From the Department of Radiology and Research Institute of Radiology, University of Ulsan College of Medicine, Asan Medical Center, 88 Olympic-ro 43 Gil, Songpa-gu, Seoul 138-736, Korea.
Radiology. 2021 Apr;299(1):202-210. doi: 10.1148/radiol.2021202803. Epub 2021 Feb 2.
Background The solid portion size of lung cancer lesions manifesting as subsolid lesions is key in their management, but the automatic measurement of such lesions by means of a deep learning (DL) algorithm needs evaluation. Purpose To evaluate the performance of a commercially available DL algorithm for automatic measurement of the solid portion of surgically proven lung adenocarcinomas manifesting as subsolid lesions. Materials and Methods Surgically proven lung adenocarcinomas manifesting as subsolid lesions on CT images between January 2018 and December 2018 were retrospectively included. Five radiologists independently measured the maximal axial diameter of the solid portion of lesions. The DL algorithm automatically segmented and measured the maximal axial diameter of the solid portion. Reader measurements, software measurements, and invasive component size at pathologic examination were compared by using intraclass correlation coefficient (ICC) and Bland-Altman plots. Results A total of 448 patients (mean age, 63 years ± 10 [standard deviation]; 264 women) with 448 lesions were evaluated (invasive component size, 3-65 mm). The measurement agreements between each radiologist and the DL algorithm were very good (ICC range, 0.82-0.89). When a radiologist was replaced with the DL algorithm, the ICCs ranged from 0.87 to 0.90, with an ICC of 0.90 among five radiologists. The mean difference between the DL algorithm and each radiologist ranged from -3.7 to 1.5 mm. The widest 95% limit of agreement between the DL algorithm and each radiologist (-15.7 to 8.3 mm) was wider than pairwise comparisons of radiologists (-7.7 to 13.0 mm). The agreement between the DL algorithm and invasive component size at pathologic evaluation was good, with an ICC of 0.67. Measurements by the DL algorithm (mean difference, -6.0 mm) and radiologists (mean difference, -7.5 to -2.3 mm) both underestimated invasive component size. Conclusion Automatic measurements of solid portions of lung cancer manifesting as subsolid lesions by the deep learning algorithm were comparable with manual measurements and showed good agreement with invasive component size at pathologic evaluation. © RSNA, 2021
背景 肺癌亚实性病变实性部分的大小在其管理中至关重要，但需要评估深度学习（DL）算法对这些病变进行自动测量的性能。目的 评估一种商用 DL 算法自动测量经手术证实的表现为亚实性病变的肺腺癌实性部分的性能。
材料与方法 回顾性纳入 2018 年 1 月至 2018 年 12 月 CT 图像上表现为亚实性病变的经手术证实的肺腺癌患者。5 名放射科医生独立测量病变实性部分的最大轴向直径。DL 算法自动分割并测量实性部分的最大轴向直径。使用组内相关系数（ICC）和 Bland-Altman 图比较读者测量值、软件测量值和病理检查的侵袭性成分大小。
结果 共评估了 448 例患者（平均年龄，63 岁±10[标准差]；264 例女性）的 448 个病变（侵袭性成分大小，3-65 mm）。每位放射科医生与 DL 算法的测量一致性均非常好（ICC 范围，0.82-0.89）。当用 DL 算法替代一位放射科医生时，ICC 范围为 0.87 至 0.90，5 位放射科医生的 ICC 为 0.90。DL 算法与每位放射科医生的平均差值范围为-3.7 至 1.5 mm。DL 算法与每位放射科医生之间最宽的 95%一致性界限（-15.7 至 8.3 mm）比放射科医生之间的两两比较（-7.7 至 13.0 mm）更宽。DL 算法与病理评估的侵袭性成分大小之间的一致性良好，ICC 为 0.67。DL 算法（平均差值，-6.0 mm）和放射科医生（平均差值，-7.5 至-2.3 mm）的测量值均低估了侵袭性成分大小。
结论 通过深度学习算法自动测量表现为亚实性病变的肺癌实性部分与手动测量值相当，与病理评估的侵袭性成分大小具有良好的一致性。
