Classification Of Temporal Lobe Epilepsy Based On Resting State Functional Magnetic Resonance Imaging

Temporal lobe epilepsy(TLE)is one of the most frequent forms of focal epilepsy,and recurrent seizures over a long period of time can be very damaging and potentially life-threatening to the patient.Due to the complex and diverse pathogenesis and clinical symptoms of TLE,it is difficult to accurately diagnose TLE at an early stage and to assess the precise epileptogenic region preoperatively,making subsequent treatment difficult.As a relatively advanced neuroimaging technique,resting-state functional magnetic resonance imaging can detect the activity between different regions of the brain and has been widely used in studies related to TLE.Therefore,based on restingstate functional magnetic resonance imaging,this thesis will analyze the abnormal changes in brain regions of TLE patients from the perspectives of topological properties of brain functional networks and signal characteristics of brain regions in the timefrequency domain.At the same time,machine learning algorithms are used to achieve automatic classification of TLE patients,which is expected to assist in the clinical diagnosis of TLE patients,shorten the diagnosis time and improve the accuracy of the diagnosis.(1)A study of temporal lobe epilepsy based on multivariate pattern analysis and Granger causal analysis.In this part of the study,regional homogeneity(Re Ho)was calculated for each subject brain region in the TLE group and the normal controls(NC)group using the REST toolbox.Then,based on the Re Ho values of individual brain regions,the sample data from the TLE and NC groups were classified using multivariate pattern analysis,and a classification accuracy of 87.27% was obtained,indicating that abnormalities in brain region Re Ho can be used as an important indicator to differentiate between TLE patients and NC.Furthermore,four brain regions were selected as seed points based on the classification weights and cluster sizes of each brain region calculated by multivariate pattern analysis,and the change of effective connectivity(EC)between seed points and whole brain was calculated by voxel-based Granger causal analysis.The results showed that the TLE group had enhanced EC from the right cuneus to the right superior occipital gyrus and right middle occipital gyrus and weakened EC from the right superior occipital gyrus and right middle occipital gyrus to the right cuneus compared to NC,and this alteration in bidirectional connectivity between brain regions provides a new direction for exploring abnormalities in connectivity of brain regions in TLE patients.In addition,correlation analysis of Re Ho and clinical intelligence quotient scores(IQ)at four seed points in the TLE group revealed a significant positive correlation between Re Ho and verbal IQ(VIQ)in the left superior temporal gyrus,and between Re Ho.A significant negative correlation was found between Re Ho in the right lingual gyrus and performance IQ(PIQ).(2)A study on the classification of temporal lobe epilepsy based on brain functional network features.The above findings confirm the validity of voxel-based Granger causal analysis in exploring abnormalities in EC between brain regions.Therefore,in this part of the study,we will construct directed and undirected brain functional networks using brain region-based Granger causal analysis and Pearson correlation,respectively,and calculate 11 relevant brain functional network topological properties,including 4 global properties and 7 local properties,based on a graph analysis approach.Then,a two-sample t-test was used to select features for the obtained directed and undirected brain functional networks properties respectively,and the filtered network topological properties were used as input features for the SVM classification model to classify the sample data of the TLE group and the NC group.The results showed that the directed brain function network constructed based on Granger causal analysis had an accuracy of 90.91% and a sensitivity of 82.61% for both groups,which was significantly higher than that of the undirected brain functional network constructed based on Pearson correlation,confirming the effectiveness and superiority of the classification model based on the topological properties of the directed brain functional network combined with SVM in the diagnosis of TLE.Based on the above classification results,the accuracy of the directed topological properties of 6 sub-networks(Dos-160 atlas division)combined with the SVM classification model to classify the accuracy separately were further calculated and found that the cerebellum network alone had the highest accuracy of 81.82%.The feature selection results based on a two-sample t-test revealed that patients with TLE showed abnormal changes in the topological properties of several brain nodes in the bilateral posterior cerebellum lobes of the cerebellum network,the left inferior parietal lobule of the fronto-parietal network,the left mid insula of the sensorimotor network and the bilateral occipital lobes of the occipital network.In addition,correlation analysis of node flow coefficient,Pagerank centrality and IQ in the directed brain functional network of the TLE group revealed a significant negative correlation between flow coefficient and PIQ in the left posterior cerebellum in the cerebellum network and a significant positive correlation between Pagerank centrality and VIQ in the left occipital gyrus in the default mode network.(3)A study of temporal lobe epilepsy classification based on signal features in the time-frequency domain of brain regions.In this part of the study,time-frequency analysis was performed on two groups of pre-processed time series using short-time Fourier transform,wavelet transform and Hilbert-Huang transform,and the signal features in the time-frequency domain of the two groups of subjects were extracted based on three signal processing methods,and the weighted frequencies under the three signal processing methods were calculated respectively.Then,a two-sample t-test was used to select features for the weighted frequencies obtained from the three signal processing methods,and the filtered weighted frequencies were used as input features for the SVM classifier to classify the sample data of the TLE group and the NC group.The results showed that the Hilbert weighted frequency(HWF)extracted from the Hilbert-Huang transform had an accuracy of 96.30% and a sensitivity of 91.30% for both groups,and after improving the method of the upfront time domain decomposition,a classification accuracy of 96.36% and sensitivity of 95.65% were obtained.The above classification results were significantly higher than the classification results based on the weighted frequencies extracted from the short-time Fourier transform and wavelet transform for both groups,and also higher than the classification results based on the topological properties of the directed brain functional network,confirming the effectiveness and superiority of the classification model using the HWF extracted from the Hilbert-Huang transform as a feature combined with SVM in the diagnosis of TLE.The results of feature selection based on a two-sample t-test also revealed that TLE patients showed abnormal changes in HWF in several brain regions including the right thalamus,right precentral gyrus,left fusiform gyrus,left parahippocampal gyrus and right insula.