More about HKUST
Towards Data Scarcity in Medical Image Analysis
PhD Thesis Proposal Defence Title: "Towards Data Scarcity in Medical Image Analysis" by Miss Shuhan LI Abstract: Medical image analysis is crucial for providing precise diagnostic support through advanced deep-learning techniques. Despite the rapid growth in methodologies and applications within this field, data scarcity remains a significant challenge. This scarcity typically manifests in three forms: limited annotation, imbalanced classes, and the absence of multi-modality data. Addressing these issues involves two primary solutions: optimizing learning from limited datasets and augmenting data through generative models. For the first solution, this thesis introduces a novel few-shot learning model for the classification of rare skin diseases, named SCAN. SCAN features a dual-branch architecture designed to learn both inter and intra-class representations, effectively leveraging as few as one to five images to train the model. This approach dramatically enhances classification accuracy, particularly in scenarios characterized by imbalanced class distributions. Addressing the second solution, we first propose an iterative online image synthesis framework designed to mitigate class imbalance. Unlike traditional methods that generate synthetic samples before model training, our framework integrates synthesis with training, dynamically adjusting the quantity of generated images based on real-time training performance. Then, we expand the scope of generative control by incorporating detailed conditions beyond class labels, enabling the production of images with higher fidelity and diversity. This enhancement is also expected to significantly improve performance in downstream tasks. Lastly, to overcome the challenges posed by the absence of multi-modality data, we introduce TransPro, a transformative model that employs a 3D Generative Adversarial Network (GAN) coupled with two auxiliary tasks to translate between Optical Coherence Tomography (OCT) and Optical Coherence Tomography Angiography (OCTA). TransPro not only facilitates modality conversion but also enriches contextual information and improves the quality of vessel regions in the generated images. Overall, this thesis contributes to the fields of few-shot learning and generative modeling by proposing innovative solutions tailored to overcome the data scarcity challenges in medical image analysis. Date: Tuesday, 22 October 2024 Time: 10:00am - 12:00noon Venue: Room 5501 Lifts 25/26 Committee Members: Prof. Tim Cheng (Supervisor) Dr. Xiaomeng Li (Co-supervisor) Dr. Hao Chen (Chairperson) Dr. Long Chen