Retrieval-Augmented Generation in Radiology: A Scoping Review of Architectures, Imaging Applications, and Directions...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology AI and Machine Learning
• Natural Language Processing in Healthcare

Background:

• Large language models (LLMs) in radiology face challenges with factual accuracy and hallucinations.
• Retrieval-augmented generation (RAG) integrates external knowledge retrieval to enhance generative models.
• Understanding RAG applications in radiology is crucial for advancing AI in medical imaging.

Purpose of the Study:

• To conduct a scoping review of Retrieval-Augmented Generation (RAG) applications in radiology and medical imaging.
• To characterize RAG system implementations, clinical tasks, retrieval strategies, and knowledge sources used.
• To assess the reported performance, limitations, and evaluation methodologies of RAG in radiology.

Main Methods:

• Systematic literature search across major databases (PubMed, Embase, Scopus, IEEE Xplore, arXiv).
• Identification and analysis of 45 studies utilizing RAG for radiology-related tasks.
• Categorization of studies by clinical application, retrieval strategy, knowledge source, and imaging modality.

Main Results:

• RAG most commonly applied to radiology report generation and question answering, primarily using dense retrieval.
• Chest radiography and X-ray tasks dominated applications; CT, MRI, and other subspecialties were under-represented.
• While RAG often improved performance over baselines, hallucinations persisted, and evaluation metrics were limited, potentially overestimating clinical readiness.

Conclusions:

• RAG demonstrates potential for enhancing factual grounding in radiology AI but faces persistent challenges.
• Current evaluation practices are insufficient for assessing real-world clinical readiness, safety, and bias.
• Future research must focus on improving retrieval quality, clinical grounding, safety, bias assessment, and efficiency for responsible translation.