Author: Vanessa Karlen Tidwell

Publisher:

Published: 2013

Total Pages: 82

ISBN-13:

Magnetic Resonance Imaging (MRI) is a key medical imaging technology. Through in vivo soft tissue imaging, MRI allows clinicians and researchers to make diagnoses and evaluations that were previously possible only through biopsy or autopsy. However, analysis of MR images by domain experts can be time-consuming, complex, and subject to bias. The development of automatic segmentation techniques that make use of robust statistical methods allows for fast and unbiased analysis of MR images. In this dissertation, I propose segmentation methods that fall into two classes---(a) segmentation via optimization of a parametric boundary, and (b) segmentation via multistep, spatially constrained intensity classification. These two approaches are applicable in different segmentation scenarios. Parametric boundary segmentation is useful and necessary for segmentation of noisy images where the tissue of interest has predictable shape but poor boundary delineation, as in the case of lung with heavy or diffuse tumor. Spatially constrained intensity classification is appropriate for segmentation of noisy images with moderate contrast between tissue regions, where the areas of interest have unpredictable shapes, as is the case in spinal injury and brain tumor. The proposed automated segmentation techniques address the need for MR image analysis in three specific applications: (1) preclinical rodent studies of primary and metastatic lung cancer (approach (a)), (2) preclinical rodent studies of spinal cord lesion (approach (b)), and (3) postclinical analysis of human brain cancer (approach (b)). In preclinical rodent studies of primary and metastatic lung cancer, respiratory-gated MRI is used to quantitatively measure lung-tumor burden and monitor the time-course progression of individual tumors. I validate a method for measuring tumor burden based upon average lung-image intensity. The method requires accurate lung segmentation; toward this end, I propose an automated lung segmentation method that works for varying tumor burden levels. The method includes development of a novel, two-dimensional parametric model of the mouse lungs and a multifaceted cost function to optimally fit the model parameters to each image. Results demonstrate a strong correlation (0.93), comparable with that of fully manual expert segmentation, between the automated method's tumor-burden metric and the tumor burden measured by lung weight. In preclinical rodent studies of spinal cord lesion, MRI is used to quantify tissues in control and injured mouse spinal cords. For this application, I propose a novel, multistep, multidimensional approach, utilizing the Classification Expectation Maximization (CEM) algorithm, for automatic segmentation of spinal cord tissues. In contrast to previous methods, my proposed method incorporates prior knowledge of cord geometry and the distinct information contained in the different MR images gathered. Unlike previous approaches, the algorithm is shown to remain accurate for whole spinal cord, white matter, and hemorrhage segmentation, even in the presence of significant injury. The results of the method are shown to be on par with expert manual segmentation. In postclinical analysis of human brain cancer, access to large collections of MRI data enables scientifically rigorous study of cancers like glioblastoma multiforme, the most common form of malignant primary brain tumor. For this application, I propose an efficient and effective automated segmentation method, the Enhanced Classification Expectation Maximization (ECEM) algorithm. The ECEM algorithm is novel in that it introduces spatial information directly into the classical CEM algorithm, which is otherwise spatially unaware, with low additional computational complexity. I compare the ECEM's performance on simulated data to the standard finite Gaussian mixture EM algorithm, which is not spatially aware, and to the hidden-Markov random field EM algorithm, a commonly-used spatially aware automated segmentation method for MR brain images. I also show sample results demonstrating the ECEM algorithm's ability to segment MR images of glioblastoma.