Unsupervised Affine and Deformable Medical Image Registration with Convolutional Neural Networks

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


PhD Thesis Defence


Title: "Unsupervised Affine and Deformable Medical Image Registration with 
Convolutional Neural Networks"

By

Mr. Chi Wing MOK


Abstract

Reliable and accurate medical image registration plays an important role 
in medical imaging studies. It provides valuable information that is vital 
for clinical diagnosis, evolutionary evaluation and planning of treatment 
strategies. The treatment quality and diagnosis precision of these 
applications highly rely on an accurate registration result of automated 
image registration algorithms. Therefore, improving the robustness, 
accuracy and runtime of image registration is essential. Recently, there 
has been rapid adoption of deep learning-based medical image registration 
applications over the past few years. While existing deep learning-based 
image registration approaches significantly accelerate the image 
registration by circumventing the costly iterative optimization process in 
conventional methods, these methods often ignore the desirable 
diffeomorphic properties of the transformation and fail spectacularly in 
images with large displacement settings and in medical images with 
pathologies. The core of this thesis is to develop a novel learning-based 
image registration framework for deformable medical image registration, 
which addresses the aforementioned issues. The contributions are linked 
under the common theme of learning-based image registration, but stand on 
their own as valuable components within the image registration framework.

In particular, a new unsupervised symmetric image registration method is 
proposed, which maximizes the similarity between images within the space 
of diffeomorphic maps and estimates both forward and inverse 
transformations simultaneously. This original concept learns the image 
registration problem within the space of diffeomorphic maps, resulting in 
securing desirable diffeomorphic properties of the solution. Furthermore, 
we introduce a pioneering conditional deformable image registration 
architecture and learning paradigm for large deformation image 
registration and rapid hyperparameter tuning. By learning the conditional 
features that are correlated with the hyperparameters and utilizing the 
advantages of the multi-resolution architecture, our formulation achieves 
end-to-end fast image registration under large deformation settings and 
enables fast hyperparameter tuning in learning-based registration, where 
usual learning-based registration solutions do not succeed. Moreover, we 
present a new deep learning-based deformable registration method that 
jointly estimates regions with absent correspondence and bidirectional 
deformation fields in pre-operative and follow-up brain tumour MRI scans. 
Finally, we present a fast and robust learning-based algorithm with the 
coarse-to-fine vision transformer for 3D affine medical image 
registration. Our method naturally leverages the global connectivity and 
locality of the convolutional vision transformer and the multi-resolution 
strategy to learn the global affine registration. The experimental results 
demonstrated that our novel methods achieve higher robustness and 
registration quality, as compared with state-of-the-art approaches, in 
affine and deformable image registration under large deformation and with 
pathological images.


Date:			Tuesday, 23 August 2022

Time:			2:00pm - 4:00pm

Zoom Meeting:
https://hkust.zoom.us/j/98298954964?pwd=aStwUmkvaWNyTFo2U0JPL1JtcWpWQT09

Chairperson:		Prof. Hong Kam LO (CIVL)

Committee Members:	Prof. Albert CHUNG (Supervisor)
 			Prof. Pedro SANDER (Supervisor)
 			Prof. Qifeng CHEN
 			Prof. Dit Yan YEUNG
 			Prof. Angela WU (LIFS)
 			Prof. Kenneth WONG (HKU)


**** ALL are Welcome ****