Development and Evaluation of a Retrieval-Augmented Generation System for Radiology Guidelines.
Authors
Affiliations (5)
Affiliations (5)
- Institute of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, School of Medicine and Health, TUM University Hospital, Technical University of Munich, Ismaninger Straße 22, 81675, Munich, Bavaria, Germany. [email protected].
- Institute of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, School of Medicine and Health, TUM University Hospital, Technical University of Munich, Ismaninger Straße 22, 81675, Munich, Bavaria, Germany.
- Department of Interventional Radiology, Medical Center - University of Freiburg, Faculty of Medicine, University of Freiburg, Hugstetter Straße 55, 79106, Freiburg, Baden-Württemberg, Germany.
- Institute of Diagnostic and Interventional Neuroradiology, Klinikum rechts der Isar, School of Medicine and Health, TUM University Hospital, Technical University of Munich, Ismaninger Straße 22, 81675, Munich, Bavaria, Germany.
- Institute for Cardiovascular Radiology and Nuclear Medicine, German Heart Center Munich, School of Medicine and Health, TUM University Hospital, Technical University of Munich, Lazarettstraße 36, 80636, Munich, Bavaria, Germany.
Abstract
Large language models (LLMs) have demonstrated remarkable capabilities in processing and generating domain-specific information. Their application in clinical decision-making is, however, still limited by unreliability and outdated knowledge. In time-sensitive medical environments, such as radiology, rapid access to accurate and up-to-date guidelines is crucial for optimal patient outcomes. The European Society of Urogenital Radiology (ESUR) guidelines provide such diagnostic and therapeutic recommendations. However, manual lookup is often time-consuming and inefficient. To address these challenges, we developed a retrieval-augmented generation (RAG) system that grounds LLM responses in authoritative guideline content. The system extracts, indexes, and retrieves information using a headline-based chunking approach and the all-mpnet-base-v2 embedding model. We evaluated its performance against both a standalone LLM and an enhanced iterative RAG system using 79 queries, assessing retrieval accuracy, factual correctness, completeness, and clinical usefulness. Both RAG systems significantly outperformed the standalone LLM in all metrics, with the enhanced model achieving the highest scores: Factual accuracy (0.89 vs. 0.68), completeness (4.20 vs. 3.05 on a 5-point Likert scale), and usefulness (3.99 vs. 3.09 on a 5-point Likert scale). The enhanced RAG pipeline showed minor but statistically not significant improvements over the standard version in terms of factual accuracy and completeness. While our results are promising, opportunities remain to improve retrieval accuracy and reduce hallucinations. Future refinements, like domain-specific embeddings and advanced query expansion, may further improve reliability. These findings suggest that grounded RAG systems have significant potential to enhance clinical guideline accessibility but require further validation before clinical deployment.