An interpretable machine learning approach using nnU-Net-based radiomics for preoperative risk stratification of thymic epithelial tumors: a multicenter study.

OBJECTIVE: This study aimed to develop and validate an interpretable machine learning (ML) model based on nnU-Net automated segmentation and computed tomography (CT) radiomics for preoperative risk stratification in thymic epithelial tumors (TETs). METHODS: In this retrospective multicenter study, 764 patients with pathologically confirmed TETs were enrolled and divided into training, internal validation, and two external validation cohorts. An nnU-Net model was trained for automatic tumor segmentation, with performance assessed by the dice similarity coefficient (DSC). Radiomic features were extracted from the automated segmentations of venous-phase CT images, and least absolute shrinkage and selection operator (LASSO) regression was applied for feature selection. Predictive models, including radiomics-only, clinical-only, and a clinical-radiomics (combined) model, were constructed using five ML algorithms (RF, SVM, KNN, DT, and LR). Model performance was evaluated using the receiver operating characteristic (ROC) curve. Delong’s test was employed to compare these ML models and select the best-performing model as the final model. Calibration curve and decision curve analysis (DCA) were performed to assess clinical efficacy of the final model. The interpretability of the optimal model was elucidated using SHapley Additive exPlanations (SHAP). RESULTS: The nnU-Net segmentation model achieved excellent performance, with a DSC of 0.979 on the test cohort. Compared to the other four combined models, the RF-based combined model demonstrated superior predictive efficacy, yielding area under the curve (AUC) values of 0.941 (training), 0.884 (internal validation), 0.867 (external validation 1), and 0.872 (external validation 2). The calibration curves indicated excellent agreement between the RF-based model’s predictions and actual outcomes, and furthermore, DCA confirmed its superior net benefit over baseline strategies across a wide range of thresholds. SHAP tool identified 11 radiomic features and 3 clinical features as the most influential features, providing transparency into the model’s decision-making process. CONCLUSIONS: The nnU-Net framework enables accurate and efficient automatic segmentation of TETs. The proposed RF-based combined model, integrating clinical and radiomic features, provides a robust and interpretable tool for identifying the high-risk TETs, holding promise for supporting clinical decision-making towards personalized therapy.