神经放射科医生对大型语言模型的认知、应用与担忧:国际调查
调查评估了医生对大型语言模型的知晓、应用偏好及对完整性和披露问题的担忧。结果显示受访者对LLM持谨慎乐观态度,倾向于流程辅助类应用而非影像解读。
要点
- 受访者对LLM知晓度高但概念有盲区,16.7%误将DALL·E归为LLM。
- 医生更认可指南推荐和方案选择等辅助用途,影像解读认可度仅22.0%。
- 对伪造数据和输出偏倚担忧普遍,多数支持AI检测工具与使用记录。
核心结论
- 受访者普遍支持LLM谨慎融入神经放射工作,偏好非影像解读类的流程辅助应用。
- 同时多数受访者强调披露、监管和针对性教育的重要性。
研究设计与样本
- 研究类型为国际性匿名自愿SurveyMonkey问卷调查。
- 265名受访者开始调查,排除后209名纳入分析,其中64.6%为男性。
- 分类变量以计数和百分比总结,Likert项目以加权均数和响应分布总结。
- 预设亚组分析采用卡方或Fisher精确检验及非参数检验,并在预设区块内进行Holm多重性控制。
诊断标准与关键数值
- ChatGPT知晓率为83.3%,14.8%对所有列出模型均不熟悉,16.7%将DALL·E误分类为LLM。
- 指南类推荐认可度为75.6%,方案选择认可度为60.8%,影像解读认可度为22.0%。
- 对抄袭或数据伪造的担忧占82.1%,对输出不准确或有偏倚的担忧占75.8%,对问责的担忧占72.1%。
- 77.0%支持使用AI检测工具,73.1%支持按照示例期刊政策记录LLM使用。
临床应用与阅片要点
- 在日常工作中可优先将LLM用于指南查询和方案选择等流程辅助任务。
- 影像解读场景对LLM的临床应用认可度低,不应作为独立阅片依据。
- 使用LLM辅助报告或科研时应注意记录使用情况并配合AI检测工具。
局限性
- 研究本身为问卷调查,结果受自愿参与和回忆偏倚影响。
- 调查样本可能存在选择偏倚,不能完全代表所有神经放射科医生。
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
OBJECTIVES: Large language models (LLMs) are increasingly used in medicine and research, but neuroradiologists' awareness, perceived utility, and concerns about integrity and disclosure remain incompletely characterized. This survey aimed to assess radiologists' awareness and perceptions of LLMs in clinical and research domains.
MATERIALS AND METHODS: An anonymous, voluntary SurveyMonkey survey was distributed internationally (October 1, 2024 to March 31, 2025) via neuroradiology society newsletters/membership channels and social media. Categorical variables were summarized as counts and percentages; Likert items were summarized using weighted means and response distributions. Item-level complete-case denominators were reported. Prespecified subgroup analyses used chi-square/Fisher exact tests (categorical) and nonparametric tests (ordinal), with Holm multiplicity control within prespecified multi-item question blocks and within each subgroup factor.
RESULTS: A total of 265 respondents started the survey; after exclusions, 209 were included in the analytic sample, of whom 64.6% were male. Awareness of LLMs was high (ChatGPT: 83.3%), but knowledge gaps persisted (14.8% unfamiliar with all listed models; 16.7% misclassified DALL·E as an LLM). Respondents most frequently endorsed bounded, workflow-adjacent clinical applications, including guideline-based recommendations (75.6%) and protocol selection (60.8%), with lower endorsement for image interpretation (22.0%). Concerns were common regarding plagiarism/data fabrication (82.1%), inaccurate or biased outputs (75.8%), and accountability (72.1%), alongside support for AI-detection tools (77.0%) and documentation of LLM use aligned with an example journal policy (73.1%).
CONCLUSION: Survey respondents reported cautious optimism toward LLM integration, favoring workflow-adjacent applications while emphasizing disclosure, oversight, and targeted education.
引用
- Rai, Soni, Ora, et al. Large language models in neuroradiology: an international survey of awareness, applications, and concerns. Neuroradiology. 2026. PMID: 42700224. doi:10.1007/s00234-026-04176-5
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。