More about HKUST
Towards Data Scarcity in Medical Image Analysis
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "Towards Data Scarcity in Medical Image Analysis" By Miss Shuhan LI Abstract: Deep learning has revolutionized computer vision, with significant applications in classification, segmentation, image translation, and super-resolution. These techniques are particularly vital in medical imaging, where they enhance disease diagnosis and treatment. However, achieving accurate and robust predictions in this field often requires large training datasets and precise annotations, which are frequently scarce. This dissertation addresses three critical challenges: (1) Class Imbalance, as data for common diseases is readily available while rare diseases are underrepresented; (2) Lack of Annotations, since pixel-level annotation of medical images demands extensive expertise and time; and (3) Absence of Multi-modalities, where the need for complementary imaging insights is hindered by equipment limitations or patient safety concerns. To tackle these issues, the proposed work focuses on three key objectives: (1) Improving the performance of tail classes in classification tasks; (2) Generating image pairs with their corresponding pixel-level annotations using advanced generative models; and (3) Translating information from existing modalities to predict and simulate missing ones. For the first objective, two approaches are introduced: one employing few-shot learning to develop a robust feature extractor for head classes, and the other utilizing diffusion models to synthesize additional samples for tail classes, thereby mitigating class imbalance. For the second objective, a text-to-image synthesis model, conditioned on segmentation masks, is proposed, incorporating an additional reward model to enhance alignment between text and generated images. The third objective addresses the modalities of Optical Coherence Tomography (OCT) and OCT Angiography (OCTA), where the latter provides information on retinal blood flow but requires expensive equipment. We propose an OCT-to-OCTA modality translation method to generate high-quality OCTA images from OCT data. Through comprehensive experiments and analyses, this thesis presents potential solutions to alleviate data scarcity challenges in medical imaging, contributing to the advancement of these critical tasks. Date: Thursday, 16 January 2025 Time: 1:30pm - 3:30pm Venue: Room 3494 Lifts 25/26 Chairman: Prof. Minhua SHAO (CBE) Committee Members: Prof. Tim CHEN (Supervisor) Prof. Albert CHUNG Dr. Long CHEN Dr. Fengbin TU (ECE) Prof. Antoni Bert Chan (CityU)