Interpretable Deep Learning Radiomics for Differentiating Pleomorphic Adenoma and Warthin Tumor.
Authors
Affiliations (5)
Affiliations (5)
- The Sixth School of Clinical Medicine, the Affiliated Qingyuan Hospital (Qingyuan People's Hospital), Guangzhou Medical University, Qingyuan, P.R. China.
- Department of Neurology, Clinical Neuroscience Center, The 7th Affiliated Hospital, Sun Yat-Sen University, Shenzhen, P.R. China.
- Department of Nuclear Medicine, the Affiliated Qingyuan Hospital (Qingyuan People's Hospital), Guangzhou Medical University, Qingyuan, P.R. China [email protected].
- The Sixth School of Clinical Medicine, the Affiliated Qingyuan Hospital (Qingyuan People's Hospital), Guangzhou Medical University, Qingyuan, P.R. China; [email protected].
- Department of Nuclear Medicine, the Affiliated Qingyuan Hospital (Qingyuan People's Hospital), Guangzhou Medical University, Qingyuan, P.R. China.
Abstract
Preoperative differentiation between pleomorphic adenoma (PA) and Warthin tumor (WT) is essential for optimizing surgical strategies. This study aimed to develop and validate an interpretable machine learning framework integrating clinical data, traditional radiomics, and deep learning features to enhance diagnostic precision <i>via</i> conventional computed tomography (CT). We retrospectively analyzed 171 patients (84 PA and 87 WT). A total of 1,561 radiomic and 2,048 deep learning features were extracted from preoperative CT scans. Following LASSO-based feature selection, models were constructed and evaluated on a 70:30 training-validation split (n=119 and n=52, respectively). Diagnostic performance was quantified using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CI). Model interpretability was achieved through SHapley Additive exPlanations (SHAP) analysis. The combined model significantly outperformed individual clinical or radiomic models. The XGBoost classifier exhibited the highest efficacy, yielding a validation AUC of 0.961 (95%CI=0.916-1.000). In the training cohort, the clinical model and the radiomics-deep learning model achieved AUCs of 0.902 (95%CI=0.843-0.962) and 0.95 (95%CI=0.91-0.99), respectively. SHAP analysis indicated that the deep-learning score, patient age, and gender were the most influential predictors. An interpretable model combining deep learning radiomics and clinical attributes provides robust accuracy in distinguishing PA from WT. The integration of SHAP values offers clinicians transparent insights, supporting personalized treatment planning.