Radiology Report Conclusion Generation by Retrieval Augmented Generation

Radiology report generation has increasingly relied on fine-tuned large language models (LLMs), but these approaches demand substantial computational resources and raise concerns about patient data privacy. Building on our prior work, GPT-Driven Radiology Report Generation with FineTuned Llama 3, this paper introduces a Retrieval-Augmented Generation (RAG) framework as a scalable, privacy-conscious alternative. Our method leverages a vector database of 21,152 anonymized radiology reports (with findings and conclusions) from the University of Medicine and Pharmacy - Craiova’s Imaging Center. By encoding these into embeddings, the system retrieves semantically similar cases at inference time, dynamically integrating domain-specific knowledge into a generic LLM via API calls—eliminating resource-intensive finetuning. This approach reduces computational overhead, accelerates deployment cycles, and safeguards patient confidentiality by avoiding direct model exposure to sensitive data. Experiments on the same dataset as our earlier study demonstrate that RAG achieves diagnostic accuracy comparable to fine-tuned models while producing more concise conclusions. A blinded evaluation by board-certified radiologists revealed no statistically significant performance gap between the two methods. Crucially, RAG introduces novel advantages: (1) chain-of-thought prompting synergizes with retrieval to enhance diagnostic reasoning clarity, and (2) seamless database updates enable continuous improvement without retraining. By decoupling knowledge storage from generation, our framework supports agile adaptation to evolving clinical guidelines and regional reporting standards. These results position RAG as a transformative paradigm for AI-assisted radiology, balancing efficiency, privacy, and adaptability—a critical step toward real-world clinical integration.