M/EEG imaging by learning mean norms in brain tiles

We present a new approach to the M/EEG inverse problem, formulated in the framework of probabilistic modeling. Given a tiling of the brain into separate regions, we define a model parametrized by the mean source power, or norm, in different regions, as well as the mean noise power. A fast algorithm learns optimal values of these region-specific norms from data, leading to higher-resolution images compared to minimum-norm methods that minimize the total norm of the solution. It also learns the noise power, facilitating automatic regularization. The algorithm produces robust reconstructions of current distributions across time, which are shown to be quite accurate.