Scholarlink

<p>Large language models (LLMs) demonstrate significant potential in healthcare applications, but clinical deployment is limited by privacy concerns and insufficient medical domain training. This study investigated whether retrieval-augmented generation (RAG) can improve locally deployable LLM for radiology contrast media consultation. In 100 synthetic iodinated contrast media consultations we compared Llama 3.2-11B (baseline and RAG) with three cloud-based models—GPT-4o mini, Gemini 2.0 Flash and Claude 3.5 Haiku. A blinded radiologist ranked the five replies per case, and three LLM-based judges scored accuracy, safety, structure, tone, applicability and latency. Under controlled conditions, RAG eliminated hallucinations (0% vs 8%; χ²₍Yates₎ = 6.38, p = 0.012) and improved mean rank by 1.3 (Z = –4.82, p &lt; 0.001), though performance gaps with cloud models persist. The RAG-enhanced model remained faster (2.6 s vs 4.9–7.3 s) while the LLM-based judges preferred it over GPT-4o mini, though the radiologist ranked GPT-4o mini higher. RAG thus provides meaningful improvements for local clinical LLMs while maintaining the privacy benefits of on-premise deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Retrieval-augmented generation elevates local LLM quality in radiology contrast media consultation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月2日

查看全文
  • Akihiko Wada,
  • Yuya Tanaka,
  • Mitsuo Nishizawa,
  • Akira Yamamoto,
  • Toshiaki Akashi,
  • Akifumi Hagiwara,
  • Yayoi Hayakawa,
  • Junko Kikuta,
  • Keigo Shimoji,
  • Katsuhiro Sano,
  • Koji Kamagata,
  • Atsushi Nakanishi,
  • Shigeki Aoki

摘要

Large language models (LLMs) demonstrate significant potential in healthcare applications, but clinical deployment is limited by privacy concerns and insufficient medical domain training. This study investigated whether retrieval-augmented generation (RAG) can improve locally deployable LLM for radiology contrast media consultation. In 100 synthetic iodinated contrast media consultations we compared Llama 3.2-11B (baseline and RAG) with three cloud-based models—GPT-4o mini, Gemini 2.0 Flash and Claude 3.5 Haiku. A blinded radiologist ranked the five replies per case, and three LLM-based judges scored accuracy, safety, structure, tone, applicability and latency. Under controlled conditions, RAG eliminated hallucinations (0% vs 8%; χ²₍Yates₎ = 6.38, p = 0.012) and improved mean rank by 1.3 (Z = –4.82, p < 0.001), though performance gaps with cloud models persist. The RAG-enhanced model remained faster (2.6 s vs 4.9–7.3 s) while the LLM-based judges preferred it over GPT-4o mini, though the radiologist ranked GPT-4o mini higher. RAG thus provides meaningful improvements for local clinical LLMs while maintaining the privacy benefits of on-premise deployment.