SEEG Pathological Signal Detection And Computer-aided Seizure Onset Zone Localization

Epilepsy is a neurological disorder characterized by frequent seizures and transient brain dysfunction,which is caused by the abnormal discharge of brain neurons.Approximately 30%of epilepsy patients still experience seizures despite receiving medication,which is known as drug-resistant epilepsy.These patients require surgical resection or minimally invasive ablation of the seizure onset zone(SOZ),and the key to successful treatment lies in accurately preoperatively localizing the SOZ.Stereotactic electroencephalography(SEEG)is a necessary tool for the preoperative assessment of epileptic foci in drug-resistant epilepsy,as it can record deep brain structures directly and provide a visual representation of intracranial electrophysiological activity.Clinicians typically visually inspect SEEG monitoring recordings taken over several days to identify abnormal discharges and subsequently locate the SOZ.However,this process is timeconsuming,laborious,and subjective.Therefore,research on the development of an artificial intelligence algorithm for pathological signal detection and SOZ localization in drug-resistant epilepsy is highly necessary.The specific research consists of two tasks:pathological signal detection and SOZ localization.To address these tasks,we propose solutions to extract the pathological signal features from SEEG data,overcome patient-specific property in clinical settings,and characterize the epileptogenic index of SOZ contacts from multiple aspects.(1)An intracranial EEG signal detection model that combines time-frequency features and deep learning is proposed,aiming to address the limited ability of traditional time-frequency analysis methods to extract intracranial EEG pathological features.To extract the temporality of traditional time-frequency features,a temporal enhancement branch is constructed.The raw signal is split for multidomain feature extraction,and the split extracted features are sequentially combined.A Bi-LSTM-AM is then used to enhance the temporal characteristics.To extract high-level hidden features,an end-to-end branch is constructed that automatically extracts deep features of the raw signal using 1D-CNN.Finally,the combined features of the two branches are fed into a shallow neural network for signal classification.The proposed model achieved an accuracy of 97.60%,a sensitivity of 97.78%,and a specificity of 97.42%on the Bern-Barcelona intracranial EEG dataset.It also achieved an average accuracy of 88.03%in cross-patient experiments on clinical SEEG datasets(5 patients),demonstrating its ability to effectively and more comprehensively extract potential pathological features and achieve high signal detection performance.(2)A cross-patient SEEG signal detection model is proposed(SEEG-Net)based on domain generalization,aiming to address the problem of poor generalization ability of models caused by strong patient differences in clinical applications.In order to overcome strong patient differences,a domain generalizer is introduced to enable the model to learn patient domain-invariant representations.Additionally,to reduce the impact of imbalanced disease physiology signals on model performance,a focal domain generalization loss function is designed.To fully exploit the pathological information in SEEG,a multi-scale convolutional neural network(MSCNN)module is designed to increase the receptive field of convolutional kernels,extract multi-frequency domain features,local,and global features of SEEG.In addition,GradCAM++is used to explain the end-to-end learning process of SEEG-Net,enhancing clinical trust in the use of this model.Cross-patient experiments achieve an average AUROC of 88.81%for multi-class detection in MAYO institution from the publicly available multi-center SEEG dataset,an average AUROC of 84.51%for multi-class detection in the FNUSA institution,and an average accuracy of 93.85%in clinical SEEG datasets(9 patients).Multiple datasets preliminarily validate the effectiveness and generalization of the model,providing a more reliable solution for SEEG pathological signal detection in clinical scenarios and laying a crucial foundation for subsequent SOZ localization.(3)A SOZ localization method is proposed based on the analysis of epileptic features of contacts in different periods,aiming to address the problem of insufficient epileptic features of contacts in a localization task.In order to analyze the epileptogenicity of contacts using multiple biomarkers,a method based on the detection of spikes and high frequency oscillations in interictal SEEG to extract the distribution and intensity of different contacts is proposed.Specifically,a pathological ripple occurrence rate feature is introduced to reduce the impact of physiological high-frequency oscillations on localization.To analyze effective connectivity between contacts and extract potential features of seizure onset,a directed transfer function is computed based on preictal SEEG,and the mean out/in flow strength features of contacts are extracted.By fusing the epileptic features across periods and performing feature selection,the method overcomes the limitation of imbalanced distribution between epileptic and non-epileptic contacts using attention mechanism and focal loss function.The proposed method achieves 91.79%sensitivity and 93.98%specificity in the cross-patient experiments with the clinical SEEG datasets(8 patients).Furthermore,the epileptic indexes derived from contact analysis are displayed on magnetic resonance images through the whole-brain mapping,providing accurate preoperative assessment and reliable auxiliary localization for clinicians.