Scoping Review of RAG in Radiology Finds Gains but Overestimates Clinical Readiness

Scoping review of 45 radiology RAG studies: most showed performance gains over non-retrieval baselines, but evaluation relied on automated metrics, safety and bias were rarely assessed, and clinical readiness is likely overestimated.

• RAG was most often applied to report generation and question answering; dense retrieval dominated over sparse/hybrid methods.
• Most studies reported gains over non-RAG baselines, but hallucinations and errors persisted; evaluation relied on automated metrics with minimal safety or bias assessment.
• Applications skewed heavily to chest X-ray; few studies addressed CT, MRI, or other modalities/subspecialties.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.