Functional MRI data analysis : Detection, estimation and modelling

Abstract

Functional Magnetic Resonance Imaging (fMRI) is a very important technique for neuroimaging in cognitive neuroscience. This thesis proposes advanced signal processing and data analysis methods to the complex fMRI data. The major objectives of the thesis are: (i) enhanced detection of the activated regions of the brain; (ii) better estimation of the haemodynamic response (HDR) and (iii) efficient modeling of the dynamics of the fMRI data. Three methods under Bayesian framework (for determination of the design matrix, for nonstationary noises, and for modified general linear model (GLM) with drift) were suggested to efficiently detect the activated regions of the brain. The HDR which reflects the properties of human brain function was estimated through both linear (spatio-temporal linear adaptive filter) and nonlinear (Radial Basis Function (RBF) neural network) methods. Finally, nonlinear autoregressive with exogenous inputs (NARX) neural network was proposed to model the dynamics of fMRI signal. Through the results of both simulated as well as real fMRI data, it is shown that these methods are robust, efficient and more importantly, flexible. The proposed methods can complement the traditional analysis methods to cope with diverse challenges of fMRI data analysis.