Brain Signal Conversion Based On Generative Adversarial Networks

As a carrier for bridging the internal activities of the brain and real-world task scenarios,EEG signals have a lot of practical application value.In the field of clinical medicine and health care,EEG signals are often used in tasks such as seizure detection,Alzheimer’s disease detection,depression diagnosis,and sleep quality assessment.How to use more efficient and accurate technology to mine the rich information and application scenarios behind EEG signals has become one of the research hotspots in various fields.According to the different acquisition methods,EEG signals can be categorized into non-invasive scalp EEG(electroencephalogram,EEG)and invasive stereo EEG(stereoelectroencephalography,SEEG).EEG has the advantages of easy acquisition and low cost,but it is susceptible to noise interference and cannot reflect neuronal activity deep within the skull.SEEG has greater clarity and clinical value,but is only acquired in limited conditions and limited patients due to its risk of invasive procedures.In order to combine the characteristics of EEG and SEEG,this work proposes the research content of converting EEG signals into simultaneous SEEG signals.In solving this task,we need to face two major challenges.First,we need to determine the mapping relationship from the set of EEG electrodes to the set of SEEG electrodes.Second,we need to design reasonable algorithms to model the complex nonlinear correlation between EEG and SEEG.In view of the above tasks and challenges,this work has tried different algorithms and verification experiments to explore a reasonable EEG signal conversion method.The main contributions of this paper are as follows:1.Two-stage EEG-SEEG Matching Strategy.For the establishment of the mapping relationship between EEG electrodes and SEEG electrodes,we propose a two-stage matching strategy.We consider the similarity between the signals recorded by two electrodes and the physical distance between two electrodes to determine the most suitable one-to-one electrode matching relationship,and require that there is a correlation between the matched electrodes,so that the validity of the generated task is promised.2.EEG-to-SEEG Conversion Based on Conditional Generative Adversarial Networks.For the conversion task,we utilize conditional generative adversarial networks to address this challenge.In order to better characterize the signal,we use the amplitude spectrum and the instantaneous frequency spectrum as the signal representation.To ensure that the generated signals can retain the correlation with the input signals,we propose a correlated spectral attention module.To address the spurious phenomenon in the generated signal,we propose a weighted patch prediction module.Finally,we demonstrate the ability of the proposed method to capture epilepsy-related abnormal micro-discharges between signals through comparison experiments.3.Python-based EEG analysis toolkit.In view of the current lack of toolkits for EEG signal conversion tasks,we encapsulate the engineering code into a set of EEG analysis toolkits based on the Python platform.The toolkit encapsulates the necessary processes involved in the EEG signal preprocessing,model training and analysis phases,saving researchers the extra labor of developing from scratch.At the same time,we provide an extensible interface.Users only need to implement the relevant base classes to implement their own methods such as preprocessing and deep learning model training processes.