Deep Learning Framework for Groupwise Medical Image Registration

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


PhD Thesis Defence


Title: "Deep Learning Framework for Groupwise Medical Image Registration"

By

Miss Ziyi HE


Abstract:

Groupwise image registration (GIR) is a fundamental task in medical imaging 
that facilitates the simultaneous deformation of cohorts of subjects towards 
a specified or implicit group center, enabling population-level analysis. 
Traditional optimization-based methods often suffer from substantial 
computational costs, which limits their application in clinical tasks. With 
the advent of deep learning methods, several learning-based studies have 
emerged which purely utilized the network as feature extractions, focusing 
on improving optimization processes instead of prediction.

This thesis aims to address these gaps by proposing a series of robust and 
efficient deep learning-based frameworks for medical image groupwise 
registration. We start with an unsupervised end-to-end groupwise 
registration framework with multi-step updating mechanism to align the group 
subjects into the implicit group center without constructing template 
images. After that, we propose a template synthesis method based on the 
generative adversarial network and an auxiliary segmentation module to 
generate high-quality template images. To extend the framework from the 
setting of fixed group size to arbitrary group sizes, we present 
InstantGroup which deploys a Siamese variational autoencoder for encoding 
pairs of inputs and generating the template image of the minimum group unit 
through latent vectors' arithmetic. The method exhibits promising 
flexibility and efficiency. To further improve the adaptability and 
performance of InstantGroup, we propose TAG to integrate test-time training 
to InstantGroup to deal with target groups of multiple resolutions. 
Experiments illustrate the method outperforms state-of-the-art benchmarks as 
well as maintains the robustness across various scenarios.


Date:                   Wednesday, 5 March 2025

Time:                   4:00pm - 6:00pm

Venue:                  Room 4472
                        Lifts 25/26

Chairman:               Dr. Yuanliang ZHAI (LIFS)

Committee Members:      Prof. Albert CHUNG (Supervisor)
                        Prof. Pedro SANDER
                        Dr. Hao CHEN
                        Prof. Shing Yu LEUNG (MATH)
                        Prof. Anqi QIU (PolyU)