肺癌筛查中AI肺结节纵向匹配性能评估
本研究评估了AI肺结节分析系统在英国肺癌筛查试验361例接受3个月随访低剂量CT的参与者中的纵向匹配性能。AI采用NELSON 2.0/EUPS方案的≥100 mm³实性成分阈值自动检测基线结节并进行全自动匹配。结果显示AI对持续存在结节的匹配成功率为83.5%,仅1.5%的持续结节需人工干预。
要点
- AI纵向匹配成功率达83.5%,仅1.5%持续结节需人工纠正。
- 单发结节匹配率91.8%,多于5个结节时降至72.8%。
- 匹配失败者中91.1%为非结节结构,以胸膜斑为主。
核心结论
- 肺AI系统在肺癌筛查纵向结节匹配中表现稳健,可显著减少随访人工追踪工作量。
- 仅1.5%(95% CI: 0.6-3.5%)的持续存在的离散实性结节需人工干预。
研究设计与样本
- 研究类型为对英国肺癌筛查试验数据的AI算法性能评估。
- 纳入361例接受3个月随访低剂量CT扫描的参与者。
- AI采用更新的体积阈值,即实性成分≥100 mm³(NELSON 2.0/EUPS方案)确定需3个月随访的病例。
- 所有AI检测到的基线候选结节(≥100 mm³)均直接进入全自动纵向匹配,无人工筛选环节。
关键影像表现/征象
- AI在181例参与者中检测到378个基线结节(≥100 mm³)。
- 39个结节在随访时自然消退。
- 匹配失败的56个发现中,91.1%(51/56)为非结节结构。
- 匹配失败的非结节结构中,胸膜斑占46.4%(26/56)。
诊断标准与关键数值
- 基线结节纳入阈值为实性成分≥100 mm³(NELSON 2.0/EUPS方案)。
- 339个持续存在结节的匹配成功率为83.5%(283/339;95% CI: 79.2-87.1%)。
- 单发基线候选结节参与者的匹配成功率为91.8%(89/97)。
- 多于5个结节参与者(占队列6.6%)的匹配成功率为72.8%(75/103)。
- 仅5个未匹配的离散实性结节(1.5%;95% CI: 0.6-3.5%)需人工干预。
临床应用与阅片要点
- AI纵向匹配可显著降低肺癌筛查中的人工结节追踪工作量。
- 需注意高结节负荷(>5个结节)时匹配性能明显下降。
- 匹配失败者多为胸膜斑等非结节结构,阅片时需重点鉴别非结节性胸膜改变。
局限性
- 原文指出需在不同人群中进行前瞻性验证。
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
Accurate longitudinal nodule matching is a critical technical prerequisite for automated growth rate (volume doubling time) assessment in lung cancer screening. This study evaluated an artificial intelligence (AI) pulmonary nodule analysis system in all 361 UK Lung Cancer Screening (UKLS) trial participants who underwent a 3-month follow-up low-dose computed tomography (LDCT) scan. The pulmonary AI independently evaluated these baseline scans using an updated volume threshold (solid component ≥ 100 mm³ per NELSON 2.0/EUPS protocol) for cases requiring 3-month follow-up. To assess true algorithmic robustness, all AI-detected baseline candidate nodules (≥ 100 mm³) proceeded to fully automated longitudinal matching without any manual selection. The pulmonary AI identified 181 participants with 378 baseline nodules ≥ 100 mm³. In total, 39 nodules had naturally resolved at follow-up. The pulmonary AI achieved an 83.5% (283/339; 95% CI: 79.2-87.1%) matching success rate for 339 persisting nodules. Matching performance was 91.8% (89/97) for participants with a single baseline candidate nodule (59.7% of the cohort) and 72.8% (75/103) for participants with more than five nodules (6.6%). Expert review of the 56/339 (16.5%) unmatched findings showed that almost all were non-nodular structures (91.1%, 51/56), predominantly pleural plaques (46.4%, 26/56). Consequently, only five unmatched discrete solid nodules (1.5%; 95% CI: 0.6-3.5% of 339 persisting findings) required manual intervention. In conclusion, the pulmonary AI demonstrates robust longitudinal matching performance, with substantial potential for follow-up manual tracking workload reduction. KEY POINTS: Question Does standalone AI longitudinal nodule tracking provide sufficient nodule-level technical reliability to avoid manual review bottlenecks in an automated lung cancer screening workflow? Findings Pulmonary AI matched 83.5% of persisting nodules (91.8% for single nodules, 72.8% for > 5 nodules); 91.1% failures were non-nodular, with only 1.5% requiring manual correction. Clinical relevance Pulmonary AI matching can potentially reduce the manual tracking workload in lung cancer screening, with only 1.5% of persisting nodules requiring manual correction. Performance is reduced in scans with high nodule burden, and prospective validation in diverse populations is needed.
引用
- Jiang, Lancaster, Davies, et al. Automated artificial intelligence performance for longitudinal pulmonary nodule matching in lung cancer screening. Eur Radiol. 2026. PMID: 42668303. doi:10.1007/s00330-026-12825-9
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。