- Starts: 1:00 pm on Tuesday, September 26, 2023
- Ends: 3:00 pm on Tuesday, September 26, 2023
Title: Label-efficient Deep Learning for 3D Medical Image Analysis: Self-supervised and Generative Approaches
Presenter: Li Sun
Advisor: Professor Kayhan Batmanghelich
Chair: Professor Archana Venkataraman,
Committee: Professor Kayhan Batmanghelich, Professor Archana Venkataraman, Professor Ioannis Paschalidis.
Abstract: Deep learning methods have achieved state-of-the-art performance in various tasks of medical image analysis. However, the success relies heavily on the expensive and time-consuming collection of large quantities of labeled data, which is not always available. Compared to natural images, medical images have domain-specific anatomical context, which is less explored in previous research. In this work, we explore self-supervised and generative methods to improve the label efficiency of deep learning methods for 3D medical image analysis. This prospectus contains three prongs. First, we introduce a two-stage contrastive framework that learns fine-grained disease-related representation by leveraging anatomical context. Further, while previous methods are constrained by the available memory in terms of generation resolution, we propose a hierarchical GAN that can generate high-resolution 3D images for data augmentation. For future work, we will explore learning location-sensitive representation and develop a generative diffusion model that leverages anatomical prior.
- PHO 339