Development and Validation of a Risk Stratification Model of Pulmonary Ground-Glass Nodules Based on Complementary Lung-RADS 1.1 and Deep Learning Scores

Abstract

Purpose: To assess the value of novel deep learning (DL) scores combined with complementary lung imaging reporting and data system 1.1 (cLung-RADS 1.1) in managing the risk stratification of ground-glass nodules (GGNs) and therefore improving the efficiency of lung cancer (LC) screening in China. Materials and Methods: Overall, 506 patients with 561 GGNs on routine computed tomography images, obtained between January 2017 and March 2021, were enrolled in this single-center, retrospective Chinese study. Moreover, the cLung-RADS 1.1 was previously validated, and the DL algorithms were based on a multi-stage, three-dimensional DL-based convolutional neural network. Therefore, the DL-based cLung-RADS 1.1 model was created using a combination of the risk scores of DL and category of cLung-RADS 1.1. The recall rate, precision, accuracy, per-class F1 score, weighted average F1 score (F1weighted), Matthews correlation coefficient (MCC), and area under the curve (AUC) were used to evaluate the performance of DL-based cLung-RADS 1.1. Results: The percentage of neoplastic lesions appeared as GGNs in our study was 95.72% (537/561) after long-period follow-up.Compared to cLung-RADS 1.1 model or DL model, The DL-based cLung-RADS 1.1 model achieved the excellent performance with F1 scores of 95.96% and 95.58%, F1weighted values of 97.49 and 96.62%, accuracies of 92.38 and 91.77%, and MCCs of 32.43 and 37.15% in the training and validation tests, respectively. The combined model achieved the best AUCs of 0.753 (0.526–0.980) and 0.734 (0.585–0.884) for the training and validation tests, respectively. Conclusion: The DL-based cLung-RADS 1.1 model shows the best performance in risk stratification management of GGNs, which demonstrates substantial promise for developing a more effective personalized lung neoplasm management paradigm for LC screening in China.