Blind source separation of functional MRI scans of the human brain based on canonical correlation analysis

A novel approach for multi-subject blind source separation (BSS) of brain functional magnetic resonance imaging (fMRI) data is proposed. Group-level comparison analysis is common in the human brain fMRI analysis. Canonical correlation analysis (CCA) for BSS (BSS-CCA) relies on the basis that all meaningful real signals are auto-correlated compared with white noise, which should generally not be considered. By merely requiring that the second-order statistic be zero, BSS-CCA is more relaxed than independent component analysis (ICA), which demands mutual statistics of all orders to be zero. Based on spatial BSS-CCA, we propose an approach termed group BSS-CCA for the analysis of multi-subject fMRI data. In terms of the simulated situation, in which “sources” were partially overlapping in space, we determined that identification using group BSS-CCA was more efficient than that using group ICA. The results from a real data experiment revealed that the proposed group BSS-CCA approach was effective for extracting functional brain networks that were functionally distinct and spatially overlapping from the fMRI data of the human brain.