More about HKUST
Deformable Medical Image Registration: From Classical Optimization to Learning-Based Frameworks
PhD Qualifying Examination
Title: "Deformable Medical Image Registration: From Classical Optimization to
Learning-Based Frameworks"
by
Mr. Yin Lai Lyndon CHAN
Abstract:
Deformable medical image registration is a fundamental problem in medical
image analysis. Its goal is to establish spatial correspondence between a
moving image and a fixed image acquired at different time points, from
different imaging modalities, or across different subjects. Accurate
deformable registration is essential for many clinical applications, including
disease monitoring, image-guided intervention, treatment planning, and
population-based studies. It also supports a wide range of downstream tasks,
such as image segmentation, atlas construction, tumor tracking, motion
estimation, and surgical navigation.
Over the past two decades, research in deformable registration has shifted
from iterative optimization algorithms to learning-based methods. Deep neural
networks have greatly reduced registration time while achieving accuracy
comparable to, and in some cases better than, conventional approaches. A
variety of architectures have been proposed, including convolutional neural
networks, cascaded networks, transformers, and diffusion-based models. More
recently, foundation models have attracted growing interest because of their
potential to improve robustness across datasets and reduce the impact of
domain shifts.
This survey reviews both the theoretical foundations and recent developments
in deformable medical image registration. We first introduce the classical
registration framework, covering transformation models, similarity measures,
and regularization strategies. We then summarize representative traditional
and deep learning-based registration methods, discussing their advantages as
well as their limitations. We introduce the learning-based frameworks in terms
of learning paradigms, architectures, training objectives, and evaluation
metrics. Finally, we examine several open challenges, including generalization
across heterogeneous datasets, pathological registration with missing
anatomical correspondences, deformation plausibility, and the practical issues
involved in clinical deployment.
Date: Wednesday, 15 July 2026
Time: 10:00am - 12:00noon
Venue: Room 4472
Lift 25/26
Committee Members: Prof. Albert Chung (Supervisor)
Dr. Dan Xu (Chairperson)
Dr. Xiaomin Ouyang