Radiomics-derived classifier performance evaluation in lung nodule characterization compared with expert radiologists.

Accurate differentiation of benign and malignant lung nodules on CT is essential for patient management. This study compares machine learning–based radiomics models with expert radiologists for this task. Histopathologically confirmed CT cases were retrospectively collected, and nodules were segmented for radiomic feature extraction using PyRadiomics. Feature selection combined LASSO, Random Forest importance, mRMR, and eBoruta, yielding 33 features. Multiple models (Random Forest, XGBoost, LightGBM, CatBoost, SVM, LDA, QDA, MLP) were evaluated using 5-fold stratified cross-validation, with hyperparameter tuning via Optuna. The tuned SVM performed best, achieving an accuracy of 0.886, precision of 0.875, recall of 0.955, F1-score of 0.913, and AUC of 0.941 on the test set. McNemar’s test showed no significant difference between SVM and radiologists (p = 1.000). SHAP analysis provided interpretability of model decisions. Radiomics-based models, particularly SVM, demonstrated performance comparable to radiologists, suggesting potential as supportive tools in lung nodule evaluation, especially in resource-limited settings. However, findings are limited by the small, single-center dataset, and require validation in larger, multicenter cohorts.