Recurrent Neural Network for M/EEG Source Localization

This paper presents the solution of the inverse M/EEG problem using machine learning algorithms. The data on which the algorithm was trained were obtained as a result of solving the forward problem. The data are generated using a recurrent model of neuronal brain activity, where the consecutive excitation of sources in time of closely spaced sources is based on a stochastic Markov chain. The dipole activation model is expected to increase the accuracy of the M/EEG inverse problem solution by reducing its indeterminacy by replacing uncorrelated excitations of sources whose magnetic field is statically distributed on the brain surface with their spatially related time series. Machine learning algorithms in this case can recognize the correlation of source excitations and solve the problem in real time. The application of a method for localization of neuronal activity of the brain that takes into account the spatiotemporal character of the excitations can presumably provide a more complete picture of the brain functioning.