Efficient Magnetic Resonance Brain Image Registration and High Performance Registration-based Brain Image Segmentation

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Efficient Magnetic Resonance Brain Image Registration and High Performance 
Registration-based Brain Image Segmentation"

By

Miss Yishan Luo


Abstract

Brain Magnetic Resonance (MR) imaging is widely used in clinical practice for 
disease diagnosis, patient follow-up, therapy evaluation and human brain 
mapping. In order to extract useful information from MR images, image 
registration and image segmentation are two crucial procedures in practice. On 
one hand, image registration is necessary in order to compare or combine 
information obtained from different images. On the other hand, image 
segmentation is commonly used to extract more meaningful representation of an 
image for analysis. More importantly, in medical image analysis, this two 
processes are not independent but closely related to each other. Due to some 
image artifacts introduced in the imaging stage, automatic segmentation relying 
on the target image alone is still challenging for brain MR images. Therefore, 
registration-based segmentation is essential and commonly applied for 
simplifying the segmentation task. In this thesis, we make contributions in 
both image registration and registration-based segmentation areas.

In the first part, we propose a novel image registration method derived from a 
physics model, i.e., the crystal dislocation model. An analogy is made between 
the registration process and the dislocation system in physics, and thus an 
elastic interaction between the reference image and the moving image is derived 
to drive the registration process. It is proved that the proposed registration 
method can not only improve the registration accuracy, but also achieve a high 
convergence rate in the optimization procedure.

In the second part, we focus on improving the performance of registration-based 
segmentation. In registration-based segmentation methods, the target image is 
segmented through registering the atlas image to the target image and 
transforming the atlas tissue labels to the target image. An atlas is defined 
as the combination of an intensity image and its pre-segmented image. In this 
thesis, we first propose a new way for atlas construction, as reliable atlases 
can provide useful prior information for the ensuing segmentation. To construct 
the atlas(es) from a population of subjects, we propose to divide the whole 
population into several subgroups. Then a newly designed tissue-wise weighted 
groupwise registration method is implemented within each subgroup and produces 
one average atlas for each subgroup. The proposed atlas construction scheme is 
evaluated through using the constructed atlas(es) for segmentation. It is 
experimentally validated that our method outperforms other conventional ways 
for building the atlas.

The second contribution in registration-based segmentation is that a new 
concept, i.e., atlas-guided groupwise segmentation, is proposed. Groupwise 
segmentation uses one single atlas image as guidance to segment a population of 
target images simultaneously. It is developed based on a Markov Random Field 
(MRF) deformation model to impose the consistency constraints among the 
population of target images and to embed the prior shape information of the 
atlas. The experiment results demonstrate that the proposed groupwise 
segmentation method can achieve higher accuracy than the state-of-the-art 
registration-based segmentation methods.


Date:			Tuesday, 21 August 2012

Time:			10:00am – 12:00noon

Venue:			Room 3501
 			Lifts 25/26

Chairman:		Prof. Bertram Shi (ECE)

Committee Members:	Prof. Albert Chung (Supervisor)
 			Prof. Huamin Qu
 			Prof. Chiew-Lan Tai
 			Prof. Oscar Au (ECE)
                         Prof. Pheng-Ann Heng (Comp. Sci. & Engg., CUHK)


**** ALL are Welcome ****