Principal Component Analysis of fMRI Data in Local Frequency Domain

Most established activation detection techniques for functional magnetic resonance imaging(fMRI) data are always based on oversimplified assumptions that no spatial or temporal correlations exist in the data. In this article, we strive for analysis of fMRI data in local frequency domain with multitaper frequency-domain singular value decomposition(MTM-SVD) technique which explicitly takes into account the intrinsic spatiotemporal correlations in the data and allows modeling patterns of spatiotemporal dynamics of brain activity. A local frequency-based representation well captures the features of BOLD signal evoked by experimental designs with periodic stimuli and the dominant physiology noise, therefore we can utilize principal component analysis to decompose fMRI data in local frequency domain and extract the spatiotemporal patterns of the brain activity. We have made experiment on real fMRI data, and the results demonstrated that our approach could detect the brain activity patterns effectively and reliably.