Abstract
We propose a technique that enables robust use of blind source separation techniques in fMRI data analysis. The fMRI temporal signal is modeled as the summation of the true activation signal, a physiological baseline fluctuation component, and a random noise component. A preprocessing denoising is used to reduce the dimensionality of the random noise component in this mixture before applying the principal/independent component analysis (PCA/ICA) methods. The set of denoised time courses from a localized region are utilized to capture the region-specific activation patterns. We show a significant improvement in the convergence properties of the ICA iteration when the denoised time courses are used. We also demonstrate the advantage of using ICA over PCA to separate components due to physiological signals from those corresponding to actual activation. Moreover, we propose the use of ICA to analyze the magnitude of the Fourier domain of the time courses. This allows ICA to group signals with similar patterns and different delays together, which makes the iteration even more efficient. The proposed technique is verified using computer simulations as well as actual data from a healthy human volunteer. The results confirm the robustness of the new strategy and demonstrate its value for clinical use.