放射科培训中的人工智能教育:有效性、障碍与未来方向的系统综述
本系统综述检索了截至2025年6月19日的PubMed、Web of Science、Embase和Google Scholar,从2646项研究中纳入14项评估放射科受训者AI培训项目效果的研究;92.9%(13/14)报告受训者表现改善,但存在过度依赖AI、复杂病例暴露有限和反馈不足等风险与担忧。
要点
- 14项研究中92.9%(13/14)报告AI培训改善受训者表现,1项(7.1%)未发现显著获益。
- 57.2%(8/14)研究报告诊断精确性和判读改善,57.2%(8/14)报告受训者信心提升。
- 5/14研究报告AI培训后敏感度、特异度和准确率中位数分别为72%、80%和81.3%,高于培训前的62.2%、78.9%和76.5%。
核心结论
- 92.9%的研究显示基于AI的培训可提升放射科受训者的知识、判读技能或诊断表现,尤其是低年资受训者;但安全应用需要标准化课程、多样化病例、导师指导、工作流程整合和稳健评估,并需要更大规模研究确认普适性。
研究设计与样本
- 系统综述,检索PubMed、Web of Science、Embase和Google Scholar截至2025年6月19日的文献,从2646项筛选研究中纳入14项评估放射科受训者AI培训项目表现的研究。
诊断标准与关键数值
- 5/14研究报告AI培训后敏感度、特异度和准确率中位数分别为72%、80%和81.3%,高于AI培训前的62.2%、78.9%和76.5%。
- 4/14研究报告AI知识评分改善。
- 92.9%(13/14)研究报告受训者表现改善,1项研究(7.1%,1/14)未发现显著获益。
临床应用与阅片要点
- 动手模拟和授课讲座是最常见的AI培训形式,分别占78.6%和71.4%。
- 85.7%(12/14)研究报告获得AI平台动手经验并增加AI知识,78.6%(11/14)报告接受AI辅助学习。
- 57.2%(8/14)报告诊断精确性和判读改善及受训者信心提升。
- 35.7%(5/14)报告参与基于AI的病例学习,7.1%(1/14)报告理解AI伦理和偏倚。
- 风险与担忧包括过度依赖AI、复杂或罕见病例暴露有限以及缺乏反馈。
局限性
- 各项研究的性能评估指标不一致。
- 1项研究(7.1%,1/14)未发现显著获益。
- 风险与担忧包括过度依赖AI、复杂或罕见病例暴露有限以及缺乏反馈。
- 需要更大规模研究确认普适性。
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
RATIONALE AND OBJECTIVES: The purpose of this systematic review study was to characterize the current landscape of various artificial intelligence (AI) education in radiology, summarizing existing curricula, outcomes, challenges, and future directions for effective integration into residency training.
MATERIALS AND METHODS: A comprehensive search of PubMed, Web of Science, Embase, and Google Scholar identified relevant published studies up to June 19, 2025.
RESULTS: Of the 2646 studies screened, 14 studies evaluated the performance of AI-based training programs for radiology trainees; among these, 92.9% (13/14) reported improvements in trainees' performance, including better diagnostic precision and interpretation (57.2%, 8/14), greater trainee confidence (57.2%, 8/14), hands-on experience with AI platforms (85.7%, 12/14), increased AI knowledge (85.7%, 12/14), engagement with AI-based case learning (35.7%, 5/14), understanding of AI ethics and bias (7.1%, 1/14), and acceptance of AI-assisted learning (78.6%, 11/14), whereas one study (7.1%, 1/14) found no significant benefit. Performance evaluation metrics varied across studies, with 35.7% (5/14) reporting a higher median of sensitivity, specificity, and accuracy (72%, 80%, and 81.3%) after AI training compared to before AI training (62.2%, 78.9%, and 76.5%, respectively), and 28.6% (4/14) showing improved AI knowledge scores. Hands-on simulations and didactic lectures were the most common AI training formats (78.6% and 71.4%). Risks and concerns included over-reliance on AI, limited exposure to complex or rare cases, and a lack of feedback. Recommendations highlighted the need for AI-faculty teaching, broader content coverage, and standardized multi-center AI-training programs to facilitate wider adoption.
CONCLUSION: 92.9% of studies showed that AI-based training can enhance radiology trainees' knowledge, interpretive skills, or diagnostic performance, especially for junior trainees; however, its safe adoption requires standardized curricula with diverse cases, mentorship, workflow integration, and robust evaluation, with larger studies needed to confirm generalizability.
引用
- Keshavarz P, Mohammadigoldar Z, Bedayat A, et al. Artificial Intelligence Education in Radiology Training: A Systematic Review of Effectiveness, Barriers, and Future Directions. Acad Radiol. 2025. PMID: 41242893. doi:10.1016/j.acra.2025.10.049
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。