Open-World Medical Image Analysis: Towards Robust and Generalizable Diagnostic Intelligence

This thesis develops artificial intelligence methods that make medical image analysis more reliable in real clinical settings. It studies how deep learning models can adapt to new hospitals, new data types, and newly emerging diseases without relying on fully labelled data. It designs techniques to reduce overconfident errors, limit bias toward known diseases, and combine images with clinical text. Experiments across multiple medical imaging tasks show improved recognition of known conditions and discovery of previously unseen diseases. This work supports safer and more adaptable use of AI systems in everyday healthcare practice.