Abstract
We have applied the eigenspace-based beamformer to reconstruct spatio-temporal activities of neural sources from MEG data. The weight vector of the eigenspace-based beamformer is obtained by projecting the weight vector of the minimum-variance beamformer onto the signal subspace of a measurement covariance matrix. This projection removes the residual noise-subspace component that considerably degrades the signal-to-noise ratio (SNR) of the beamformer output when errors in estimating the sensor lead field exist. Therefore, the eigenspace-based beamformer produces a SNR considerably higher than that of the minimum-variance beamformer in practical situations. The effectiveness of the eigenspace-based beamformer was validated in our numerical experiments and experiments using auditory responses. We further extended the eigenspace-based beamformer so that it incorporates the information regarding the noise covariance matrix. Such a prewhitened eigenspace beamformer was experimentally demonstrated to be useful when large background activity exists.