Efficient Blind Source Separation Method for fMRI Using Autoencoder and Spatiotemporal Sparsity Constraints

Diversity measures exploited by blind source separation (BSS) methods are usually based on either statistical attributes/geometrical structures or sparse/overcomplete (underdetermined) representations of the signals. This leads to some inefficient BSS that derived from a mixing matrix (mm), sparse weight vectors (sw), code (sc). In contrast, proposed efficient method, spatiotemporal (ssBSS), avoids computational complications associated with lag sets, deflation strategy, and repeated error computation using whole dataset. It solves data reconstruction model (STEM) $l_{1}$ -norm penalization notation="LaTeX">$l_{0}$ constraints Neumann’s alternating projection lemma block coordinate descent approach yield desired bases. Its specific solution allows incorporating three-step autoencoder univariate soft thresholding for update source/mixing matrices. Due utilization both spatial temporal information, it can better distinguish between sources interpretable results. These steps also make ssBSS unique because, best my knowledge, no method incorporates sparsity features multilayer network structure. The is validated synthetic various functional magnetic resonance imaging (fMRI) datasets. Results reveal superior performance compared existing mmBSS swBSS. Specifically, overall, 14% increase in mean correlation value 91% reduction time over ssICA algorithm was discovered.