Dictionary learning based on sparse representations for resting-state functional MRI data analysis

Abstract

Functional Magnetic Resonance Imaging (fMRI) has been valuable to the current understanding of brain function and pre-operative evaluation of patients. In the recent years, the technique has been increasingly applied to the cases when the subject is at rest, also referred to as the resting-state fMRI. Resting-state fMRI measures spontaneous fluctuations in the blood oxygen level-dependent (BOLD) signal to investigate the functional topology of the brain. It is possible to identify various anatomically distinct areas of the brain that demonstrate synchronous BOLD fluctuations at rest, also referred to as the brain functional networks. Conventional approach to extract these functional dynamics is the datadriven Independent Component Analysis (ICA) method. In this work, we propose to utilize sparse representations for identifying functional connectivity networks. Specifically, fMRI signals are decomposed into morphological components which have sparse spatial overlap. Allowing sparse spatial overlap between components is a more physically plausible assumption to the statistical independence assumption of the conventional ICA method. The dictionary is learnt from the data using a K-SVD algorithm. Experimental results show that the proposed MCA-KSVD method can be used as an alternative to the conventional ICA method.