Towards Real-World Generalization in Medical AI: From Task-Specific Models to Foundation Models

This thesis develops artificial intelligence (AI) systems that can work reliably in real-world hospitals, not just in laboratory tests. Many current medical AI models fail when used on patients from different hospitals or populations. To solve this, the research identifies and fixes hidden biases in medical data and designs new methods that help AI models adapt to different clinical settings. It also builds large, diverse medical image datasets and advanced multimodal foundation models that learn from millions of examples. These models show strong, trustworthy performance across hospitals and tasks, offering safer, more generalizable AI tools for doctors and patients.