Generative AI in Radiotherapy Imaging: From Cross-Modality Synthesis to Foundation Model-Based Segmentation

Imaging underpins nearly every stage of radiation therapy, yet the images obtained closest to treatment delivery are often the least reliable. Cone-beam CT is degraded by scatter artifacts and inaccurate Hounsfield unit values, whereas on-treatment MRI is limited by time constraints that preclude standard protocols. Generative artificial intelligence provides new solutions to these limitations. This review examines three areas of direct impact: cross-modality image synthesis, in which diffusion models generate CT-quality images foundation model-based segmentation and target delineation, where large-scale pretrained models generalize across modalities without task-specific retraining and multimodal architectures improve delineation by integrating imaging with clinical text; and generative approaches to post-treatment assessment. Persistent challenges, including image hallucinations, algorithmic biases, and data privacy, must be addressed before these tools can be deployed safely.