Inverse / Forward modeling and solution ( 5 ) 7 - 19 : An enhanced particle filter for estimating neural currents from magnetoencephalographic data

We present a dynamical Bayesian method to estimate source currents from MEG data. We assume that the sources can be modeled by a small set of current dipoles whose number, position, orientation, and amplitude may vary over time. A particle filter tracks the probability densities of these parameters by a two-step procedure comprising a filtering step in which the particles best fitting the data survive, and a prediction step in which particles evolve. We improved our previous particle filter algorithm [1] by using i) a grid of precomputed forward solutions that reduces the computational cost substantially and allows a larger number of simultaneous sources (up to 5); ii) a Probability Hypothesis Density instead of the conditional mean for more reliable estimates; iii) a clustering analysis on the entire set of estimated dipoles to keep track of the identities of the sources throughout the analysis epoch.This procedure was applied both to simulated data (sources mimicked the response to a complex visual stimulus with real background noise [2]), and real data. We compared our results (simulated data) with those obtained by traditional multi-dipole fitting and minimum current estimates [2], and conclude that particle filter performs better than the other two methods. [1] Sorrentino et al. Particle filters: a new method for reconstructing multiple current dipoles from MEG data. New Frontiers in Biomagnetism (ICS) 1300 (2007) pp 173-6 [2] Stenbacka et al. Comparison of minimum current estimate and dipole modeling in the analysis of simulated activity in the human visual cortices. Neuroimage 16 (2002) 936-943