Bayesian Modeling for fMRI Brain Activation and Connectivity
- Author(s): Yu, Zhe;
- Advisor(s): Ombao, Hernando;
- et al.
Functional magnetic resonance imaging (fMRI) is a non-invasive technique that measures the associated changes in the cerebral blood flow, which reflects the neuronal activities in the brain. In this dissertation, novel Bayesian modeling approaches are developed for studying brain activation and connectivity using fMRI data. Brain activation modeling aims to identify the activated brain regions in response to the experimental condition; brain connectivity modeling aims to study the communications between brain regions. Findings from brain activation and connectivity analyses can be applied to medicine and neuroscience, as they may explain the pathological pathways of mental disorders and other neurological diseases, identify potential risk factors, explore hidden symptoms, and help improve disease diagnosis and treatment plans.
The proposed modeling approaches have the following advantages. (1) Local activation and global connectivity are simultaneously modeled in the framework of general linear model (GLM) and Granger-causality through vector-autoregressive models or state-space models. (2) The effect of variation in the hemodynamic response functions (HRF) among brain regions, which could confound the estimation for connectivity, is accounted for by estimating region-specific HRFs. (3) Sparsity is imposed on connectivity to allow for convenient inference on connectivity network, without having to estimate multiple models of different connectivity networks followed by model comparisons. (4) Connectivity is modeled to have the flexibility to vary across experimental conditions. (5) Using Bayesian approach, available scientifically relevant prior information can be incorporated. (6) The hierarchical modeling in the proposed multi-subject model allows the information to be shared across subjects to increase statistical power, and allows to compare differential activation and connectivity patterns across subject groups. Simulation studies show the proposed approaches perform well in detecting brain activation and inferring connectivity network. These models were applied to real fMRI data sets from the stroke study at UCI Neurorehabilitation Lab (PI: Cramer), and suggested potential compensatory effect that requires the involvement of the primary motor region from the unaffected brain hemisphere and the secondary motor regions to aid in executing the simple motor task using the stroke-affected hand.