Narrative comprehension is a fundamental cognitive skill that involves the coordination of different functional brain regions. To investigate the network structure among the brain regions supporting this cognitive function, a Spectral Bayesian Network with Bayesian model averaging is developed based on the spectral density estimation of the functional Magnetic Resonance Imaging (fMRI) time series recorded from multiple brain regions. In this approach, the neural interactions and temporal dependence among different brain regions are measured by spectral density matrices after a Fourier transform of the fMRI signals to the frequency domain. A Bayesian model averaging method is then applied to build the network structure from a set of candidate networks. Using this model, brain networks of three distinct age groups are constructed to assess the dynamic change of network connectivity with respect to age. Networks of multivariate time series are also simulated from vector autoregressive models to compare the performances of the SBN with existing methods in learning network structure from time series data.
In addition to the network modeling of the functional interactions among brain regions, the quantification of the functional connectivity between two brain regions is also very important for understanding how the functions of the human brain develop. Using spectral coherence and partial spectral coherence, the overall and direct functional connectivity strengths among the language-related neural circuits are computed based on fMRI time series data collected in 313 children ranging in age from 5 to 18 years in a story comprehension experiment. The age or gender effects on both the pair wise direct link and connection strength are studied to access children’s development of brain functions for story comprehension. In addition, the connectivity differences between the left and right hemispheres, and the connections in both hemispheres that are directly related to the children's story comprehension performance, are also studied using spectral connectivity.
The selection of the fMRI preprocessing pipeline plays a critical role in any modeling attempt using fMRI data. The third part of this dissertation focuses a novel residual bootstrap framework for the evaluation of fMRI preprocessing pipelines based on the repeatability of the activation maps generated from a statistical model. Based on the bootstrapping of residual time series from regressions, this framework define image reproducibility using the similarity of the activation maps generated from resampled fMRI images. The multiple-level linear model combined with the residual bootstrap scheme is very flexible and can accommodate non-homogenous populations such as case-control or pediatric data sets. The superior performance of this method is demonstrated through synthetic fMRI data sets with different degrees of motion artifacts, and the evaluation of the preprocessing pipelines from a longitudinal fMRI study.