Abstract Clear cell renal cell carcinoma (ccRCC) prognosis is guided by a tumor’s pathological features. While radiomics shows promise for non-invasive tumor characterization, it remains unclear whether incorporating the peritumoral region and leveraging multiphase CT imaging improves model performance beyond standard tumor-region analysis. This study aimed to determine whether radiomic features derived from the peritumoral area in standard four-phase CT protocol improve the prediction of ccRCC stage, grade, and aggressiveness compared to intratumoral features alone. A retrospective cohort of 250 patients with ccRCC was analyzed. Using tumor and 5-mm peritumoral regions on four-phase CT, 1874 radiomic features were extracted per region per phase. Elastic Net and Random Forest models were trained with tenfold cross-validation to predict binarized stage, grade, and aggressiveness outcomes using three feature sets: tumor region, peritumoral region, and a combined region including both the tumor and peritumoral regions. Models using only tumor-region features consistently outperformed those using the peritumoral and combined regions. The best tumor-region model achieved area under the receiver operating characteristic curve (AUC) values of 0.875 for stage, 0.697 for grade, and 0.915 for aggressiveness. Multiphase analysis provided no benefit over a single-phase approach, with the nephrographic phase alone yielding equivalent or superior performance to multiphase models across all outcomes. A simplified radiomic approach using features from the tumor region in the nephrographic phase alone provides optimal performance for predicting key ccRCC characteristics. The added complexity of peritumoral and multiphase analysis did not enhance predictive accuracy, potentially streamlining future ccRCC radiomic research.
Muellner et al. (Mon,) studied this question.