Deep transfer learning radiomics combined with explainable machine learning for predicting malignancy risk in parotid gland tumors based on ultrasound.

OBJECTIVE: This study aimed to develop and validate an ultrasound (US)-based deep transfer learning radiomics model, integrated with explainable machine learning, for the preoperative malignant risk prediction of parotid gland tumors (PGTs). METHODS: Data from 1,191 patients were retrospectively collected from three medical centers, and postoperative histopathological examination was used as the reference standard. Radiomics features and deep transfer learning features (ResNet50, Inception_V3, Vgg19) were extracted from the US images. Key predictive variables were selected using principal component analysis (PCA) and the least absolute shrinkage and selection operator (LASSO). Six classifiers-decision tree, gradient boosting machine, k-nearest neighbors, logistic regression, naïve Bayes, and random forest-were employed to construct models based on five feature sets: Clinical model, radiomics (Rad) model, deep transfer learning radiomics (DTLR) model, combined deep transfer learning and radiomics (DTLR-Rad) model, and a comprehensive combined model (CM Clinical + DTLR-Rad). Model performance was evaluated using the area under the curve (AUC). Feature importance was interpreted using SHapley Additive exPlanations (SHAP). A web application for real-time, personalized risk prediction was developed. RESULTS: In external test sets 1 and 2, the CM Clinical + DTLR-Rad model based on the random forest classifier achieved the highest AUCs among the evaluated models, with 0.922 (95% CI: 0.890-0.954) and 0.959 (95% CI: 0.932-0.985), respectively. The integrated model outperformed the clinical-only and single-modality models in both external test sets. SHAP visualizations demonstrated the contribution of individual features. The web application provided both prediction probabilities and feature-level interpretability. CONCLUSION: The CM Clinical + DTLR-Rad model demonstrated good predictive performance. The integration of interpretable machine learning and a web-based application may enhance preoperative risk stratification for PGTs and support clinical decision-making.