Research On Deep Learning-based Algorithms Of Automatic Epileptic EEG Detection

Epilepsy is a common chronic disease caused by the disorder of the central nervous system and abnormal activity of neurons in the brain.Its corresponding seizure attack is characterized by repetitive and sudden,which poses a severe threat to the quality of life and health safety of patients.At present,the cause and mechanism of epileptic seizures in most patients are still unclear,and some refractory epilepsy can not be effectively treated with antiepileptic drugs.As the primary tool for studying the human brain,EEG has been widely used in the clinical diagnosis and treatment of epilepsy.In most cases,clinicians still manually examine and analyze multi-channel long-term EEG signals through vision,which is boring and time-consuming.Therefore,the development of the EEG-based automatic epilepsy detection system can significantly reduce the work burden of doctors and contribute to timely diagnosing and interventing epileptic seizures.Meanwhile,it is also of great significance to improve the life quality of patients and alleviate the economic burden on patients’ families.Most of the existing automatic epilepsy detection algorithms are based on the combination of EEG feature extraction and traditional machine learning classifiers.Although they are usually simple and interpretable,they have limited learning ability and cannot be well-trained on the large-scale database.With the development of deep learning theory and the improvement of the computing capability of the computer,epileptic EEG detection and classification using the deep neural network has become a hot spot of current research.In this thesis,the time-frequency representation of EEG signals and deep learning theory for automatic EEG feature extraction are studied,and several end-to-end automatic epilepsy detection models are proposed.The research contents and innovations of this thesis include the following aspects:(1)A model combining S-transform spectrogram and deep convolution neural network is proposed to realize end-to-end seizure detection.Firstly,the original EEG signal is band-pass filtered by discrete wavelet transform,and then the EEG signal is segmented.S-transform gives the time-frequency representation of multi-channel EEG segments.The obtained S-transform spectrogram is input as multi-channel images into the deep convolutional neural network with multiple convolutional layers for classification,and the seizure probability for each EEG segment is acquired.Finally,the output probability values of the model are fused,and the detection performance is further improved by post-processing.S-transform has the advantages of Short-Time Fourier Transform(STFT)and Continuous Wavelet Transform(CWT).Experimental results show that the S-transform with appropriate window parameters can effectively capture the characteristics of EEG signals in the ictal period and interictal period.Compared with the traditional seizure detection algorithms that combine manual feature extractor and classifier,the proposed model does not need feature engineering steps,and the deep neural network can achieve automatic EEG feature extraction.The results on the Freiburg Epilepsy EEG database show that the proposed algorithm can reach the world-leading detection performance with few training data.(2)At present,most seizure detection algorithms are highly dependent on individual EEG data,and the generalization performance of the algorithm needs to be further improved due to the complex and diverse characteristics of epileptic EEG signals.In order to address these problems,this thesis proposes to use a one-dimensional convolutional neural network as an automatic feature extractor to capture the morphological features of EEG signals and then feed these features into Bidirectional Long Short-Term Memory(BiLSTM)to further extract the temporal characteristics of EEG signals,achieving an end-to-end patient-independent seizure detection model.The channel-perturbation layer is introduced in the training stage,which can significantly improve the generalization performance of the model by randomly perturbating the channel order of the input EEG signals in each training iteration.In order to verify the detection performance of the model on clinical data,the proposed model was comprehensively evaluated on the publicly available CHB-MIT epileptic EEG database and SH-SDU clinical epileptic EEG database collected in the hospital.The experimental results show that the model with a channel perturbation layer can achieve satisfactory performance not only in the cross-subject evaluation scenario but also in the cross-database evaluation scenario.(3)A data-driven learnable time-frequency transform combined with a deep neural network is proposed for the first time.All complex windows in the learnable time-frequency transform are optimized by the same loss function in the deep neural network and updated by the back-propagation algorithm.The learnable time-frequency transform eliminates the manual design,selection,and hyperparameter adjustment of window function or mother wavelet in the traditional time-frequency transform.In this thesis,a recurrent neural network based on learnable time-frequency transform is constructed for the seizure detection task.Taking the learnable time-frequency transform as the first layer of the network,the spectrogram of the EEG signal is extracted.Then it is input into the BiLSTM network to obtain the classification results.In order to verify the effectiveness of the proposed learnable time-frequency transform,the Gaussian window function utilized in S-transform is leveraged to initialize the window parameters of the learnable time-frequency transform.Combined with BiLSTM networks with different structures and training settings,several EEG classification and comparison experiments are carried out on the Bonn Epilepsy EEG database to verify the performance.The experimental results show that the learned time-frequency spectrogram features of learnable time-frequency transform are significantly better than STFT,S-transform,and CWT with various mother wavelets.At the same time,the results on the long-term EEG database show that the algorithm can achieve high seizure detection performance.The time-frequency spectrogram obtained from the training of learnable time-frequency transform has great interpretability.It can express more discriminative time-frequency spectrogram characteristics,which provides a new tool for visualizing EEG features and studying seizure mechanisms.(4)According to the periodicity of the EEG signal,the convolutional kernel modulated by cosine function is used to replace the traditional convolutional kernel,and the cosine convolution operator is proposed for the first time.The cosine convolution kernel has only two learnable parameters,i.e.,amplitude and frequency,making the cosine convolutional neural network with the cosine convolutional operator have fewer parameters than the traditional convolution neural network.In this thesis,quantities of comparative experiments are designed on the epileptic EEG classification task to assess the cosine convolution neural network.The results show that the accuracy and generalization performance of cosine convolution neural networks are significantly better than traditional convolution neural networks.The cosine convolution operator can be used as a generalized convolution operation unit to replace the conventional convolution operator and can easily be embedded into other existing convolution neural network structures.At the same time,this thesis also proposes a post-training quantization algorithm of cosine convolution network based on KL divergence and cosine look-up table,which can reduce the computational space complexity of cosine convolutional network nearly four times without precision loss.In addition,for the quantized cosine convolution operator,a cosine convolution accelerator is designed on the Field Programmable Gate Array(FPGA),and different structures of deep cosine convolution neural networks are deployed on the Xilinx Zedboard development board to verify the performance of the seizure detection system.The hardware assessment results show that the proposed automatic seizure detection hardware system based on a cosine convolution neural network can achieve real-time seizure detection with low power consumption and high accuracy.The work of this thesis can effectively promote the research and application of deep learning theory in epileptic EEG signal classification.Although this thesis has made some progress in the automatic epilepsy detection algorithm research,due to the limitations of experimental data,the seizure detection performance still needs more EEG data to verify.