An extended block sparse Bayesian learning algorithm for EEG source imaging

Electroencephalographic (EEG) source imaging reconstructs brain activity patterns from noninvasive electrode recordings. A persistent challenge is that identical EEG recordings can correspond to different source locations. To solve this problem, a novel spatial smoothing (SS) and discrete cosine transform (DCT) based block sparse Bayesian learning (BSBL) algorithm (i.e., the SS-DCT-BSBL algorithm) is explored for EEG source imaging. First, SS is exerted on brain voxel signals to generate clustered, sparse signals. Next, DCT transforms EEG signals into coefficients with main source information in low-frequency coefficients and noises in high-frequency coefficients, and captures temporal correlations between voxels through the block structure. A boundary optimization (BO)-based SS-DCT-BSBL algorithm is then derived for the update of hyperparameters. Experiment comparisons verify that the proposed algorithm is more superior over state-of-the-art methods using synthetic and clinical data (epilepsy and face processing) across various signal-to-noise ratios (-10 dB to 20 dB) and source configurations (1 to 6).