Decision fusion model for predicting WHO/ISUP grade of clear cell renal cell carcinoma based on multi-phase CT radiomics: a multicenter study

The pathological grading of clear cell renal cell carcinoma (ccRCC) plays a critical role in both its diagnosis and treatment. This study aimed to develop a decision fusion model based on multi-phase CT to predict the WHO/ISUP grade of ccRCC. This retrospective study included clinical and radiological data from patients with ccRCC from three centers. Radiomics features from both the arterial phase and venous phase of CT images were extracted. After feature dimensionality reduction and selection, optimized features from the arterial phase, venous phase, and their combination were input into six machine learning classifiers to construct 18 radiomics models. The receiver operating characteristic (ROC) curve analysis was used to select the optimal radiomics model. Univariate and multivariate logistic regression was applied to select clinical-radiological features and establish a clinical model. A combined model then was developed using stacking ensemble learning to integrate radiomics and clinical model. The performance of the models was evaluated using ROC curves, and the clinical utility of the models was assessed using decision curve analysis (DCA). A total of 884 ccRCC patients were ultimately included in the study (training set: n = 459; internal validation set: n = 198; external validation set: n = 227). The ExtraTrees model from the CTA_CTV group demonstrated the best performance, achieving an AUC of 0.691 in the external validation set. Subsequently, ExtraTrees classifier was used to construct the clinical model. By integrating the clinical and radiomics models using LightGBM, the combined model exhibited superior performance in the external validation cohort, with an AUC value of 0.703 (95% CI: 0.628–0.778). The combined model outperformed the clinical and radiomics models. DCA revealed that the combined model provided the highest clinical net benefit for pathological grading prediction. The combined model showed potential for preoperative WHO/ISUP grading of ccRCC and may provide complementary non-invasive information for risk stratification.