人工智能驱动的胸部X线报告起草:韩国胸部放射学会基于专家调查的2025年立场声明
韩国胸部放射学会基于20名胸部放射科医师使用AI工具(KARA-CXR 1.0.0.3)审阅60例胸部X线(每个临床场景10例)的两轮Delphi调查(12个关键问题、72个条目,共识需≥70%一致)制定立场声明。72个条目中41个(56.9%)达成共识,受访者对准确性和临床整合多持中立态度,仅健康筛查场景获得正面评价,反对在常规实践中单独使用该工具。学会支持仅在健康筛查场景中经放射科医师验证后使用AI自动起草胸部X线报告,反对常规实践中单独使用,建议在采用前优化性能并开展学会认可的教育和指南。
要点
- 仅健康筛查场景对AI自动起草胸部X线报告持正面评价。
- 反对在常规实践中单独使用该AI工具,须经放射科医师验证。
- 72个条目中41个(56.9%)达成共识,共识阈值为≥70%一致。
核心结论
- KSTR支持仅在健康筛查场景中、经放射科医师验证后使用AI自动起草胸部X线报告,反对在常规实践中单独使用。
- KSTR建议在采用前优化性能并开展学会认可的教育和指南。
研究设计与样本
- 制定机构为韩国胸部放射学会(KSTR),方法为基于专家调查的立场声明。
- 20名胸部放射科医师在各自工作场所独立审阅60例胸部X线病例作为调查准备,随后进行两轮Delphi调查。
- 60例病例均匀分布于健康筛查、住院、急诊科、重症监护病房、呼吸科门诊和非呼吸科门诊6个临床场景,每个场景10例。
- 参与者分别选择自己曾处理过的胸部X线病例,整个选择和审阅流程在1个月内完成。
- 参与者使用AI自动报告起草工具(KARA-CXR,版本1.0.0.3;KakaoBrain,首尔,韩国),对12个关键问题(72个条目)按9点Likert量表评分,共识要求≥70%一致。
诊断标准与关键数值
- 共识判定标准为≥70%一致,72个条目中41个(56.9%)达成共识。
- 受访者对大多数关于准确性和临床整合的问题持中立立场,对工具既不印象深刻也不失望。
- 仅健康筛查检查获得正面评价;对在常规实践中单独使用该AI工具持反对意见。
- 参与者强调在部署前需要进一步优化性能,并主张在采用前开展学会认可的教育和指南。
临床应用与阅片要点
- KSTR支持仅在健康筛查场景中、经放射科医师验证后使用AI自动起草胸部X线报告。
- KSTR反对在常规实践中单独使用该AI工具。
- KSTR建议在采用前优化性能并开展学会认可的教育和指南。
局限性
- 该立场声明源自专家对软件在适度病例集上使用的Delphi调查,工具的准确性、可靠性和临床效用仍不确定。
- 参与者评分为主观Likert量表,共识仅覆盖72个条目中的41个(56.9%)。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: Generative artificial intelligence (AI) systems can be used to draft automated chest X-ray (CXR) reports. Although promising in terms of efficiency and workforce shortages, their accuracy, reliability, and clinical utility remain uncertain. This article presents the Korean Society of Thoracic Radiology (KSTR) position statement on AI-assisted CXR report drafting, derived from a Delphi survey of experts who used the software on a modest case set.
MATERIALS AND METHODS: Twenty thoracic radiologists completed a Delphi survey after reviewing 60 CXR cases using an AI-based tool for automated report drafting (KARA-CXR, version 1.0.0.3; KakaoBrain, Seoul, Republic of Korea). Prior to the Delphi survey, the participants individually reviewed 60 CXR cases at their respective workplaces as part of the survey preparation process. The 60 cases were distributed evenly across six clinical settings (health screening, inpatient, emergency department, intensive care unit, respiratory outpatient, and non-respiratory outpatient), with 10 cases in each setting. The participants individually selected CXR cases in which they had worked. The entire selection and review processes were completed within 1 month. Subsequently, two Delphi rounds were conducted. Participants rated 12 key questions (72 items) regarding the clinical applicability of the AI-based tool on a 9-point Likert scale. Consensus required ≥70% agreement.
RESULTS: Consensus emerged for 41 of 72 items (56.9%). Respondents adopted a neutral stance on most questions concerning accuracy and clinical integration; they were neither impressed nor disappointed with the tool. A favorable view emerged only for health-screening examinations. Conversely, the stand-alone use of the AI-based tool in routine practice was opposed. Participants stressed the need for further performance optimization before deployment and advocated society-endorsed education and guidelines before adoption.
CONCLUSION: The KSTR supports the use of an AI-based automated CXR report-drafting tool only in health-screening settings with radiologist validation and opposes its standalone use in routine practice, recommending performance optimization and society-endorsed education and guidelines before its adoption.
引用
- Jeong WG, Hwang EJ, Jin GY, et al. Artificial Intelligence-Driven Drafting of Chest X-Ray Reports: 2025 Position Statement From the Korean Society of Thoracic Radiology Based on an Expert Survey. Korean J Radiol. 2025. PMID: 41078017. doi:10.3348/kjr.2025.0457
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。