Prediction of metastatic risk of renal clear cell carcinoma based on CT radiomics analysis.

Objective: To investigate the value of using imaging histological models to non-invasively assess the risk of metastasis in patients with clear cell renal cell carcinoma (ccRCC). Methods: This study retrospectively enrolled 273 clear cell renal cell carcinoma (ccRCC) patients from three hospitals, with 57 cases allocated as an independent test cohort. High-throughput imaging histomic features (n=2,264) were extracted from triphasic CT (non-enhanced, corticomedullary, and nephrographic phases) using Pyradiomics. Three monophasic radiomics models were developed following dimensionality reduction, with feature contributions quantified via Shapley Additive exPlanations (SHAP) framework to enhance interpretability. A triphasic radiomics model was subsequently established by ensembling phase-specific prediction probabilities. Metastasis risk factors identified through univariate/multivariate logistic regression informed a clinical predictor model. The final combined model integrated triphasic radiomics signatures with clinical parameters, visualized through a nomogram. Diagnostic performance was evaluated via ROC analysis, while calibration curves validated prediction consistency. Results: In this study, SHAP analysis revealed that radiomics features quantifying intratumoral heterogeneity (e.g., necrosis patterns in medullary-phase CT) synergized with clinical factors (tumor size >3 cm, creatinine levels) to drive predictions. Key biological insights included threshold effects of necrosis volume (linked to hypoxia) and tumor diameter (critical threshold: 3 cm), aligning with known metastatic pathways. The clinical model achieved an area under the ROC curve (AUROC) of 0.752 (95% confidence interval [CI]: 0.679-0.826) in the training dataset and 0.681 (95% CI: 0.529-0.833) in the testing dataset. Among the single-phase radiomics models, the CT_Medullary model demonstrated good prediction performance, with an AUROC of 0.785 (95% CI: 0.645-0.924) in the testing dataset. The three-phased CT model exhibited improved diagnostic performance, with a testing AUROC rising to 0.812 (95% CI: 0.680-0.943). Notably, the combined model integrating clinical and radiomics features yielded the best prediction, achieving a further improvement in testing AUROC to 0.824 (95% CI: 0.704-0.944). Conclusion: Radiomics technology provides a quantitative, objective method for predicting the risk of metastasis in patients with ccRCC. Nonetheless, the clinical indicators persist as irreplaceable.