Automatic Recognition of Resting State fMRI Networks with Dictionary Learning

Abstract

Resting state functional magnetic resonance imaging (rs-fMRI) is a functional neuroimaging technique that investigates the spatially remote yet functionally linked neuronal coactivation patterns of the brain at rest. Non-invasiveness and task-free characteristics of rs-fMRI make it particularly suitable for aging, pediatric and clinical population. Researchers typically follow a source separation strategy to efficiently reconstruct the concurrent interacting resting state networks (RSN) from a myriad of whole brain fMRI signals. RSNs are currently identified by visual inspection with prior knowledge of spatial clustering of RSNs, as the variability and spatial overlapping nature of RSNs combined with presence of various sources of noise make automatic identification of RSNs a challenging task. In this study, we have developed an automated recognition algorithm to classify all the distinct RSNs. First, in contrast to traditional single level decomposition, a multi-level deep sparse matrix factorization-based dictionary leaning strategy was used to extract hierarchical features from the data at each level. Then we used maximum likelihood estimates of these spatial features using Kullback-Leibler divergence to perform the recognition of RSNs. Experimental results confirmed the effectiveness of our proposed approach in accurately classifying all the RSNs.