Learning sparse methods for multi-subject brain data analysis via dictionary learning

Abstract

Functional Magnetic Resonance Imaging (fMRI) is a non-invasive neuroimaging technique that has emerged as one of the most utilized imaging modalities for mapping brain regions involved in the cognitive processes. In recent years, various hypothesis-driven and data-driven methods have been proposed to perform single-subject as well as multi-subject fMRI analysis enabling researchers to perform population level inferences. The most widely used data-driven methods include PCA, CCA, ICA, and more recently Sparsity based methods. In this thesis, we focused on developing Multi-subject fMRI analysis methods using a sparsity prior of the latent functional networks as our main underlying assumption. To this end, we utilized the sparse representation based technique called Dictionary Learning.

The multi-subject analysis methods developed in this thesis assume that the time course of a single voxel from a single subject can be modeled as a sparse linear combination of component features from two distinct basis sets (dictionaries), a shared basis and a subject-specific one. Based on this assumption, we developed novel dictionary learning methods which can simultaneously learn not only the features which are shared across multiple subjects but also extracts the unique features which are only present in a specific subject dataset. After successful validation of the proposed methods via simulations, these methods have been applied to real-world task fMRI datasets as well and they have been shown to outperform the existing methods in terms of temporal dynamic and functional brain network recovery.

We further developed a complex-valued adaptive sequential dictionary learning method to learn from the complex-valued fMRI data directly. We show that the proposed method is able to achieve much superior accuracy performance in terms of true and false positive rates of the task-related and DMN components.