A novel variational Bayesian method for spatiotemporal decomposition of resting-state fMRI

Introduction We apply a new variational Bayesian factor partition (VBFP) method to the sparse spatiotemporal decomposition of resting state fMRI data. The VBFP method estimates sources with sparse distributions in both spatial and temporal domains and incorporates automatic relevance determination in a fully Bayesian inference framework. Hence it achieves dimension reduction as an integrated part of the inference. We apply VBFP to resting state fMRI data and compare it with a maximum likelihood independent component analysis (ICA) algorithm [Bell and Sejnowski, 1995] and show that VBFP identifies similar functionally coherent brain networks and their temporal fluctuations. The potential advantages of VBFP on the integrated inference of the noise model and on robustness for small sample sizes motivate further investigation.