Differentiation of benign and malignant parotid gland tumors using T2-weighted MRI-based radiomics: the role of feature selection and dimensionality reduction.

OBJECTIVE: To identify the optimal machine learning strategy, including feature selection and dimensionality reduction methods, for differentiating benign and malignant parotid gland tumors using radiomic features from T2-weighted magnetic resonance images (MRI). MATERIAL AND METHODS: This retrospective study included 95 patients with histopathologically confirmed parotid gland tumors (49 benign, 46 malignant). A total of 107 radiomic features (morphological, first-order, and texture features) were extracted using PyRadiomics. Seven feature selection and dimensionality reduction methods were evaluated in combination with five classifiers. Model performance was assessed using five-fold cross-validation, all data preprocessing and feature selection steps were entirely embedded within the cross-validation framework for data leakage. Performance metrics included accuracy, precision, recall, F1-score, specificity, and area under the receiver operating characteristic curve (AUC). RESULTS: The full feature set showed limited performance (AUC: 0.65). PCA, PCR, and mRMR did not significantly improve results, whereas recursive feature elimination (RFE) consistently identified more informative features. The RFE-SVM model achieved the best performance with an AUC of 0.85. CONCLUSION: RFE combined with SVM improves the differentiation of benign and malignant parotid gland tumors on T2-weighted MRI. These findings highlight the importance of model-based feature selection for developing clinically applicable radiomic models.