Development of a CT-based comprehensive model with deep learning for differentiating pathological types of pulmonary ground-glass nodules.

Background: The lack of reliable clinical features for differentiating benign from malignant pulmonary pure ground-glass nodules (pGGNS) leads to potential misdiagnosis and unnecessary invasive examinations. Although radiomics and deep learning approaches have shown potential in nodule characterization, the diagnostic performance of integrated models combining clinical features, radiomics, and deep learning remains insufficiently defined. This study aimed to develop and validate an integrated model to distinguish benign from malignant pGGNs and to further differentiate pathological subtypes. Materials and methods: This retrospective study included 1,067 patients with pulmonary pGGNs from Shandong First Medical University Cancer Hospital. Clinical and imaging data were collected, and radiomics features and deep learning (DL) derived features were extracted using Python (version 3.7). Patients were randomly divided into training and validation cohorts. Multiple machine-learning classifiers were constructed, and diagnostic performance was assessed using receiver operating characteristic (ROC) curve analysis. Results: For distinguishing benign from malignant pGGNS (Model 1), clinical features such as age, nodule multiplicity, CEA levels, and amylase were identified as clinically relevant features. Thirty-eight valuable features were selected for model development. Among individual classifiers, the Support Vector Machine (SVM) achieved the highest performance with a validation receiver operating characteristic curve (AUC) of 0.840, followed by random forest (0.829), stochastic gradient descent (0.828), k-nearest neighbors (0.814), XGBoost (0.798), and LightGBM (0.818). The integrated model combining clinical features, radiomics, and deep learning achieved a validation set AUC of 0.871. For pathological subtype classification of pGGNs (Model II), clinical features such as gender, Pro-Gastrin-Releasing-Peptide (ProGRP), AST/ALT ratio (De Ritis ratio), creatine kinase-MB (CKMB), and globulin were identified as informative clinical variables. Twelve valuable features were selected The SVM classifier again showed the best individual performance (validation AUC = 0.831), while the integrated model achieved a superior AUC of 0.853. Conclusion: An integrated model incorporating clinical characteristics, radiomics, and deep learning demonstrates robust performance in distinguishing benign from malignant pulmonary pGGNs and in identifying pathological subtypes, suggesting potential clinical utility for non-invasive decision support.