Dissertation/Thesis Abstract

Longitudinal Tissue Segmentation Methods for the Aging Brain
by Schwarz, Christopher George, Ph.D., University of California, Davis, 2012, 148; 3540732
Abstract (Summary)

Automatic methods for tissue segmentation are critical to scientific study of the aging brain. Seeking to understand the biology of aging and the many disease processes that hinder cognition among the elderly population, researchers have embarked upon many large-scale longitudinal neuroimaging studies. In such studies, subjects are recruited to receive Magnetic Resonance Imagery (MRI) scans of their brains, along with tests of cognition and other biomarkers, regularly over time. As these studies produce large volumes of imaging data, they require state of the art computer algorithms for processing and extracting features and biomarkers that are of interest to medical science.

The task of tissue segmentation, determining which type of tissue is present in each location of a brain image, is fundamental to this line of research. Many aging-related brain changes are visible under MRI, and so researchers require tissue segmentation algorithms that accurately measure these changes over time. Although research depends on these algorithms to accurately segment longitudinal data, many of the most popular, state of the art segmenters apply only to individual images, rather than modeling multiple images simultaneously. While joint-longitudinal segmentation would allow for direct modeling of change over time and facilitate more robust, accurate results, computational and domain-specific challenges make this a particularly challenging task.

In this work, we begin with a review of Markov Random Fields (MRFs), the mathematical model underlying many of the most popular algorithms for tissue segmentation and low-level image processing. We then present and validate our own state of the art MRF-based methods for segmentation of the major classes of brain tissue from MR Images: Gray Matter (GM), White Matter (WM), Cerebrospinal Fluid (CSF), and White Matter Hyperintensities (WMH). Additionally, we discuss and present experimental evidence of the many challenges that make these tasks difficult, and how our proposed methods address them. We then present a novel method to reduce the computational cost of MRF-based algorithms and demonstrate its ability to provide significant cost reduction with minimal detriment to classification performance in a WMH segmentation task. By taking advantage of the reduced cost afforded by this procedure, we are able to perform joint-longitudinal segmentation of WMH tissue, which would otherwise be intractable under such conditions. We then demonstrate that such joint-longitudinal segmentation improves the biological feasibility of segmentation outputs for longitudinal data.

Finally, we review the findings of many studies that have already employed techniques presented in this document, in order to provide a broader context for understanding their many applications, and propose several avenues for continuing future research.

Indexing (document details)
Advisor: Carmichael, Owen T.
Commitee: Amenta, Nina, Joy, Kenneth I.
School: University of California, Davis
Department: Computer Science
School Location: United States -- California
Source: DAI-B 74/02(E), Dissertation Abstracts International
Subjects: Neurosciences, Medical imaging, Computer science
Keywords: Brain tissue segmentation, Computer vision, Image processing, Longitudinal anlysis, Markov random fields, Medical image analysis
Publication Number: 3540732
ISBN: 978-1-267-66297-2
Copyright © 2020 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy