Deep Residual Neural Networks for Spatial EEG Source Imaging

EEG source imaging is an indispensable tool for non-invasive study of brain function. Existing methods mainly directly deal with the EEG inverse problem by imposing prior constraints. However, different brain activation patterns may produce similar potential distributions on scalp EEG, which makes the inverse solution process ill-posed. In this paper, we proposed a framework called DeepMapper for EEG spatial source imaging. It consisted of two parts: dataset simulation and neural network training. In the dataset simulation section, we dealt with the EEG forward problem by using the 3-shell realistic head model and the boundary element method (BEM). The cortical functional atlas were used to provide physiological constraints. The nonlinear difference equation was used to simulate the brain information. In the neural network training section, we simulated a large amount of EEG-cortex training data. The prior information in weight layers and nonlinear connections was stored using supervised learning. Simulation results showed that DeepMapper's performance was better than traditional methods in recovering EEG spatial sources. Therefore, it suggested that the proposed method may provide a novel approach to EEG source imaging.