A comprehensive model with dual plane ultrasound based deep learning radiomics to distinguish parotid pleomorphic adenoma from Warthin tumor
Yi Wang, An Dai, Zhaolin Yin, Jiening Gao, Meng Sun, Yufei Xiao, Jingyu Chen, Ruoling Han
Objective To develop and validate a dual-plane ultrasound (US)-based deep learning radiomics model for the noninvasive discrimination between parotid pleomorphic adenoma (PA) and Warthin tumor (WT). Methods This retrospective two-center study enrolled 656 patients with pathologically confirmed parotid tumors (361 PA, 295 WT). Cases from Center 1 (n = 556) were randomly split into training and internal validation sets (7:3), while Center 2 cases (n = 100) formed the external test set. Radiomics features were extracted from long-axis and short-axis plane US images respectively to construct dual-plane radiomics models. For deep learning, pre-trained ResNet/DenseNet was adopted, and two fusion strategies were compared: multi-channel image fusion (long-axis, short-axis, and their pixel-wise mean) versus feature fusion (concatenation of deep features extracted from each plane). A comprehensive prediction model (nomogram) was further constructed by integrating clinical information and US features. Performance was evaluated using AUC, calibration curves, and DCA, and compared against three US physicians of varying experience. Diagnostic performance with and without model assistance was also assessed. Results The multi-channel fusion deep learning model (AUC = 0.860, 95%CI: 0.790-0.931) outperformed feature fusion (AUC = 0.818, 95%CI: 0.735-0.904). Nomogram achieved the highest AUC: 0.907(95%CI: 0.850-0.964) in testing set, outperforming individual clinical, radiomics and deep learning models. It also exceeded the diagnostic performance of all physicians. Model assistance improved junior physicians’ AUC by 0.062 and specificity by 0.220; senior and expert physicians showed higher sensitivity and accuracy. Conclusions The proposed dual-plane US-based deep learning radiomics model effectively differentiates PA from WT noninvasively, enhances diagnostic performance—particularly for less experienced physicians—and demonstrates potential for clinical translation.