Has computer-aided diagnosis now outperformed radiologists in the accurate localization of lung nodules?-a...

Globally, lung cancer is one of the most frequently diagnosed cancers and the leading cause of cancer-related death. Manual radiologist (RAD) reading brings inconsistent localization results for subtle nodules, while existing computer-aided diagnosis (CAD) tools lack verified three-dimensional (3D) positioning accuracy compared with human readers. Thus, the aim of this study was to compare 3D localization accuracy of CAD and RAD evaluations for pulmonary nodules on computed tomography (CT) images.

This multicenter retrospective study included 1,275 patients' chest thin-section CT images from 3 centers. CT images were evaluated via a radiologist (RAD) assessment and a CAD tool. The CAD system was developed using a high-precision convolutional neural network (U-Net) for 3D localization of pulmonary nodules. The gold standard for nodule localization was established based on surgical findings, postoperative histopathological confirmation of the resected specimens, and correlation with preoperative CT images.

In Center 1 (921 pulmonary nodules), CAD and RAD correctly located 856 and 758 pulmonary nodules, with accuracy rates of 92.9% [95% confidence interval (CI), 91.1-94.4%] and 82.3% (95% CI, 79.7-84.6%), respectively. In Center 2 (145 pulmonary nodules), CAD and RAD correctly located 141 and 112 pulmonary nodules, with accuracy rates of 97.2% (95% CI, 94.54-99.90%) and 77.2% (95% CI, 70.39-84.05%), respectively. In Center 3 (209 pulmonary nodules), CAD and RAD correctly located 196 and 163 pulmonary nodules, with accuracy rates of 93.8% (95% CI, 90.53-97.07%) and 78% (95% CI, 72.38-83.62%), respectively. In all three centers, the accuracy rate using CAD was significantly greater than that using RAD (all P<0.001).

Compared with manual detection by radiologists, the use of CAD increased the accuracy rate for 3D localization of pulmonary nodules.