3D Deep Convolutional Neural Network Based Childhood Epilepsy Syndrome Analysis

Childhood epilepsy syndrome is defined as a specific group of clinical manifestations and electroencephalograms(EEG)changes consisting of epilepsy,which seriously affects the growth and development of the patients.In clinical practice,different childhood epilepsy syndromes have different age of onset and seizure characteristics,along with typical EEG changes.Neurologists can diagnose the type of childhood epilepsy syndrome and make targeted treatment plans by analyzing the EEG.However,due to the lack of professional medical resources,accurate diagnose childhood epilepsy syndrome becomes challenging.In addition,current analysis of childhood epilepsy syndrome mainly focuses on seizure detection.With these objectives,the following research are conducted in the thesis:(1)A novel 3D residual-attention-module-based deep network(AR3D)is proposed.In the algorithm,the short-time fast Fourier transform(STFT)is used to extract the time-frequency feature of interictal EEG.Compared with the shortcoming of 2D depth network in multichannel feature representation,the proposed 3D residual-attention-module-based deep network uses multi-channel time-frequency features as inputting,and can better capture the information of different EEG channels.The algorithm achieves the accuracy of 98.62% on CHZU.(2)A novel two-stream 3D attention module based deep network(TSA3D)is developed.In the algorithm,the time-frequency feature is extracted by continuous wavelet transform(CWT),and the frequency-space feature is extracted by power spectral density(PSD)combined with the spatial distribution of EEG electrodes.Compared with single feature inputting,TSA3 D fuses the multi-domain feature information of time-frequency-space,thus realizing the automatic feature extraction,and the improved attention module can not only mine the channel information of features,but also capture the spatial information of features effectively.In terms of accuracy,the proposed method is improved by 7.11% and 0.9% respectively compared with the two-stream VGG and the AR3 D algorithms.(3)An intelligent analysis system for childhood epilepsy syndrome is developed.The system can load and extract the features of multi-channel EEG,and combine the pre-training depth model of the proposed AR3 D network to realize the intelligent classification of childhood epilepsy syndrome.This is of great auxiliary value for clinicians to analyze the EEG of children with epilepsy syndrome,which makes the research content more useful.