Foundation Models in Radiation Oncology: Applications, Challenges, and Path to Trustworthy AI
Description
The rapid emergence of foundation AI models, large-scale pre-trained architectures such as vision transformers, diffusion models, and multimodal encoders, has ushered in a transformative era in medical image analysis. Leveraging massive natural and/or medical datasets, these models exhibit strong zero-shot and few-shot learning capabilities, cross-modality generalization, and the potential to unify diverse imaging tasks under a common representation framework. This session will critically examine the evolving role of foundation models in medical imaging and medical physics, with applications spanning classification, segmentation, registration, and automated report and treatment plan generation. The session will also emphasize the importance of interdisciplinary collaboration, bridging medical physics, medical imaging, language, computer vision, and clinical practice, and explore how foundation models may serve as the backbone for digital twin frameworks, image-guided radiotherapy, or real-time adaptive therapies. Special focus will be given to risks including model over-reliance, hallucinations, and domain shifts, with calls for robust validation, benchmarking, and ethical deployment, to enable trustworthy AI in radiation oncology.