RadioRAG: Online Retrieval-augmented Generation for Radiology Question Answering.

Abstract

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To evaluate diagnostic accuracy of various large language models (LLMs) when answering radiology-specific questions with and without access to additional online, up-to-date information via retrieval-augmented generation (RAG). Materials and Methods The authors developed Radiology RAG (RadioRAG), an end-to-end framework that retrieves data from authoritative radiologic online sources in real-time. RAG incorporates information retrieval from external sources to supplement the initial prompt, grounding the model's response in relevant information. Using 80 questions from the RSNA Case Collection across radiologic subspecialties and 24 additional expert-curated questions with reference standard answers, LLMs (GPT-3.5-turbo, GPT-4, Mistral-7B, Mixtral-8 × 7B, and Llama3 [8B and 70B]) were prompted with and without RadioRAG in a zero-shot inference scenario (temperature ≤ 0.1, top- P = 1). RadioRAG retrieved context-specific information from www.radiopaedia.org. Accuracy of LLMs with and without RadioRAG in answering questions from each dataset was assessed. Statistical analyses were performed using bootstrapping while preserving pairing. Additional assessments included comparison of model with human performance and comparison of time required for conventional versus RadioRAG-powered question answering. Results RadioRAG improved accuracy for some LLMs, including GPT-3.5-turbo [74% (59/80) versus 66% (53/80), FDR = 0.03] and Mixtral-8 × 7B [76% (61/80) versus 65% (52/80), FDR = 0.02] on the RSNA-RadioQA dataset, with similar trends in the ExtendedQA dataset. Accuracy exceeded (FDR ≤ 0.007) that of a human expert (63%, (50/80)) for these LLMs, while not for Mistral-7B-instruct-v0.2, Llama3-8B, and Llama3-70B (FDR ≥ 0.21). RadioRAG reduced hallucinations for all LLMs (rates from 6-25%). RadioRAG increased estimated response time fourfold. Conclusion RadioRAG shows potential to improve LLM accuracy and factuality in radiology question answering by integrating real-time domain-specific data. ©RSNA, 2025.