Magnetoencephalographic source localization and reconstruction via deep learning.

Within this manuscript a deep learning algorithm designed to achieve both spatial and temporal source reconstruction based on signals captured by MEG devices is introduced. Brain signal estimation at source level is a significant challenge in magnetoencephalographic (MEG) data processing. Traditional algorithms offer excellent temporal resolution but are limited in spatial resolution due to the inherent ill-posed nature of the problem. Nevertheless, many applications require precise localization of pathological tissues to provide reliable information for clinicians. In this context, deep learning solutions emerge as promising candidates for high resolution signals estimations. The proposed approach, termed "Deep-MEG," employs a hybrid neural network architecture capable of extracting both temporal and spatial information from signals captured by MEG sensors. The algorithm is capable to handling the entire brain and, therefore, is not limited to cortical sources imaging. To validate its efficacy, the Authors conducted simulations involving multiple active sources using a realistic forward model, and subsequently compared the results with those obtained using various state-of-the-art reconstruction algorithms. Finally Deep-MEG has been tested also with real MEG data.