Date of Award

4-4-2017

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Elizabeth Pierce

Second Advisor

Keith Bush

Abstract

To better understand the brain’s structure and function, we need to be able to measure the brain’s activity. The field of measuring the brain is neuroimaging. Neuroimaging methodology predominantly relies on the functional magnetic resonance imaging (fMRI) blood oxygen level dependent (BOLD) contrast signal. Although the BOLD signal is a valid measure, it is an indirect measure of neural activity. Rather, two interacting latent components comprise BOLD: 1) neural events, which encode cognitive processes, and 2) physiological responses to neural events, which are well-modeled by the Hemodynamic Response Function (HRF). In this dissertation, we explore models of these components via realistic simulation. First, we model the structure of neural events using a Hidden Markov Model (HMM), a statistical model in which the system being modeled is assumed to be a Markov chain process involving latent (unobserved) states. We explore both static and dynamic variants of these models suggested by the literature. We conduct a series of experiments using both static and dynamic models, and the results show that the dynamic model best represents the properties of the fMRI BOLD signal regardless of whether the neural event generation process was static or dynamic. Here the criteria of best explaining the fMRI BOLD also includes the model complexity alongside the goodness of the fit. We measure the model complexity using the Akaike Information Criterion (AIC). Second, we explore variance-based methods for identifying the latent HRF function. To achieve this, first we show the existence of the functional relationship between the shape of the HRF and the distribution of the filtered noise. We show this by deconvolving BOLD signal using a set of different HRFs. This variety in HRF time-to-peak characteristics creates an approximately linear increase in mean residual magnitude. We then build a classifier using this functional relationship to identify the true, underlying HRF function.

Download

Share

COinS