Fusion of 2.5D deep transfer learning and radiomics for predicting benign and malignant Lung Imaging Reporting and Data System (Lung-RADS) 3 and 4A nodules.

Background: Pulmonary nodules categorized as Lung Imaging Reporting and Data System (Lung-RADS) 3 and 4A constitute a substantial proportion of radiologically indeterminate lesions, and current imaging criteria demonstrate limited accuracy for distinguishing benign from malignant nodules. To develop and validate an integrated predictive model combining radiomics and two-and-a-half (2.5D) deep transfer learning (DTL) for differentiating benign from malignant pulmonary nodules classified as Lung-RADS categories 3 and 4A on computed tomography (CT). Methods: This retrospective study included 298 patients with Lung-RADS 3 and 4A nodules from three centers. The cohort from Center 1 (n=247) was divided into training (n=172) and test (n=75) sets, while patients from Centers 2 and 3 (n=51) formed an independent validation set. We constructed three models: a radiomics (Rad) model, a DTL model, and an integrated deep transfer radiomics (DTR) model. Model performance was evaluated using the area under the receiver operating characteristic curve, and clinical utility was assessed using decision curve analysis. Results: The DTR model demonstrated superior performance in the training [area under the curve (AUC): 0.975, 95% confidence interval (CI): 0.9529-0.9981], testing (AUC: 0.851, 95% CI: 0.7328-0.9683), and external validation cohorts (AUC: 0.727, 95% CI: 0.5876-0.8663), outperforming both the Rad model (training: AUC: 0.743; testing: AUC: 0.642; validation: AUC: 0.613) and the DTL model (training: AUC: 0.843; testing: AUC: 0.757; validation: AUC: 0.701). The DTR model exhibited high specificity (0.895) and positive predictive value (0.956) in the test cohort. SHapley Additive exPlanations (SHAP) analysis revealed that the DTR model effectively leveraged complementary features from both Rad and DTL. Conclusions: The integration of Rad and 2.5D DTL significantly improves the diagnostic accuracy for differentiating between benign and malignant Lung-RADS 3 and 4A nodules. This approach provides a robust decision-support tool that could potentially reduce unnecessary interventions for benign nodules while facilitating earlier detection and treatment of malignant lesions.