Characterizing and differentiating task-based and resting state FMRI signals via two-stage dictionary learning

A relatively underexplored question in fMRI is whether there are intrinsic differences in terms of signal composition patterns that can effectively characterize and differentiate task-based and resting state fMRI (tfMRI or rsfMRI) signals. In this paper, we propose a novel two-stage sparse representation framework to examine the fundamental difference between tfMRI and rsfMRI signals. In the first stage, subject-wise whole-brain tfMRI and rsfMRI signals are factorized into dictionary matrix and the corresponding loading coefficients via dictionary learning method. In the second stage, dictionaries learned at the first stage across multiple subjects are aggregated into the matrix which serve as the input for another round of dictionary learning, obtaining groupwise common dictionaries and their loading coefficients. This framework had been applied on the recently publicly released Human Connectome Project (HCP) data, and experimental results revealed that there exist distinctive and descriptive atoms in the groupwise common dictionary that can effectively differentiate tfMRI and rsfMRI signals, achieving 100% classification accuracy. Moreover, certain common dictionaries learned by our framework have a clear neuroscientific interpretation. For example, the well-known default mode network (DMN) activities can be recovered from the heterogeneous and noisy large-scale groupwise whole-brain signals.