Interpretable machine learning based on intratumoral and peritumoral ultrasound radiomics for predicting central lymph node metastasis in papillary thyroid carcinoma.
Authors
Affiliations (3)
Affiliations (3)
- Department of Medical Ultrasound, West China Hospital, Sichuan University, Chengdu, Sichuan, China.
- Department of Medical Ultrasound, West China Hospital, Sichuan University, Chengdu, Sichuan, China. Electronic address: [email protected].
- Department of Medical Ultrasound, West China Hospital, Sichuan University, Chengdu, Sichuan, China. Electronic address: [email protected].
Abstract
This retrospective and single-center study aimed to develop machine learning (ML) model integrating clinical features, ultrasound (US) features, and radiomics signatures extracted from both intratumoral and peritumoral regions to predict central lymph node metastasis (CLNM) in papillary thyroid carcinoma (PTC). The SHapley Additive exPlanations (SHAP) method was applied to visualize the prediction process and enhance clinical interpretability. A total of 879 patients with PTC who underwent preoperative US examination between January 2023 and January 2024 were retrospectively analyzed. Patients were randomly divided into training (n = 615) and test (n = 264) sets. Radiomics signatures were extracted from intratumoral regions and peritumoral regions extending 3 mm and 5 mm beyond the tumor margin. After feature selection, Radscore were computed. Five ML models incorporating clinical features, US features and Radscore were developed. Model performance was evaluated using receiver operating characteristic (ROC) curves, decision curve analysis (DCA), calibration curves, and the Hosmer-Lemeshow test. SHAP was used to explain ML model predictions. CLNM occurred in approximately 52% of PTC. Patients with CLNM were younger and more often male (p < 0.001). Multifocal tumors, extrathyroidal extension, and suspicious lymph nodes on US were also associated with higher CLNM risk (p < 0.05). The Radscore derived from intratumoral and peritumoral regions were significantly different between patients with and without CLNM (p < 0.05). Combined ML models outperformed those based on clinical and US features (p < 0.05). The best performing model (XGB) achieved an AUC of 0.868 (sensitivity = 0.777, specificity = 0.803 and accuracy = 0.749) in the training set and an AUC of 0.787 (sensitivity = 0.704, specificity = 0.695 and accuracy = 0.713) in the test set. The XGB model demonstrated superior clinical utility and well-calibrated for CLNM prediction. SHAP analysis identified the Radscore from the combination of intratumoral and 3-mm peritumoral regions as the most CLNM predictor and provided patient-level interpretability. Intratumoral and peritumoral radiomics features based on US show potential for predicting CLNM in PTC. The integration of SHAP analysis enhances model transparency and may support individualized treatment decision-making.