Abstract
To develop and validate two models for convolutional neural networks (CNNs), namely EfficientNet-B0 and Res2Net, constructed from non-contrast CT images for discriminate malignant and benign pulmonary nodules.We recruited 3579 cases of solitary pulmonary nodules, among which 335 cases were benign and 3244 cases were malignant. The EfficientNet-B0 model and Res2Net model were constructed as two-dimensional(2D), and three- dimensional (3D) models, respectively. Furthermore, 4 clinical factors (sex, age, smoking status, and tumor marker) and all clinical factors were combined with Res2Net model to build Res2Net-4F model, and Res2Net-20F model, respectively. The receiver operating characteristic (ROC) curves were utilized to evaluate the diagnostic efficiency and discriminative capability of these models, and ROC curves of these models were compared with Delong test.The diagnostic accuracy of Res2Net, Res2Net-4F, and Res2Net-20F [areas under ROC curves (AUC) = 0.9301, AUC = 0.9811, and AUC = 0.9357, respectively] were higher than that of the EfficientNet B0 (AUC = 0.8801) in the training data set. The results were confirmed by the validation data set (AUC = 0.8282 for the Res2Net; AUC = 0.8299 for Res2Net-4F; AUC = 0.8468 for Res2Net-20F; AUC = 0.7737 for the EfficientNet B0). There was a significant difference between Res2Net model and EfficientNet-B0 model in discriminating malignant and benign pulmonary nodules in both the training data set and validation set (Delong test, both p < 0.05). We developed two novel deep learning models to distinguish malignant and benign pulmonary nodules, and the Res2Net model showed better differentiation accuracy and sensitivity than EfficientNet-B0 model.