文献速递胸部骨肌腹部DRUS

本地大语言模型生成患者友好型影像报告:质量与可读性的专家评估

Generating patient-friendly radiology reports with a local large language model: expert evaluation of quality and readability
RSS 订阅 · PubMed 顶刊 · 影像 AI · 2026-10-30

本研究回顾性纳入荷兰初级保健的360份胸部、膝、腰椎、肩部X线和肩部、腹部超声报告,用本地部署的Gemma-3(12B参数)将报告主体和结论部分翻译为患者友好版本。结论部分的准确度中位数高于主体报告,但12.5%的报告出现临床相关幻觉,提示需临床监督。

要点

  • 本地LLM可生成患者友好型报告且避免数据外传,结论部分质量优于主体报告。
  • 患者友好型报告更长,可读性评分更高,但12.5%存在临床相关幻觉。
  • 临床落地前需医生审核,结论部分是最有希望的初始应用场景。

核心结论

  • 本地部署的LLM可生成荷兰语X线和超声报告的患者友好型翻译,结论部分表现优于主体报告。
  • 临床相关幻觉发生于45份报告(12.5%),仍是安全性顾虑,需临床监督和进一步评估。

研究设计与样本

  • 回顾性研究,纳入荷兰初级保健的360份胸部、膝、腰椎、肩部X线及肩部、腹部超声报告。
  • 使用本地部署的Gemma-3(12B参数)翻译报告主体(n=359)和结论(n=350)为患者友好版本。
  • 以5点Likert量表评估准确性和完整性,用文本长度和Flesch-Douma阅读难易度评分评估可读性,系统记录幻觉。
  • 统计方法包括Kruskal-Wallis、Wilcoxon符号秩和配对t检验。

诊断标准与关键数值

  • 患者友好型结论的准确度中位数为5(Q1-Q3 5-5),主体报告为4(3-5)。
  • 主体和结论部分的完整性评分均为5(5-5)。
  • 临床相关幻觉发生于45份报告(12.5%)。

临床应用与阅片要点

  • 可将本地LLM用于生成患者友好型报告结论,但须医生审核以识别临床相关幻觉。
  • 报告主体翻译质量低于结论部分,日常应用时应优先关注结论部分的生成。

局限性

  • 回顾性研究设计,样本来自荷兰初级保健,结论外推需谨慎。
  • 临床相关幻觉发生率12.5%,临床实施前需进一步患者中心验证。
英文原文(摘要)· 点击展开

Eur Radiol . 2026 Oct 29. doi: 10.1007/s00330-026-12971-0. Online ahead of print. ABSTRACT OBJECTIVES: To assess the feasibility of using a locally hosted large language model (LLM) to generate patient-friendly radiology reports while avoiding cloud transfer of report text. MATERIALS AND METHODS: In this retrospective study, 360 chest, knee, lumbar spine, and shoulder X-rays, and shoulder and abdominal ultrasound reports from Dutch primary care were included. Main report sections (n = 359) and conclusion sections (n = 350) were translated into patient-friendly versions using Gemma-3 (12B parameters) and assessed for accuracy and completeness on 5-point Likert scales. Readability was evaluated using text length and Flesch-Douma reading ease (FDRE) scores. Hallucinations were systematically documented. Statistical analyses included Kruskal-Wallis, Wilcoxon signed-rank, and paired t-tests. RESULTS: Patient-friendly conclusions achieved higher median accuracy scores than main reports (5 [Q1-Q3 5-5] vs 4 [3-5]), while completeness scores were high for both sections (5 [5-5]). Compared with original reports, patient-friendly reports were longer (median increase 65 words [48-84], p 60). Clinically relevant hallucinations occurred in 45 reports (12.5%). CONCLUSION: A locally hosted LLM demonstrated feasibility for generating patient-friendly translations of Dutch X-ray and ultrasound reports while avoiding cloud transfer of data, with better performance for conclusion sections than main reports. However, clinically relevant hallucinations remain a safety concern and indicate the need for clinician oversight and further evaluation before clinical implementation. KEY POINTS: Question Can a locally hosted LLM generate patient-friendly radiology reports while preserving data privacy and maintaining acceptable accuracy, completeness, and readability? Findings A locally hosted LLM generated patient-friendly report conclusions with higher quality scores than main reports, while clinically relevant hallucinations remained an important safety concern. Clinical relevance A locally hosted LLM has potential to generate patient-friendly radiology report translations without external cloud data transfer. Report conclusions appear to be the most promising initial application, although clinician oversight and further patient-centered validation remain essential before clinical implementation. PMID: 42957801 | DOI: 10.1007/s00330-026-12971-0

来源:PubMed 顶刊 · 影像 AI,查看原文。版权归原作者所有。

引用

  1. PubMed 顶刊 · 影像 AI:Generating patient-friendly radiology reports with a local large language model: expert evaluation of quality and readability

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。