Abstract: 

Recent advances in deep learning and the growing availability of labeled medical data have propelled significant progress in medical image analysis. However, critical challenges persist, impeding the widespread adoption of medical AI technologies. My research group has concentrated on three key obstacles: label scarcity, label variability, and modality incompleteness. To tackle label scarcity, we have developed semi-supervised segmentation techniques that effectively leverage abundant unlabeled data. For label variability, our approach enables both diversified and personalized segmentation, adapting to the unique needs of individual cases. Additionally, we have designed models capable of generating full-modality images from incomplete inputs, addressing modality incompleteness. This talk will provide insights into these innovations, highlighting their potential to bridge gaps in medical image analysis and advance the field.