Sparse MEG inverse solutiona via hierarchical Bayesian modeling: Evaluation with a parallel fMRI study

Here we demonstrate how sparse cortically constrained MEG inverse solutions can be obtained via the hierarchical Bayesian minimum-norm estimation approach, wherein Gaussian priors with individual precisions (inverse variances) are first assumed for the distributed current amplitudes at each cortical location. A common second-level prior (hyperprior) of specific form is then imposed on the prior precisions to cause most of the current amplitudes to vanish, thus resulting in sparse inverse solutions. We validated the approach using an empirical dataset, wherein identical visual stimulation experiments were carried out in MEG and fMRI. The physiologically highly feasible fMRI statistical parametric maps were used as a reference for the MEG solutions. The results show that the hierarchical Bayesian approach is capable of producing solutions concordant with the fMRI data in a rather automated fashion, despite the characteristic complexity of the visual evoked magnetic fields. The proposed method of selecting the hyperprior to obtain sparsity provides an effective and straightforward way to regularise the solutions. We also discuss the possibility of utilising fMRI information in the MEG source reconstruction.