商用AI肺结节CT评估产品的科学证据系统综述
研究按RADAR分级模型分类,早期以技术效能和诊断准确性为主,至2024年诊断思维、治疗效能和患者结局级别研究占比超过三分之一。所有研究至少一个领域存在高偏倚风险,近三分之二涉及厂商资助或合著。
要点
- 早期AI研究聚焦量化与检出,恶性预测占比低,临床解读报告时需注意证据级别差异。
- 2024年恶性预测研究占28.6%,结节特征化占4.5%,提示AI适用范围正在扩展。
- 多数研究存在厂商利益关联,阅片时不宜将AI结果作为独立诊断依据。
核心结论
- 商用AI在肺结节CT评估中的研究数量大幅增长,效能分级从早期技术效能向更高临床结局级别演进。
- 患者结局和成本效益级别研究仍然有限,需补齐以充分实现AI在临床实践和公共卫生中的价值。
研究设计与样本
- 纳入95项研究,对象为CE标记和/或FDA批准的商用AI系统用于CT肺结节评估。
- 采用RADAR效能分级模型对研究进行系统分类,评估不同层级效能随时间的演变。
诊断标准与关键数值
- 2012至2017年共14项研究,其中1级技术效能和2级诊断准确性合计占83.3%。
- 早期研究中46.7%聚焦量化,仅6.7%涉及恶性预测。
- 至2024年,3级诊断思维效能、4级治疗效能和5级患者结局合计占比超过三分之一。
- 至2024年恶性预测研究增至28.6%,结节特征化占4.5%。
- 所有纳入研究至少一个领域存在高偏倚风险,近三分之二涉及厂商资助或合著。
临床应用与阅片要点
- 阅片采用商用AI结节评估结果时应关注其对结节量化、检出、恶性预测的具体任务定位,区分应用场景。
- AI研究证据向诊断思维、治疗效能和患者结局级别扩展,可用于评价对临床决策路径的潜在影响。
- 报告AI结节评估结果时需注意偏倚风险及厂商关联问题,避免单纯依赖AI输出。
局限性
- 所有纳入研究至少一个领域存在高偏倚风险。
- 近三分之二研究涉及厂商资助或合著。
- 患者结局和成本效益级别(5级和6级)研究仍然有限。
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: This systematic review evaluates the available evidence on the efficacy of commercially available AI-software applications for lung nodule assessment on CT scans.
MATERIALS AND METHODS: In adherence to PRISMA guidelines, a thorough search of electronic databases was conducted (January 2012-November 2024) to identify studies on CE-marked and/or FDA-cleared AI-based systems for evaluating pulmonary nodules on CT scans. Articles were systematically categorised using the Radiology AI Deployment and Assessment Rubric (RADAR) to assess the evolution of efficacy over time across various hierarchical levels.
RESULTS: A total of 95 studies were included. Between 2012 and 2024, the number of studies increased from 14 in 2012-2017 to a total of 95 by 2024. The studies were categorised using the RADAR efficacy model. In the early period (2012-2017), most studies focused on lower efficacy levels, with Level-1 (technical efficacy) and Level-2 (diagnostic accuracy) dominating at 83.3%. During this period, AI applications were mainly focused on quantification (46.7%), while malignancy prediction was addressed in only 6.7% of studies. By 2024, Levels-3 (diagnostic thinking efficacy), 4 (therapeutic efficacy), and 5 (patient outcomes) accounted for over a third of all studies. Malignancy prediction further increased to 28.6%, and nodule characterisation emerged in 4.5% of studies. Despite advancements, research on patient outcomes and cost-effectiveness efficacy (Levels-5 and 6) remains limited. Also, all included studies demonstrated high risk of bias in at least one domain, and nearly two-thirds involved vendor funding or co-authorship.
CONCLUSION: In conclusion, the growing interest and investment in AI technologies for thoracic radiology have driven significant advancements in lung nodule assessment on CT-scans. Gaps remain in assessing patient outcomes and societal implications, which must be addressed to fully realise AI's potential in clinical practice and public health.
KEY POINTS: Question What is the current level of scientific evidence supporting commercially available AI tools for pulmonary nodule assessment on CT, and is this sufficient to support their clinical implementation? Findings Research on AI for pulmonary nodule assessment has surged, with the focus evolving from basic technical performance in the early years to higher clinical outcomes by 2024. Clinical relevance This study provides critical insights into the evolving role of AI in lung nodule assessment on CT-scans, highlighting advancements in technology and identifying key gaps in current research. Addressing these is essential for optimising AI's clinical application and improving patient outcomes.
引用
- Paramasamy, Leliveld, Groen, et al. Scientific evidence of commercial artificial intelligence products for pulmonary nodule assessment on CT scans: a systematic review. Eur Radiol. 2026. PMID: 42448899. doi:10.1007/s00330-026-12745-8
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。