreover,the SVM-RFE strategy outperformed the t-test method,and the cortical features of CSA and GMV exhibited more prominent discriminative ability.Especially,the brain regions with strongest discriminative power prevailingly located in the occipital pole,rectus gurus,suborbital sulcus,transverse frontopolar gyri and sulci.The CaffeNet and GoogleNet architectures with transfer learning strategy were applied for classification,the classifier achieved 91.16% and 86.56% accuracy in the TLE differential diagnosis.This study illustrated that there existed morphologic changes of various cortical features in the brain of patients with TLE.Different cortical features and their changes,different classifiers had significant differences in discriminative power,and the patterns of salient cortical features could be effectively applied for automatic recognition and focus locatelization,which had significant potential to facilitate the clinical diagnosis of the TLE.