Improved extraction of temporal basis from fully sampled fMRI data using a combined ICA and SVD approach

Motivation: Low rank methods such as partial separability (PS) reconstructions extract a temporal basis from a low spatial resolution navigator signal. PS reconstructions require long scan times to allow for high enough rank to capture the data while preventing ill-conditioning. Goal(s): Extract an improved lower rank representation (temporal basis) of the data from a navigator. Approach: Use a combined ICA/SVD basis to preserve nongaussian fMRI signals at very low rank relative to an SVD basis alone. Results: Improved temporal basis extraction was achieved for BOLD fMRI data with visual task activation. Future work includes application to the PS model. Impact: Basis functions from SVD and ICA can be combined to capture nongaussian signals at very low ranks when SVD alone does not. Potential applications include reducing scan time for low TR partial separability MRI reconstructions.