A Probabilistic Algorithm for Meg Source Reconstruction

We present a novel algorithm for source localization based on probabilistic modeling of stimulus-evoked MEG/EEG data. This algorithm localizes multiple dipoles with the computational complexity equivalent to a single dipole scan, and is therefore more efficient than traditional multidipole fitting procedures. The algorithm assumes that the activity of multiple dipolar sources can be characterized by a linear combination of known temporal basis functions with unknown coefficients. We model the sensor data as arising from activity in each voxel of interest, plus background activity. We estimate temporal basis functions from the data using a probabilistic algorithm called partitioned-factor analysis, previously developed in our lab. We model background activity outside the voxel of interest as an unknown linear mixture of unobserved background factors plus diagonal sensor noise. We use an expectation-maximization algorithm to calculate MAP estimates of unknown basis function coefficients, background mixing matrix, sensor noise covariance and the likelihood of a dipole in each voxel of interest. In simulations, the algorithm is able to accurately localize several simultaneously-active dipoles, at SNRs typical for averaged MEG data. The algorithm performs well even in configurations that include deep sources and highly correlated sources, and thus is superior to MUSIC and beamforming techniques which are sensitive to correlated sources. The algorithm also correctly localizes real somatosensory and auditory evoked fields to the postcentral sulcus and lower bank of the lateral sulcus, respectively