KEY POINTS
- This scoping review covers generative AI, foundation models, large language models, vision foundation models and multimodal systems in radiation therapy physics. The authors performed a non-exhaustive search of PubMed, Web of Science and Google through March 2026, focusing on applications, knowledge gaps and translational challenges.
- The review distinguishes the technologies by function: generative AI produces new content, foundation models provide broadly pretrained representations adaptable across tasks, and large language models primarily address language-heavy workflows. Vision and vision-language models extend this framework to imaging, segmentation and multimodal reasoning.
- Proposed or emerging applications span virtually the full radiotherapy pathway: synthetic imaging, image reconstruction, segmentation, registration, treatment planning, clinical documentation, TG-263 structure-name standardization, patient communication, protocol interpretation, decision support, incident analysis, clinical-trial matching and education.
- One highlighted radiation oncology study extracted 15,724 clinical cases and fine-tuned LLaMA-2 and Mistral models. For LLaMA-2, modality-selection accuracy increased from 0.499 to 0.705, ICD-10 prediction from 0.180 to 0.642, and treatment-regimen ROUGE-1 from 0.075 to 0.531; more than 60% of generated regimens were judged clinically acceptable.
- Foundation and generative models are also moving into imaging tasks including MRI-to-synthetic-CT generation, promptable segmentation and dose-distribution synthesis. The authors emphasize that much of this evidence still comes from small, curated, single-centre datasets, potentially overstating generalizability.
- The major translational barriers are not simply model accuracy: hallucination, interpretability, reproducibility, demographic bias, privacy, cybersecurity, regulatory compliance and integration into existing clinical workflows all require explicit management.
- The review does not provide a pooled estimate of performance or demonstrate improved patient outcomes. Its search was intentionally non-exhaustive, and many cited applications remain proof-of-concept, making the paper better viewed as a roadmap than a validation standard.
CLINICAL TAKEAWAY
Generative and foundation models are no longer relevant only to documentation - they are beginning to touch imaging, planning, QA and decision support across radiotherapy. The review makes an equally important second point: impressive benchmark performance is not sufficient for clinical autonomy, and human oversight, local validation and governance remain essential.