A Deep Learning Solution for Brain Source Localization and Signal Reconstruction in Magnetoencephalography

Magnetoencephalography represents one of the main state-of-the-art methods for brain functional analysis. It is a non-invasive device where a helmet equipped with sensors is located around the scalp in order to measure magnetic fields originated from the electrical brain activity. In many studies magnetoencephalography signals are exploited to map the brain function in order to identify sources locations and reconstruct their temporal signals. This is possible through the resolution of an inverse problem that passes from the recorded signals to the brain activity. Several reconstruction algorithms have been proposed and exploited in literature but, due to the inherent ill-posedness of the inverse problem, they offer good temporal resolution but a limited spatial resolution. Many clinical applications require precise localization of pathological tissues to provide reliable information for neurologists. Within this manuscript, an alternative inversion algorithm is proposed in order to estimate a more refined identification of source locations. The proposed one is a deep learning algorithm named deep-MEG. It demonstrates to overcome other state-of-the-art reconstruction approaches being effective in precisely locating the active brain area and in reconstructing its temporal evolution with a good level of accuracy, even in a challenging noisy scenario. Moreover, deep-MEG can reconstruct active areas belonging to the whole brain and it is not limited to cortical sources.