Unsupervised Domain Adaptation for Medical Image Segmentation using Generative Adversarial Alignment

PhD Thesis Proposal Defence


Title: "Unsupervised Domain Adaptation for Medical Image Segmentation using 
Generative Adversarial Alignment"

by

Miss Wen JI


Abstract:

Image segmentation is a fundamental task in medical image analysis as it 
provides the labeled regions of interest for the subsequent diagnosis and 
treatment. In recent years, deep neural networks have greatly improved 
medical image segmentation performance. However, the training of deep neural 
networks is highly dependent on data-driven methods. For many prediction 
tasks, especially dense-estimate tasks, the labeling process is usually 
expensive, making it difficult to collect large-scale and diverse datasets. 
Particularly, fully-supervised segmentation methods for medical images 
require precise annotations from experts, making them more 
resource-intensive. In addition, due to the domain shift between different 
modalities, when training a model on one domain and testing this model on 
another domain, the performance of the model will decrease significantly. 
This restricts the model from training on large-scale labeled datasets and 
then transferring the model to new datasets from different fields. 
Unsupervised domain adaptation (UDA) is an important research field in 
weakly supervised learning based on the idea of transfer learning. It 
provides an effective solution to reduce the domain shift between labeled 
source data and unlabeled target data.

Traditional UDA methods usually use a metric as the objective function to 
measure and minimize the discrepancy of high-dimensional features between 
the source domain and the target domain. Recently, the UDA methods based on 
generative adversarial learning have made some breakthroughs. Some UDA 
methods use the generation approaches as image generators to learn a mapping 
between the source and the target domains, which has also shown promising 
results in various applications.

This thesis proposal intends to utilize generative adversarial learning to 
establish alignments on the inter-domain and intra-domain perspectives, 
exploring UDA technologies to enhance the generalizability of deep models 
for cross-modality medical image segmentation.


Date:                   Friday, 9 May 2025

Time:                   4:30pm - 6:30pm

Venue:                  Room 2612B
                        Lifts 31/32

Committee Members:      Prof. Albert Chung (Supervisor)
                        Prof. Pedro Sander (Chairperson)
                        Dr. Dan Xu