Classification Of Temporal Lobe Epilepsy Based On Structural Magnetic Resonance Imaging

Magnetic resonance imaging makes it possible to study the structural information and the functional information of the human brain.It helps us to explore the abnormal brain regions from the structure and function,and then reveal the physiological mechanism of brain diseases.It is of great significance for the diagnosis and treatment of brain diseases.Epilepsy is a brain dysfunction disease caused by different causes,which is characterized by the dysfunction of central nervous system resulted from excessive discharge of brain neurons.Temporal lobe epilepsy（TLE）is the most common epilepsy syndrome in adults,accounting for more than 1/3 of partial epilepsy and more than half of intractable epilepsy.The diagnosis of TLE has become one of the hot in the field of epilepsy research.Since the brain is a highly complex system,how to effectively and comprehensively describe the brain structure and accurately diagnose brain diseases is still facing great challenges.In recent years,machine learning technology has promoted the human’s understanding of brain and made outstanding contributions to the research of brain disease based on image data.Centering on the topic of TLE automatic classification,this thesis makes full use of the magnetic resonance image information of brain.The main research contents are as follows:（1）Automatic classification of TLE based on three brain atlases.This part focus on the influence of three atlases in TLE classification,namely Desikan killiany（DK）atlas,Destrieux（DS）atlas and Brainnetome（BN）atlas.The brain regions were defined according to three brain atlases,and seven morphological features of each brain region were extracted,including numvert,volume,area,thickness,thicknessstd,meancurv and gauscurv.Then,the automatic classification of TLE was realized using SVM.The study found that the classification performance of BN atlas was better than DK atlas and DS atlas,and the accuracy was 90.57%.The reasons are as follows.On the one hand,the number of different brain regions may reflect different brain tissue levels.Comparing to DK atlas and DS atlas,BN atlas has more fine-grained subregions,including 210 cortical and 36 subcortical subregions.On the other hand,different segmentation methods reflect different information.BN atlas is not only consistent with the anatomical characteristics of the brain in clinic,but also contains information on both anatomical and functional connections.Therefore,BN atlas is expected to provide a more accurate tool for clinical diagnosis of TLE patients.（2）Automatic classification of TLE based on individual morphological brain network（MBN）features.On the basis of morphological features,such as volume,area,thickness,etc.,210 cortical structures of Brainnetome atlas was used as nodes of brain network,and the connection edges of network were obtained by using the correlation of morphological feature vectors between brain regions,and then the structural network was constructed.The topological features of network were obtained by graph theory analysis.Then,the automatic classification of TLE was realized using SVM.The aim of this part is mainly to compare the influence of MBN constructed by LASSO algorithm or Pearson correlation on TLE classification.The study found that the accuracy,sensitivity,specificity,F1-score and AUC of multiple networks based on LASSO algorithm were higher than the network based on Pearson correlation.The reasons are as follows.On the one hand,Pearson correlation considers the correlation of paired brain regions in constructing of MBN.LASSO algorithm considers the correlation across brain regions,which is more consistent with the correlation between brain structures.On the other hand,LASSO algorithm expands the research scope of MBN due to the selection ofλparameters.The best average classification results were obtained in the range from 2×10^-2to 9×10^-2ofλwith accuracy of 93.87%.Therefore,LASSO algorithm can be used the construction of MBN,providing a promising predictive ability for the diagnosis of TLE.（3）Automatic classification of TLE based on the radiomics of hippocampal subregions.The left and right hippocampus were divided into 19 subregions respectively,and 107 features were extracted from each subregion,incuding 18first-order features,14 shape features and 75 second-order features.Then the effect of each subregion in TLE classification was tested.The results show that the classification accuracy of granular cell layer body of dentate gyrus was the highest in the left hippocampal subregion,and the classification accuracy of molecular layer head is the highest in the right hippocampal subregion.The classification accuracy of CA4,CA1 and hippocampal tail on the left and right sides of the hippocampus was relatively good.The classification accuracy of the right hippocampus was higher than that of left hippocampus.Therefore,radiomics,as a quantitative description of imaging,radiomics can more carefully show the structual abnormalities of hippocampal and assist in the diagnosis of TLE.