Tugwell-Allsup J R, Owen B W, Hibbs R, England A
Betsi Cadwaladr University Health Board, Radiology Department, Penrhosgarnedd Road, Bangor, Gwynedd, LL57 2PW, UK.
Statistical Support by Integral Business Support Limited, Wrexham, UK.
Radiography (Lond). 2025 Mar;31(2):102876. doi: 10.1016/j.radi.2025.01.011. Epub 2025 Jan 30.
AI software in the form of deep learning-based automatic detection (DLAD) algorithms for chest X-ray (CXR) interpretation have shown success in early detection of lung cancer (LC), however, there remains uncertainty related to clinical validation.
CXRs and their corresponding chest-CT scans were retrospectively collated from a single institution between January 2019-2020. A commercially available AI software was used to evaluate 320 CXRs (<6 years prior-to-diagnosis) from 105 positive LC patients and 103 negative controls. Clinical reports were extracted and coded to correlate against AI findings.
Of 105 LC patients, (57[55 %] men, median [IQR] age 73[68-83] years), clinical reports identified LC in 64 (61 %) whereas AI identified LC in 95 (90 %). AI diagnostic (image-level) and prognostic (patient-level) sensitivities were 57.6 % and 90.0 %, (81 % in correct location), respectively. On CXRs performed >12 months prior to LC diagnosis, the AI detected nodules in 24(23 %) cases of which 22/24 had negative clinical reports for lung nodule/mass. The potential median reduction in time-to-diagnosis for cases where AI identified nodule(s) on previous CXR, but clinical reports negative, was 193[IQR 42-598] days. Of the 103 'negative' controls (48[47 %] men, median [IQR] age 69[61-77] years) 20 patients had a nodule abnormality score above the threshold, generating a false-positive rate of 19 %.
The AI software showed excellent performance in detecting LCs that initially went undetected on CXR. The algorithm has potential to increase LC detection rates and reduce time-to-diagnosis. Using the AI, in conjunction with a trained observer, could increase reporting accuracy and potentially improve clinical outcomes.
This study demonstrated the benefits and pitfalls associated with using AI in a clinical setting. It provides further evidence for utilising decision-support aids within clinical practice.
基于深度学习的自动检测（DLAD）算法形式的人工智能软件在胸部X光（CXR）解读中已成功用于肺癌（LC）的早期检测，然而，临床验证方面仍存在不确定性。
回顾性整理了2019年1月至2020年期间来自单一机构的胸部X光片及其相应的胸部CT扫描。使用一款商用人工智能软件对105例阳性肺癌患者和103例阴性对照的320张胸部X光片（诊断前<6年）进行评估。提取临床报告并编码，以便与人工智能的结果进行关联。
105例肺癌患者中（57例[55%]为男性，年龄中位数[四分位间距]为73岁[68 - 83岁]），临床报告确诊肺癌64例（61%），而人工智能确诊肺癌95例（90%）。人工智能诊断（图像层面）和预后（患者层面）敏感性分别为57.6%和90.0%（正确位置为81%）。在肺癌诊断前>12个月进行的胸部X光片中，人工智能在24例（23%）病例中检测到结节，其中22/24例临床报告的肺结节/肿块为阴性。对于人工智能在先前胸部X光片中识别出结节但临床报告为阴性的病例，潜在的诊断时间中位数减少为193天[四分位间距42 - 598天]。103例“阴性”对照中（48例[47%]为男性，年龄中位数[四分位间距]为69岁[61 - 77岁]），20例患者的结节异常评分高于阈值，假阳性率为19%。
该人工智能软件在检测胸部X光片上最初未被发现的肺癌方面表现出色。该算法有潜力提高肺癌检测率并缩短诊断时间。将人工智能与训练有素的观察者结合使用，可以提高报告准确性并可能改善临床结果。
本研究展示了在临床环境中使用人工智能的益处和陷阱。它为在临床实践中使用决策支持辅助工具提供了进一步的证据。
