Evaluating the performance of BSBL methodology for EEG source localization on a realistic head model

Source localization in EEG represents a high dimensional inverse problem, which is severely ill-posed by nature. Fortunately, sparsity constraints have come into rescue as it helps solving the ill-posed problems when the signal is sparse. When the signal has a structure such as block structure, consideration of block sparsity produces better results. Knowing sparse Bayesian learning is an important member in the family of sparse recovery, and a superior choice when the projection matrix is highly coherent (which is typical the case for EEG), in this work we evaluate the performance of block sparse Bayesian learning (BSBL) method for EEG source localization. It is already accepted by the EEG community that a group of dipoles rather than a single dipole are activated during brain activities; thus, block structure is a reasonable choice for EEG. In this work we use two definitions of blocks: Brodmann areas and automated anatomical labelling (AAL), and analyze the reconstruction performance of BSBL methodology for them. A realistic head model is used for the experiment, which was obtained from segmentation of MRI images. When the number of simultaneously active blocks is 2, the BSBL produces overall localization accuracy of less than 5 mm without the presence of noise. The presence of more than 3 simultaneously active blocks and noise significantly affect the localization performance. Consideration of AAL based blocks results more accurate source localization in comparison to Brodmann area based blocks.