Subspace based Multiple Constrained Minimum Variance (SMCMV) Beamformers

Abstract

In MEG and EEG brain imaging research two popular approaches are often used for spatial localization of focal task- or stimuli-related brain activations. One is a so called MUSIC approach applied in the form of RAP or TRAP MUSIC algorithms. Another one is the beamformer approach, specifically multiple constrained minimum variance (MCMV) beamformer when dealing with significantly correlated activations. Either method is using its own source localizer functions. Considering simplicity, accuracy and computational efficiency both approaches have their advantages and disadvantages. In this study we introduce a novel set of so called Subspace MCMV (or SMCMV) beamformers whose localizer functions combine MUSIC and MCMV localizers. We show that in ideal situations where forward modeling, data recording and noise measurements are error-free, SMCMV localizers allow precise identification of n arbitrarily correlated sources irrespective to their strength in just n scans of the brain volume using RAP MUSIC type algorithm. We also demonstrate by extensive computer simulations that with respect to source localization errors and the total number of identified sources SMCMV outperforms both the TRAP MUSIC and MIA MCMV (which is the most accurate MCMV algorithm to our knowledge) in non-ideal practical situations, specifically when the noise covariance cannot be estimated precisely, signal to noise ratios are small, source correlations are significant and larger numbers of sources are involved.