Type of Document Dissertation Author Hainline, Allison Elisabeth Author's Email Address email@example.com URN etd-09182018-183037 Title Advanced Statistical Techniques in DW-MRI and fMRI Data Analysis Degree PhD Department Biostatistics Advisory Committee
Advisor Name Title Jeffrey Blume Committee Chair Bennett Landman Committee Member Hakmook Kang Committee Member Matthew Shotwell Committee Member Keywords
- medical imaging
- bias and variance
- diffusion MRI
Date of Defense 2018-09-06 Availability unrestricted AbstractAs neuroimaging studies become more numerous and data are increasingly available, the need for improved understanding of the statistical properties of such data increases as well. In this dissertation, we focus on the application of advanced statistical methods to medical image analysis of both diffusion-weighted MRI (DW-MRI) and resting-state functional MRI (rs-fMRI).
We present a method for the estimation of bias and variance in individual high angular resolution diffusion imaging (HARDI) acquisitions that allows for both bias-correction and quality assurance of empirical data acquisitions. Systematic bias of imaging metrics can lead to improper inference in a research setting or misdiagnosis in a clinical setting. These bias and variance estimates can be used for data quality assurance as demonstrated through an application of the proposed methods to a study involving traveling subjects who were scanned repeatedly in up to 4 independent scanners. We also provide a suggested workflow for the use of these metrics in assessing data quality and fitting models for statistical inference. As an extension, we provide a set of deep neural networks that can effectively estimate these bias and variance values 200x faster than the traditional statistical techniques proposed. These networks are ideal for inclusion in quality-assurance pipelines as a way to determine scan quality both quickly and quantitatively. Finally, we provide methodology for detection of functionally connected areas of the brain via an application of the Likelihood Paradigm to resting-state fMRI data. The proposed technique allows for the control of both Type I and Type II error, resulting in improved inference when compared to traditional frequentist techniques.
Taken together, this dissertation aims to expand the understanding of the human brain through the application of modern statistical techniques via multiple imaging modalities that discern both the structural and functional properties of the brain.
Filename Size Approximate Download Time (Hours:Minutes:Seconds)
28.8 Modem 56K Modem ISDN (64 Kb) ISDN (128 Kb) Higher-speed Access Hainline.pdf 11.59 Mb 00:53:38 00:27:35 00:24:08 00:12:04 00:01:01