Date of Award

Summer 6-27-2019

Publication Type

Doctoral Thesis

Degree Name



Electrical and Computer Engineering

First Advisor

Wu, Jonathan




Quantitative analysis of the brain structures on magnetic resonance (MR) images plays a crucial role in examining brain development and abnormality, as well as in aiding the treatment planning. Although manual delineation is commonly considered as the gold standard, it suffers from the shortcomings in terms of low efficiency and inter-rater variability. Therefore, developing automatic anatomical segmentation of human brain is of importance in providing a tool for quantitative analysis (e.g., volume measurement, shape analysis, cortical surface mapping). Despite a large number of existing techniques, the automatic segmentation of brain MR images remains a challenging task due to the complexity of the brain anatomical structures and the great inter- and intra-individual variability among these anatomical structures. To address the existing challenges, four methods are proposed in this thesis. The first work proposes a novel label fusion scheme for the multi-atlas segmentation. A two-stage majority voting scheme is developed to address the over-segmentation problem in the hippocampus segmentation of brain MR images. The second work of the thesis develops a supervoxel graphical model for the whole brain segmentation, in order to relieve the dependencies on complicated pairwise registration for the multi-atlas segmentation methods. Based on the assumption that pixels within a supervoxel are supposed to have the same label, the proposed method converts the voxel labeling problem to a supervoxel labeling problem which is solved by a maximum-a-posteriori (MAP) inference in Markov random field (MRF) defined on supervoxels. The third work incorporates attention mechanism into convolutional neural networks (CNN), aiming at learning the spatial dependencies between the shallow layers and the deep layers in CNN and producing an aggregation of the attended local feature and high-level features to obtain more precise segmentation results. The fourth method takes advantage of the success of CNN in computer vision, combines the strength of the graphical model with CNN, and integrates them into an end-to-end training network. The proposed methods are evaluated on public MR image datasets, such as MICCAI2012, LPBA40, and IBSR. Extensive experiments demonstrate the effectiveness and superior performance of the three proposed methods compared with the other state-of-the-art methods.