Bottom-up control of leakage in spectral electrophysiological source imaging via structured sparse bayesian learning

Abstract

Brain electrical activity in different spectral bands has been associated with diverse mechanisms underlying Brain function. Deeper reconnoitering of these mechanisms entails mapping in grayordinates (Gray Matter coordinates), the spectral features of electrophysiological Brain signals. Such mapping is possible through MEG/EEG signals, due to their wide Brain coverage and excellent temporal resolution in reflecting neural-electrical-activity. This process-coined Electrophysiological Source Imaging (ESI)-can only produce approximated images of Brain activity, which are severely distorted by leakage: a pervasive effect in almost any imaging technique. It has been proposed that leakage control to tolerable levels can be achived through using priors or regularization within ESI, but their implementation commonly yields meager statistical guaranties. We introduce bottom-up control of leakage: defined as maximum Bayesian evidence search braced with priors precisely on the spectral responses. This is feasible due to an instance of Bayesian learning of complex valued data: spectral Structured Sparse Bayesian Learning (sSSBL). "Spectral" refers to specific spatial topologies that are reflected by the MEG/EEG spectra. We also present a new validation benchmark based on the concurrency between high density MEG and its associated pseudo-EEG of lower density. This reveals that prevealing methods like eLORETA and LCMV can fall short of expectations whereas sSSBL exibits an exellent performance. A final qualitative assesment reveals that sSSBL can outline brain lessions using just low density EEG, according to the T2 MRI shine through of the affected areas.