Recovering HRFs from overlapping ROIs in fMRI data using thresholding correlations for sparse dictionary learning

Recovering region-specific hemodynamic response function (HRF) in noisy fMRI data is essential to characterize the temporal dynamics of functionally coherent brain regions during activation. Data-driven techniques not based on sparsity fails to recover sub-region HRFs from overlapping regions of interest (ROIs) in task-related activations. This paper exploits spatial sparsity for recovering distinct HRFs from un-delineated overlapping ROIs in fMRI data. Spatial sparsity is realized using thresholding correlation for dictionary learning. The effectiveness of the proposed procedure is illustrated on both simulated and an experimental fMRI data obtained during a visual-task.