Medical Image Analysis with Less and Noisy Labels

Abstract

Deep learning has revolutionized medical image analysis, but its success heavily relies on supervised training with large, clean-labeled datasets. Acquiring such datasets is both costly and labor-intensive, often requiring expert annotations from medical professionals. To address this challenge, Medical image analysis with Less and Noisy labels (Med-LN) has emerged as a promising solution, enabling deep learning on less or noisy labeled datasets. However, the challenges posed in the medical imaging domain remain under-explored. In this thesis, we first address the issue of less labeled datasets by leveraging both labeled and unlabeled data. Then, we explore the common issue of noisy labels in medical datasets. Specifically, we focus on three key tasks in medical image analysis under Med-LN conditions: medical image enhancement, medical image segmentation, and medical visual question answering.