文献速递技术与安全

2025年放射学中作为医疗器械的人工智能:欧盟、美国与中国的监管图景

Artificial intelligence as medical device in radiology in 2025: the regulatory scenario in the EU, USA, and China.
PubMed 文献 · Eur Radiol · 2026-03-15

本文综述欧盟、美国和中国将放射学人工智能工具作为医疗器械的监管框架。欧盟涉及2017年医疗器械法规、2018年通用数据保护条例和2024年人工智能法案的风险分级方法;美国涉及FDA三条上市前路径及生命周期管理、上市后监测和自适应算法要求;中国涉及国家药品监督管理局2022年人工智能医疗器械软件分类定义指南及六项上市前技术审查指导原则。三地监管在患者数据保护、风险评估与分类、公平性与泛化性、透明性与可解释性以及人类监督需求方面呈现显著趋同。

要点

  • 欧盟、美国和中国监管均要求放射学AI工具作为医疗器械进行风险分类管理。
  • 三地监管趋同于患者数据保护、透明性、偏倚缓解和人类监督等核心要求。
  • 美国FDA强调真实世界证据对AI工具性能监测的重要性。

核心结论

  • 欧盟、美国和中国对放射学AI医疗器械的监管虽存在差异,但在患者数据保护、风险评估与分类、公平性与泛化性、透明性与可解释性以及人类监督需求方面呈现有意义的趋同,放射学界将面对比预期更为同质化的全球监管环境。

研究设计与样本

  • 综述性文章,描述欧盟、美国和中国三个司法辖区的监管场景;欧盟纳入2017年医疗器械法规(将AI工具列为「主动」医疗器械)、2018年通用数据保护条例和2024年人工智能法案的风险分级方法;美国纳入FDA三条上市前路径(510(k)实质性等同、De Novo新器械分类、Premarket Approval高风险应用)及近期关于生命周期管理、上市后监测和自适应算法的法规;中国纳入国家药品监督管理局2022年人工智能医疗器械软件分类定义指南及六项放射学与医学影像AI医疗器械上市前技术审查指导原则。

诊断标准与关键数值

  • 欧盟2017年医疗器械法规将AI工具归为「主动」医疗器械,2024年人工智能法案采用基于风险的方法进行分级。
  • 美国FDA 510(k)路径通过实质性等同获得许可,De Novo分类适用于无谓词器械的新器械,Premarket Approval适用于需要严格临床证据的高风险应用。
  • 中国国家药品监督管理局2022年指南用于判断工具是否为AI医疗器械并划分相关风险等级,已发布六项放射学与医学影像AI医疗器械上市前技术审查指导原则。

临床应用与阅片要点

  • 放射科医生和开发者可参考三地监管趋同的核心要求(数据保护、风险分类、透明性、偏倚缓解、人类监督)来规划AI工具的安全整合。
  • 中国法律强制要求对制造商进行去标识化以保护患者隐私。
  • 美国近期法规强调AI工具性能的真实世界证据,提示需关注上市后监测和自适应算法的生命周期管理。

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

In the last decade, advanced AI methods were applied to radiology, providing tools for clinical practice. Regulations across countries are a relevant topic, considering that AI tools must be regarded as medical devices. We describe the regulatory scenarios in the EU, USA, and China. For the EU, we considered the 2017 Medical Device Regulation, including AI tools as "active" medical devices, the 2018 General Data Protection Regulation, protecting data privacy, and the risk-based approach by the 2024 AI Act. For the USA, we considered the three FDA premarket pathways: the 510(k)-clearance demonstrating substantial equivalence, the De Novo classification for novel devices without predicates, and the Premarket Approval process for high-risk applications demanding rigorous clinical evidence; recent regulations regarded lifecycle management, post-marketing surveillance and adaptive algorithms, underscoring the importance of real-world evidence of AI tool performance. For China, the role of the 2022 Guidance for classification and definition of AI medical software by the National Medical Products Administration is illustrated, describing how to determine whether a tool is an AI-enabled medical device, categorizing the associated risk level. The NMPA published six premarket technical review guides related to AI-enabled medical devices in radiology and medical imaging; protection of patient privacy is enforced by the law and de-identification is mandatory for manufacturers. Regulations in these three scenarios show meaningful convergences about patient's data protection, risk assessment and classification, ensuring equity and generalizability, transparency and explainability, and the need of human oversight. The radiology community will act in a world scenario more homogeneous than expected. KEY POINTS: Question Regulatory fragmentation across the EU, USA, and China creates uncertainty for radiology AI development, validation, and clinical adoption, requiring clearer international harmonization. Findings Despite differences, regulations in the EU, USA, and China converge on core requirements: patient data protection, risk classification, transparency, bias mitigation, and human oversight. Clinical relevance By highlighting convergences across major jurisdictions, this review informs radiologists and developers on safe integration of AI tools, ensuring patient safety, equity, and trustworthy adoption in clinical practice.

来源:Eur Radiol,查看原文。版权归原作者所有。

引用

  1. Pesapane F, De Cecco C, Wang H, et al. Artificial intelligence as medical device in radiology in 2025: the regulatory scenario in the EU, USA, and China. Eur Radiol. 2026. PMID: 41832928. doi:10.1007/s00330-026-12460-4

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。