Computed tomography radiomics-derived nomogram for prediciting early renal function decline after partial nephrectomy in renal cell carcinoma: a multicenter development/validation study.

backgroundTo develop and externally validate CT-based radiomics-clinical nomograms integrating dual-region radiomics features and preoperative-intraoperative clinical factors for predicting early renal function decline in patients with localized renal cell carcinoma (RCC) undergoing partial nephrectomy (PN).

methodsThis multicenter retrospective study included 1440 patients with localized RCC who underwent PN and preoperative contrast-enhanced CT. Radiomics features were extracted from both the tumor and ipsilateral normal renal parenchyma across corticomedullary, nephrographic, and excretory phases. After reproducibility filtering, high-correlation removal, and least absolute shrinkage and selection operator (LASSO) regression, three radiomics signatures were constructed. Independent clinical predictors identified via multivariate logistic regression were combined with radiomics signatures to develop a preoperative nomogram and a preoperative-intraoperative nomogram. Model performance was assessed using ROC analysis, AUC, calibration curves, Hosmer-Lemeshow tests, decision curve analysis (DCA), and 1000-iteration bootstrap validation. External validation was conducted across 10 independent medical centers and an online public dataset (KITS23).

resultsTwelve tumor-derived and 8 kidney-derived radiomics features were selected, and the combined radiomics signature demonstrated superior predictive ability across cohorts. Independent predictors included age, diabetes, preoperative eGFR, RENAL score, ischemia time, and the radiomics signature. The preoperative nomogram showed excellent discrimination (AUC = 0.952, 0.909, 0.931, 0.914, 0.926), robust calibration, and favorable DCA performance. Bootstrap analysis confirmed its internal stability (mean AUC = 0.948; 95% CI: 0.927-0.972).The preoperative-intraoperative nomogram further improved discriminative performance (AUC = 0.962, 0.926, 0.947, 0.947, 0.933), with strong calibration and consistent clinical utility across all cohorts. Bootstrap validation also demonstrated excellent internal robustness (mean AUC = 0.960; 95% CI: 0.947-0.990).

conclusionsRadiomics signatures derived from CT images of both tumor and normal renal tissue, when integrated with clinical parameters, can accurately predict early renal function decline following PN. The proposed nomograms may facilitate individualized risk stratification and optimize surgical decision-making.