An explainable radiomics-based machine learning model for preoperative differentiation of parathyroid carcinoma and atypical tumors on ultrasound: a retrospective diagnostic study.

Background: Parathyroid carcinoma (PC) and atypical parathyroid tumors (APT), constituting rare endocrine malignancies, demonstrate overlapping clinical-radiological presentations with benign adenomas. This study aimed to investigate the predictive performance of three radiomics-based machine learning models for the identification of PC/APT from solitary parathyroid lesions using ultrasound.

Methods: This retrospective diagnostic study analyzed 913 surgically-confirmed parathyroid neoplasms (mean age 54.2 ± 13.7 years; 694 females, 219 male) from Nanjing Drum Tower Hospital (n = 730) and Jinling Hospital (n = 183). The cohort comprised 90 malignant lesions and 823 benign adenomas, divided into training (Hospital I) and external test cohort (Hospital II). A radiomic signature derived from 544 quantitative ultrasound features was developed using three machine learning classifiers: Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR). The performance of the predictive models was evaluated based on the pathological diagnosis.

Results: The RF-based radiomics model showed excellent diagnostic performance. The AUC of this model (0.933) was higher than that of SVM (0.900, < 0.05) and LR (0.901, < 0.05). The accuracy, precision, recall, and F1-score of RF model in distinguishing PA from APT/PC were 0.940, 0.683, 0.638 and 0.660. The explainable bar chart, heatmap and Shapley Additive exPlanations (SHAP) values were used to explain and visualize the main predictors of the optimal model.

Conclusion: This radiomics framework provides a promising tool to support doctors in the clinical management of parathyroid lesions.