Markov random field image prior models for map reconstruction of magnetoencephalogram images

In this paper the maximum a posteriori (MAP) reconstruction of magnetoencephalograms (MEG) is investigated. The solution is cast as a classical inverse imaging problem, which for MEG is notoriously ill posed and requires strong regularization. Two different Markov random vector field (MRF) models suitable for MEG regularization are developed. The first model uses a potential function which encourages solution sparseness on a rectilinear sample grid. The second model permits simulating an MRF over the non-uniform grid required for hemispherical sampling of the brain. Both methods utilize a new Markov vector field structure where neuron current dipole orientations are explicitly included in the model. MAP reconstructions are presented using simulated and real MEG data.