MUltiple SIgnal Classification (MUSIC) is a standard localization method which is based on the idea of dividing the vector space of the data into two subspaces: signal subspace and noise subspace. The brain, divided into several grid points, is scanned entirely and the grid point with the maximum consistency with the signal subspace is considered as the source location. In one of the MUSIC variants called Recursively Applied and Projected MUSIC (RAP-MUSIC), multiple iterations are proposed in order to decrease the location estimation uncertainties introduced by subspace estimation errors. In this paper, we suggest a new method called Self-Consistent MUSIC (SC-MUSIC) which extends RAP-MUSIC to a self-consistent algorithm. This method, SC-MUSIC, is based on the idea that the presence of several sources has a bias on the localization of each source. This bias can be reduced by projecting out all other sources mutually rather than iteratively. While the new method is applicable in all situations when MUSIC is applicable we will study here the localization of interacting sources using the imaginary part of the cross-spectrum due to the robustness of this measure to the artifacts of volume conduction. For an odd number of sources this matrix is rank deficient similar to covariance matrices of fully correlated sources. In such cases MUSIC and RAP-MUSIC fail completely while the new method accurately localizes all sources. We present results of the method using simulations of odd and even number of interacting sources in the presence of different noise levels. We compare the method with three other source localization methods: RAP-MUSIC, dipole fit and MOCA (combined with minimum norm estimate) through simulations. SC-MUSIC shows substantial improvement in the localization accuracy compared to these methods. We also show results for real MEG data of a single subject in the resting state. Four sources are localized in the sensorimotor area at f=11Hz which is the expected region for the idle rhythm.
