BRAIN SEGMENTATION VIA DIFFEOMORPHIC LIKELIHOOD FUSION AND ITS APPLICATIONS TO CLINICAL ANALYSES
Embargo until
Date
2014-05-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Johns Hopkins University
Abstract
The human brain is composed of a variety of structures, or regions of interest (ROIs), that are responsible for a range of functions. It is therefore an essential step in quantitative analyses to define these ROIs in the human brain anatomy. Due to its intricate formation and function, this anatomy is highly complex and that makes it challenging to segment the human brain. During the past decade, the development of neuroimaging techniques (Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and Positron Emission Tomography (PET)) has facilitated mathematical modeling algorithms for the reconstruction and representation of various human brain structures.
The primary goal of this thesis is to develop automated algorithms to accurately segment the human brain into anatomical regions based on two types of MR images -- the T1-weighted image and the diffusion tensor image (DTI). We propose an automated segmentation algorithm in the framework of Bayesian parameter estimation. This algorithm is posed in the augmented orbit of a union of multiple atlases. The estimator of the segmentation label is obtained by maximizing the posterior probability of the segmentation label given the observable image of the to-be-segmented subject. We assume that the posterior probability of interpreting a single observable image is shared across multiple different atlases, the anatomical information of which is optimally fused via a diffeomorphic likelihood fusion.
The ultimate goal of designing algorithms for neuroimaging studies is to be applicable to clinical studies. It is, therefore, natural and essential to apply the proposed segmentation algorithm to clinical studies and perform statistical analyses to study the functional mechanism of certain structures in the development of specific disorders. The most widely used metric for analysis is the volume measurement of an anatomical structure. In part, this thesis aims to develop statistical methods to examine additional, manifold-based, metrics in the setting of computational anatomy (CA) that can differentiate and predict disease states. We use the 2-D surface that contours the 3-D subvolume of the structure of interest to construct our manifold-based comparison metrics.
This thesis starts with a derivation of a Bayesian parameter estimation algorithm in the framework of multi-atlas random diffeomorphic orbit model. The maximum a posteriori (MAP) estimator is obtained by maximizing the fused likelihoods from multiple atlases, in an expectation maximization (EM) fashion, which we term multi-atlas likelihood fusion (MALF). We design a two-level hierarchical segmentation pipeline for T1-weighted brain images based on MALF. Validation results on a variety of datasets are presented. Following that, we apply the segmentation pipeline to two large-scale neuroimaging studies and demonstrate its power in the clinical study of various diseases. We also describe an extension of this algorithm to the segmentation of multi-contrast DTI images. Subsequent to that, we illustrate a statistical shape analysis method to contrast two groups in terms of localized shape area, as well as an extension to longitudinal neuroimaging studies, examining the temporal dynamics of these differences -- the local shape change rates. In the final portion of this thesis, we present a deformation-pattern based manifold learning and clustering approach applied to differentiating and predicting neurodegenerative diseases. Specifically, classification and prediction results on dementia of the Alzheimer type, based on the deformation pattern, will be presented.
Description
Keywords
Brain Segmentation, Diffeomorphic Likelihood Fusion, Statistical Shape Analysis, Image Registration, Manifold Learning and Clustering, Neuroinformatics