Locating highly correlated sources from MEG with (recursive) (R)DS-MUSIC

Abstract

We introduce a source localization method of the MUltiple Signal Classification (MUSIC) family that can locate brain-signal sources robustly and reliably, irrespective of their temporal correlations. The method, double-scanning (DS) MUSIC, is based on projecting out the topographies of source candidates during topographical scanning in a way that breaks the mutual dependence of highly correlated sources, but keeps the uncorrelated sources intact. We also provide a recursive version of DS-MUSIC (RDS-MUSIC), which overcomes the peak detection problem present in the non-recursive methods. We compare DS-MUSIC and RDS-MUSIC with other localization techniques in numerous simulations with varying source configurations, correlations, and signal-to-noise ratios. DS-and RDS-MUSIC were the most robust localization methods; they had a high success rate and localization accuracy for both uncorrelated and highly correlated sources. In addition, we validated RDS-MUSIC by showing that it successfully locates bilateral synchronous activity from measured auditory-evoked MEG.

Subspace based Multiple Constrained Minimum Variance (SMCMV) beamformers

“…Such truncation can be easily included but is unnecessary as simulations confirm, because residuals of the out-projected sources have negligible SNRs and by design do not produce false maxima of SMCMV localizers. Also, more sophisticated "combined" approaches could be constructed using multiloop iterations both on the MUSIC side as in [25] and on the MCMV side as with MIA [18]. However we believe that while possibly yielding improvements in accuracy in certain situations, a manifold increase in computational times would generally outweigh the benefits.…”

MEG Source Localization via Deep Learning

“…Some multi-source variants are also non-recursive (e.g., [ 44 , 45 , 46 , 47 ]), and as a result they use brute-force optimization, assume that the approximate locations of the neuronal sources have been identified a priori, or still require the identification of the largest local maxima in the localizer function. To overcome these limitations, non-recursive methods have recursive counterparts, such as RAP MUSIC [ 52 ], TRAP MUSIC [ 53 ], Recursive Double-Scanning MUSIC [ 54 ], and RAP Beamformer [ 7 ]. The idea behind the recursive execution is that one finds the source locations iteratively at each step, projecting out the topographies of the previously found dipoles before forming the localizer for the current step [ 7 , 52 ].…”

MEG Source Localization via Deep Learning

“…[46], [38], [40], [41]), and as a result they use brute-force optimization, assume that the approximate locations of the neuronal sources have been identified a priori, or still require the identification of the largest local maxima in the localizer function. To overcome these limitations, non-recursive methods have recursive counterparts, such as RAP MUSIC [47], TRAP MUSIC [48], Recursive Double-Scanning MUSIC [49], and RAP Beamformer [7]. The idea behind the recursive execution is that one finds the source locations iteratively at each step, projecting out the topographies of the previously found dipoles before forming the localizer for the current step [47], [7].…”