Department of Radiology, The Royal Melbourne Hospital, Melbourne, Victoria, Australia.
J Med Imaging Radiat Oncol. 2024 Sep;68(6):659-666. doi: 10.1111/1754-9485.13734. Epub 2024 Aug 9.
Early-stage lung cancer diagnosis through detection of nodules on computed tomography (CT) remains integral to patient survivorship, promoting national screening programmes and diagnostic tools using artificial intelligence (AI) convolutional neural networks (CNN); the software of AI-Rad Companion™ (AIRC), capable of self-optimising feature recognition. This study aims to demonstrate the practical value of AI-based lung nodule detection in a clinical setting; a limited body of research.
One hundred and eighty-three non-contrast CT chest studies from a single centre were assessed for AIRC software analysis. Prospectively collected data from AIRC detection and characterisation of lung nodules (size: ≥3 mm) were assessed against the reference standard; reported findings of a blinded consultant radiologist.
One hundred and sixty-seven CT chest studies were included; 52% indicated for nodule or lung cancer surveillance. Of 289 lung nodules, 219 (75.8%) nodules (mean size: 10.1 mm) were detected by both modalities, 28 (9.7%) were detected by AIRC alone and 42 (14.5%) by radiologist alone. Solid nodules missed by AIRC were larger than those missed by radiologist (11.5 mm vs 4.7 mm, P < 0.001). AIRC software sensitivity was 87.3%, with significant false positive and negative rates demonstrating 12.5% specificity (PPV 0.6, NPV 0.4).
In a population of high nodule prevalence, AIRC lung nodule detection software demonstrates sensitivity comparable to that of consultant radiologist. The clinical significance of larger sized nodules missed by AIRC software presents a barrier to current integration in practice. We consider this research highly relevant in providing focus for ongoing software development, potentiating the future success of AI-based tools within diagnostic radiology.
通过计算机断层扫描（CT）检测结节进行早期肺癌诊断仍然是患者生存的关键，这促进了使用人工智能（AI）卷积神经网络（CNN）的国家筛查计划和诊断工具；该软件是 AI-Rad Companion™（AIRC），能够自我优化特征识别。本研究旨在展示基于人工智能的肺结节检测在临床环境中的实际价值；这是一个有限的研究领域。
对来自单一中心的 183 例非对比 CT 胸部研究进行了 AIRC 软件分析评估。前瞻性收集了 AIRC 检测和肺结节特征（大小：≥3mm）的数据，并与参考标准（报告的盲法顾问放射科医生的发现）进行了评估。
共纳入 167 例 CT 胸部研究；52%的研究为结节或肺癌监测。在 289 个肺结节中，两种方法均检测到 219 个（75.8%）结节（平均大小：10.1mm），AIRC 单独检测到 28 个（9.7%），放射科医生单独检测到 42 个（14.5%）。AIRC 漏诊的实性结节比放射科医生漏诊的结节大（11.5mm 比 4.7mm，P<0.001）。AIRC 软件的灵敏度为 87.3%，特异性有显著的假阳性和假阴性率，为 12.5%（PPV 0.6，NPV 0.4）。
在结节患病率较高的人群中，AIRC 肺结节检测软件的灵敏度与顾问放射科医生相当。AIRC 软件漏诊的较大结节的临床意义是当前实践中整合的一个障碍。我们认为这项研究非常重要，为正在进行的软件开发提供了重点，为人工智能工具在诊断放射学中的未来成功奠定了基础。
