Deep-learning-based radiomics of intratumoral and peritumoral CT images to differentiate benign from malignant parotid tumors.

This study aimed to explore the value of radiomics and deep learning features derived from intratumoral and peritumoral regions based on CT images in distinguishing benign from malignant parotid gland tumors (PGTs), and to evaluate whether integrating clinical, radiomics, and deep learning features could improve differential diagnostic efficiency. We retrospectively collected CT imaging and clinical data from 357 patients with confirmed parotid gland tumors (PGTs) after surgery. The dataset was randomly divided into two sets at a ratio of 7:3. Radiomics and deep learning features were extracted from both intratumoral and peritumoral regions. The least absolute shrinkage and selection operator (LASSO) regression method was applied for feature selection. Traditional radiomics (RAD), deep learning (DL), and fusion (RAD+DL) models were constructed using logistic regression (LR). A combined predictive model was further developed by integrating the optimal model with clinical and CT features. Model performance was assessed using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). The fusion model (Tumor+External2_RAD+DL) incorporating both radiomics and deep learning features from intratumoral and peritumoral regions achieved the best performance, with an AUC of 0.881 (95% CI: 0.809-0.943) in the test set. After integrating clinical features (including symptoms, margin, and enlarged lymph nodes), the combined model showed the highest predictive performance, with an AUC of 0.913 (95% CI: 0.842-0.966) and an accuracy of 0.880, significantly outperforming the clinical model alone and models using only intratumoral or peritumoral features from a single region. In conclusion, the fusion model integrating radiomics and deep learning features from intratumoral and peritumoral regions on CT images effectively distinguishes benign from malignant PGTs. The combined model incorporating clinical factors achieves superior predictive performance, demonstrating good potential for clinical application.