Fast physics model driven method for brain image registration and robust single atlas guided methods for registration-based brain image segmentation

PhD Thesis Proposal Defence


Title: "Fast physics model driven method for brain image registration and
robust single atlas guided methods for registration-based brain image
segmentation"

by

Miss Yishan Luo


ABSTRACT:

Brain image registration and segmentation are two intensively studied
topics in medical image analysis field. The process of accurate
registration and segmentation of the images is crucial for accurate
diagnosis by clinical tools.

On one hand, image registration plays an important role in adding new
values to images, e.g., combination of structural and functional
information, disease diagnosis, statistical atlas model construction and
so on. In this proposal, we first propose one novel intensity-based image
registration method. A new similarity metric derived from a physics model
is designed for solving image registration problem. The proposed method,
namely registration with crystal dislocation energy, utilizes an elastic
interaction between the reference image and the moving image to drive the
registration process, which not only improves the registration accuracy,
but also provides a high convergence rate in the optimization procedure.

On the other hand, image registration can also facilitate segmentation.
Due to the low quality of medical brain images, it is not easy to rely on
the images alone to distinguish different brain structures, especially
those deep brain structures with weakly visible boundaries. Using a
pre-labeled atlas for segmenting target images is thus more preferable. In
the second part of this proposal, we propose two registration-based
segmentation methods. The first method explores the spatial dependency
relations among deep brain structures and builds a prior spatial
dependency tree in order to constrain their inter-relationships and
determine the structure-by-structure segmentation sequence. In the second
method, a new concept, i.e., groupwise segmentation, which uses one atlas
image to segment a population of target images simultaneously, is proposed
for the first time. It is based upon a Markov Random Field (MRF) model to
impose the consistency constraints among the population of target images
and to embed the prior shape information of the atlas. It is
experimentally demonstrated that the two proposed segmentation methods can
achieve relatively higher accuracy than the state-of-the-art methods.


Date:                   Thursday, 30 June 2011

Time:                   10:00am - 12:00noon

Venue:                  Room 3588
                         lifts 27/28

Committee Members:      Dr. Albert Chung (Supervisor)
                         Prof. Long Quan (Chairperson)
 			Dr. Huamin Qu
 			Dr. Chiew-Lan Tai


**** ALL are Welcome ****