More about HKUST
DOMAIN-GENERALIZED RECOGNITION AND GENERATION FOR MEDICAL IMAGE ANALYSIS
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "DOMAIN-GENERALIZED RECOGNITION AND GENERATION FOR MEDICAL IMAGE ANALYSIS" By Mr. Haoxuan CHE Abstract: Medical image analysis plays a pivotal role in modern healthcare by facilitating automated interpretation and enhancing clinical decision-making. While deep learning has achieved state-of-the-art performance in various medical imaging tasks, its deployment in real-world clinical settings remains challenging, particularly due to issues related to domain data limitation and domain data isolation. Domain data limitation stems from the limited availability and diversity of training data within individual medical institutions, leading to poor model generalization and performance degradation when applied to unseen clinical environments. Simultaneously, although vast amounts of valuable medical data are distributed across institutions, strict privacy regulations and proprietary constraints result in domain data isolation, preventing effective cross- institutional knowledge sharing and limiting the development of robust and generalizable models. To address these challenges, this thesis contributes in two key directions. First, it introduces a series of frameworks designed to enhance model robustness and generalization in scenarios characterized by domain data limitation. These approaches focus on learning domain-invariant representations, mitigating quality variations, and adapting to distributional shifts, thereby improving performance across diverse clinical settings. Second, it proposes privacy-preserving collaborative learning strategies to mitigate domain data isolation by enabling knowledge transfer across institutions without compromising data privacy. These strategies range from generative model- based solutions to large-scale model-based knowledge sharing, facilitating collaboration while maintaining compliance with privacy constraints. Through extensive experimental validation across various medical recognition and generation tasks, this thesis demonstrates that the proposed methodologies significantly enhance cross-domain performance while maintaining computational efficiency and privacy safeguards. The findings contribute to both the theoretical advancement and practical deployment of AI- driven medical image analysis, moving towards the development of more generalizable and clinically deployable systems. Date: Monday, 12 May 2025 Time: 4:30pm - 6:30pm Venue: Room 5501 Lifts 25/26 Chairman: Prof. Dan LI (CBE) Committee Members: Dr. Hao CHEN (Supervisor) Dr. Long CHEN Dr. Dan XU Dr. Jiguang WANG (LIFS) Prof. Jing QIN (PolyU)