指南与共识技术与安全胸部神经CTMR

医学影像诊断AI的伦理与治理挑战:系统综述与政策框架

Ethical and Governance Challenges of AI in Medical Imaging and Diagnostics: A Systematic Survey and Policy Framework Recommendations.
指南 / 共识 · Healthcare (Basel) · 2026-07-02

本研究采用PRISMA-ScR方法系统分析了156篇2018至2025年间的文献、政策文件与专业指南,整合伦理风险领域与AI全生命周期治理措施。研究提出以公平与偏差缓解、可解释性与透明性、问责与监督、隐私保护基础设施、适应性监管协调为五大支柱的伦理治理框架。

要点

  • 影像AI治理需覆盖数据采集至持续学习的全生命周期,而非仅关注技术性能指标。
  • 阅片中使用AI辅助诊断时,需关注模型校准度与临床医师诊断一致性。
  • AI决策缺乏透明性与可解释性时,仍应保留医师对诊断结论的复核义务。

核心结论

  • 提出包含公平与偏差缓解、可解释性与透明性、问责与监督、隐私保护基础设施、适应性监管协调五大支柱的影像AI伦理治理框架。
  • 信任量化是临床AI治理中尚未充分发展但必需的维度,需整合模型校准度、临床医师与AI一致性及患者接受度等可测量指标。
  • 该框架虽以澳大利亚医疗系统为例,但设计上适用于联邦制和多辖区卫生体系。

研究设计与样本

  • 系统综述遵循PRISMA-ScR方法,纳入156篇2018至2025年间同行评议文献、监管文件、政策报告与专业指导材料。
  • 通过主题分析与生命周期映射方法进行综合,覆盖数据采集、模型开发、部署、监测及持续学习各阶段。

诊断标准与关键数值

  • 伦理治理核心领域确定为:公平与偏差缓解、可解释性与透明性、问责与监督、隐私保护基础设施、适应性监管协调。
  • 信任量化指标应包括模型校准度、临床医师与AI诊断一致性及患者接受度。

临床应用与阅片要点

  • 在部署AI辅助诊断工具前,应评估其训练数据与目标人群的适配性,防范系统性偏差。
  • 影像报告环节使用AI结果时,应记录AI输出与医师最终结论不一致的情形。
  • 科室引入AI系统时需建立持续监测机制,覆盖部署后的性能漂移与不良事件。

局限性

  • 该框架以澳大利亚医疗系统为主要示例,其他司法辖区的适用性尚未经实证验证。
  • 信任量化维度的具体测量方法当前尚缺乏成熟标准。

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

Background/Objectives: Artificial intelligence (AI) is increasingly embedded within diagnostic imaging workflows, reshaping clinical decision-making, health system governance, and regulatory oversight. While technical advances in radiological AI have accelerated, governance mechanisms have struggled to keep pace with issues of bias, transparency, accountability, and lifecycle oversight. This study examines ethical, regulatory, and implementation challenges in AI-enabled diagnostic imaging, building on prior reviews that have often emphasised technical performance by integrating ethical risk domains with governance responses across the AI lifecycle. Methods: This study presents a PRISMA-ScR-informed systematic survey of 156 sources, including peer-reviewed publications, regulatory documents, policy reports, and professional guidance materials (2018-2025), synthesised through thematic analysis and lifecycle mapping spanning data acquisition, model development, deployment, monitoring, and continuous learning. Results: Drawing on both thematic insights derived from the reviewed literature and established ethical and regulatory frameworks, we propose a literature-derived conceptual ethical-governance framework organised around five pillars: equity and bias mitigation, explainability and transparency, accountability and oversight, privacy-preserving infrastructure, and adaptive regulatory alignment. Although illustrated through the Australian healthcare context, the framework is designed to be transferable to federated and multi-jurisdictional health systems. This review further identifies trust quantification as an underdeveloped but essential dimension of clinical AI governance, emphasising the need to integrate measurable indicators such as calibration, clinician-AI concordance, and patient acceptance into lifecycle-based evaluation. Conclusions: By bridging technical, ethical, and policy perspectives, this review proposes a structured conceptual governance framework to support safe, equitable, and trustworthy AI integration in digital health systems.

来源:Healthcare (Basel),查看原文。版权归原作者所有。

引用

  1. Athukorala, Ahmed, Nowrozy, et al. Ethical and Governance Challenges of AI in Medical Imaging and Diagnostics: A Systematic Survey and Policy Framework Recommendations. Healthcare (Basel). 2026. PMID: 42450985. doi:10.3390/healthcare14131975

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。