A preoperative clinical–radiomics nomogram based on contrast-enhanced CT for differentiating thymic epithelial tumors from thymic cysts: a retrospective study

Background Thymic epithelial tumors (TETs) and thymic cysts are the most common anterior mediastinal lesions but differ markedly in biological behavior and treatment strategies. Misdiagnosis of thymic cysts often leads to non-therapeutic thymectomy. Accurate preoperative and non-invasive differentiation is therefore essential. Previous radiomics studies mainly relied on non-contrast CT, whereas contrast-enhanced CT (CECT) provides richer information on tissue contrast and vascularity. This study aimed to develop and validate a CECT-based clinical–radiomics nomogram to improve diagnostic accuracy and reduce unnecessary surgery. Methods We retrospectively enrolled 306 patients with pathologically confirmed TETs (n = 168) or thymic cysts (n = 138) who underwent CECT between June 2018 and June 2024. Patients were randomly divided into a training cohort (n = 214) and a validation cohort (n = 92). Radiomics features were extracted from manually segmented arterial-phase ROIs using PyRadiomics. After reproducibility testing and correlation filtering, least absolute shrinkage and selection operator (LASSO) regression was applied to select features and calculate the radiomics score (Rad-score). Clinical and semantic features were identified using logistic regression. Ten machine learning algorithms were compared. The combined model was constructed by integrating the multilayer perceptron (MLP)-derived Rad-score with independent clinical features into logistic regression. Model performance was assessed by receiver operating characteristic (ROC) analysis, calibration curves, decision curve analysis (DCA), and the DeLong test. SHapley Additive exPlanations (SHAP) were used to evaluate feature contributions. Results The final nomogram incorporated the Rad-score and five independent features: lesion length, width, area, density homogeneity, and myasthenia gravis. The combined model achieved the best overall performance, with an AUC of 0.988 (95% CI: 0.979–0.998) in the training cohort and 0.914 (95% CI: 0.856–0.972) in the validation cohort. Accuracy, sensitivity, and specificity were 94.4%, 92.1%, and 97.0% in the training cohort, and 83.7%, 77.8%, and 92.1% in the validation cohort, respectively. Calibration showed good agreement between predicted and observed outcomes, and DCA confirmed greater net clinical benefit of the combined model. SHAP analysis indicated that the Rad-score, lesion size parameters, and density homogeneity were the most influential predictors, consistent with radiological knowledge. Conclusions The CECT-based clinical–radiomics nomogram provides an accurate, interpretable, and non-invasive tool for preoperative differentiation between TETs and thymic cysts. By incorporating SHAP analysis, the model enhances transparency and clinical applicability, with potential to reduce unnecessary thymectomy and support individualized surgical decision-making.