Prediction study of surrounding tissue invasion in clear cell renal cell carcinoma based on multi-phase enhanced CT radiomics.

Objective: To examine the effectiveness of a nomogram model that combines clinical-image features and CT radiomics in predicting surrounding tissue invasion (STI) in clear cell renal cell carcinoma (ccRCC) patients before surgery.

Methods: Postoperative pathological data of 248 ccRCC patients from two centers were retrospectively collected. Univariate and multivariate regression analyses were used to identify clinical and image features of ccRCC patients to construct a clinical model. Radiomics features were extracted from three CT scans, including tumoral, intratumor, and peritumoral regions. A nomogram was developed by integrating clinical model with optimal radiomics signature. The Shapley Additive Explanations (SHAP) method was used for interpretation.

Results: This study included 65 ccRCC patients with STI and 183 patients without STI. The AUC of the clinical model was 0.766, 0.765, and 0.698 in the training cohort, internal validation cohort, and external validation cohort, respectively. The AUCs were higher in the radiomics signature based on ROI4 in NP than other radiomics (training cohort: 0.837 vs. 0.775-0.847; internal validation cohort: 0.831 vs. 0.695-0.811; external validation cohort: 0.762 vs. 0.623-0.731). Integrating the optimal radiomics signature with the clinical model to construct a combined model resulted in an AUC of 0.890, 0.886, and 0.826 in the training cohort, internal validation cohort, external validation cohort, respectively. SHAP values analysis revealed the top three radiomics features to be Small Dependence Low Gray Level Emphasis, Maximum 3D Diameter, and Maximum Probability.

Conclusion: A nomogram based on preoperative CT and clinical image features is a reliable tool for predicting STI in ccRCC patients. The use of SHAP values can help popularize this tool.