Interictal And Pre-ictal Pathological Marker Label Generation Algorithm And Study On Lesion Localization

High Frequency Oscillations(HFOs)recorded in intracranial electroencephalography(iEEG)are considered to be a potential clinical biomarker of brain epileptogenic regions.However,the existing method of automatically detecting HFOs requires a large number of labeled data to establish an effective computational model,while reliable expert labeling is an expensive,subjective,time-consuming and laborious job.In order to solve the above problems,this thesis proposes an automatic generation method in a semi-supervised learning-based manner regarding high frequency oscillation labels.On this basis,a method for localizing the epileptogenic zone based on the pre-ictal of seizures was proposed,which combined the statistical characteristics of HFOs and iEEG signal characteristics of each lead of the patient and applied them to the localization of the epileptogenic zone.First of all,this thesis collected the raw iEEG data of epilepsy patients during the interictal period and onset period,with the collected data preprocessed and labeled.Therefore,a data set of clinical high frequency oscillation containing 7700 HFOs signals of 5 patients and a data set of pre-ictal including 4 patients were constructed.Then,this thesis designs an automatic generation method in a semisupervised learning-based way with respect to high frequency oscillation labels,namely,the CRM method.A network architecture composed of recurrent neural network and convolutional neural one was constructed,and a semi-supervised algorithm based on consistency regularization designed to realize the automatic generation of high frequency oscillation labels.The research validated the method on public datasets and clinical datasets,and the accuracy rates reached 91.84%and 87.83%has indicated that the CRM method can effectively alleviate the problem of lack of HFOs labels in clinical applications.Furthermore,the CRM method is compared with supervised algorithms,demonstrating the effectiveness,superiority,and clinical applicability.Finally,this thesis designs an epileptogenic zone localization method based on the pre-ictal of seizures,namely,the LRD method.Firstly,the HFOs signals at the pre-ictal of seizures were marked,and the statistical features of HFOs were calculated,and then the time domain,frequency domain and time-frequency domain features of iEEG signals were extracted,and feature selection was performed.Finally,a classifier was used to classify the leads in the epileptogenic zone.The study validated the method using a clinical dataset,and the accuracy and sensitivity reached 89.58%and 79.47%,proving the role of the LRD method in clinical applications.In addition,the effects of different time windows on the localization of the epileptogenic zone leads were compared.It is basically confirmed the conclusion that the SEEG signal closer to the onset can carry more lead information of epileptogenic zone.