Analytical techniques towards identification of human brain functional networks

Abstract

Human brain is a large scale complex network of different but functionally coupled brain regions. These functionally coupled brain regions together form functional brain networks (fBNs). Identification of these networks help in the diagnosis of various neuropsychiatric disorders. In the field of neuroscience, there are three larger issues to be handled using functional Magnetic Resonance Imaging (fMRI) data. Firstly, reliable functional networks are required to be identified using the spatio-temporal recordings of the brain. Secondly, the dynamic time-varying information processing in the brain results in a faster reorganization of fBNs. Thus, development of advanced methodology to detect dynamic fBNs is one of the biggest challenge in neuroscience. Thirdly, longer scan duration results in poor subjects attentivity to the intended task. This problem relates to poor fMRI data quality. This raises a question as to whether the data can be captured in smaller time such that good data quality fMRI data can be captured.

The above three issues were the main motivations of this dissertation. In this dissertation, we aim to adapt tools from signal processing, such as multivariate regression with constrained optimization technique, to identify fBNs. Proposed multivariate regression method considers weighted combination of different brain regions and captures sparse and dense fBNs simultaneously. We name the proposed method as Multivariate Vector Regression-based Connectivity (MVRC). Further, we extend the proposed method to identify fBNs at the group-level comprising of multiple subjects using group-fused constrained optimization.

Second contribution of the proposed work is to identify overlapping fBNs that are often ignored in the literature. By overlapping, we mean that one brain region may be a part of multiple fBNs. This sounds plausible because one stimulus, say auditory, may stimulate memory and other fBNs apart from the auditory network. This indicates a need for identifying overlapping fBNs compared to the commonly identified disjoint fBNs.

Thirdly, in order to identify dynamic fBNs, we utilize state-of-the-art sliding time window approach and consider fBNs to be static within each window. Further, we propose a technique to identify dynamic overlapping fBNs and show the efficacy of proposed approach on openly available Autism dataset.

Recently, some studies have attempted to compute fBNs via estimated intrinsic stimulus at brain regions instead of commonly used activity time-series. These methods are based on the assumption that the functional connection between brain regions is due to the intrinsic stimulus. In this dissertation, we present novel intrinsic stimulus estimation method which overcomes certain limitations of existing methods and can be further utilized to extract fBNs.

In the end of this dissertation, we present work carried out on developing new method for compressed fMRI for faster data acquisition. We aim to build methods that are able to better preserve fBNs compared to the existing methods. This is essential to show that the proposed method is accurate since extraction of fBNs is one of the the crucial motive for studying fMRI data.