Adaptive canonical correlation analysis methods for effective fMRI data analysis

© 2018 Dr. Muhammad Ali QadarFunctional magnetic resonance imaging (fMRI) is a powerful non-invasive technique that enables monitoring blood oxygenation level dependent (BOLD) contrasts as a proxy for neuronal activity. fMRI data analysis is an emerging field that addresses various problems concerning neuroimaging using fMRI data, including functional connectivity analysis, which entails identifying brain regions associated with certain cognitive tasks as well as investigating the temporal inter-connectivity between brain regions. Functional connectivity networks are usually determined by employing statistical data analysis techniques which are categorized into hypothesis-driven and data-driven methods. Among hypothesis-driven techniques, the most widely used approach is statistical parameter mapping (SPM) which employs the general linear model (GLM). One major limitation of hypothesis-driven techniques is that they require prior knowledge about the form of hemodynamic response function (HRF) and the experimental paradigm. On the other hand, data-driven techniques do not require such specifics, making them effective for activation detection in task-related datasets and functional connectivity analysis in resting-state datasets. In general, conventional data-driven techniques suffer from problems such as computational inefficiency, and poor performance in retrieving spatiotemporal information. To address these problems, the present thesis considers fMRI data decomposition while introducing assumptions such as prior information about the temporal structure, sparseness on the spatial characteristics, and autocorrelation on temporal characteristics of the fMRI data. In this thesis, we develop variants of a special type of data-driven method known as canonical correlation analysis (CCA). The strength of CCA lies in its multivariate nature and ability to analyse multiple datasets simultaneously. However, the majority of conventional CCA methods do not account for the above mentioned characteristics of fMRI data. Our goal in this work is to propose novel CCA algorithms to better estimate spatiotemporal components by introducing assumptions about the underlying structure of the data. For every proposed algorithm, extensive simulation studies using artificially generated data as well as experimental fMRI data are carried out in this thesis. The statistical performance of the proposed algorithms has been thoroughly evaluated on four different fMRI datasets quantitatively and by comparative analysis with existing data-driven techniques. Results have shown that the proposed algorithms outperform the traditional data-driven techniques in retrieving activation maps and temporal signals, leading to overall improvements in activation detection and functional connectivity analysis