More about HKUST
DOMAIN-GENERALIZED RECOGNITION AND GENERATION FOR MEDICAL IMAGE ANALYSIS
PhD Thesis Proposal Defence Title: "DOMAIN-GENERALIZED RECOGNITION AND GENERATION FOR MEDICAL IMAGE ANALYSIS" by Mr. Haoxuan CHE Abstract: Medical image analysis plays a crucial role in modern healthcare, enabling automated interpretation and assistance in clinical decision-making. While deep learning has demonstrated remarkable potential and achieved state-of-the-art performance in various medical tasks in this field, significant challenges emerge when deploying these models in real-world clinical settings, particularly in the context of domain insufficiency and domain isolation. Domain insufficiency arises from the limited availability and diversity of training data within individual medical centers, which contributes to poor generalization and significant performance degradation when models are deployed in new clinical environments. Additionally, while valuable medical data exists across institutions, it remains isolated in separate silos owing to strict privacy regulations, and proprietary concerns, hindering the development of robust, generalizable models. To address these challenges, our research contributions are organized into two complementary directions. First, we develop a series of frameworks that learn under domain insufficiency by enhancing model robustness and generalization capability when training with limited source domains. Our approaches focus on learning domain-invariant representations, adapting to quality variations, and handling data distribution shifts. Second, we propose privacy-preserving collaborative learning solutions that overcome domain isolation by enabling effective knowledge sharing across institutions while maintaining data privacy. These approaches span from generative model-based to large model-based knowledge sharing, creating a bridge between isolated medical data silos. Through extensive experiments across various medical recognition and generation tasks, we demonstrate that our proposed approaches significantly improve cross-domain performance while maintaining practical efficiency and privacy requirements. These contributions advance both the technical framework and practical applications of medical image analysis, working towards more generalizable and deployable AI systems in real-world clinical settings. Date: Friday, 7 March 2025 Time: 3:00pm - 5:00pm Venue: Room 4472 Lifts 25/26 Committee Members: Dr. Hao Chen (Supervisor) Dr. Dan Xu (Chairperson) Dr. Long Chen