Background: This study investigated a radiomics and machine learning methodology for differentiating chromophobe renal cell carcinoma (ChRCC) from renal oncocytoma (RO) using nephrographic phase computed tomography (CT) im-ages, aiming to enhance preoperative diagnostic strategies.Methods: A retrospective cohort of 66 patients (36 ChRCC, 30 RO) undergoing contrast-enhanced CT was analyzed. Man-ual 3D segmentation of renal masses on nephrographic phase images yielded 107 radiomics features. LASSO regression identified the most discriminative features for dimensionality reduction. Four machine learning algorithms (RFC, SVM, Decision Tree, XGBoost) were trained and validated using a 70:30 data split and 10-fold cross-validation. Diagnostic per-formance was quantified by sensitivity, specificity, and AUC.Results: LASSO regression identified 10 pivotal radiomics parameters, including first- and second-order features (e.g., GLCM, GLRLM), reflecting subtle architectural differences. XGBoost showed superior diagnostic performance (AUC: 0.979, 95% CI: 0.967–0.991, sensitivity: 89.75%, specificity: 94.55%). SVM achieved an AUC of 0.939 (95% CI: 0.909–0.968, sensitivity: 91.5%, specificity: 89.25%). Decision Tree (AUC: 0.906, sensitivity: 92.44%, specificity: 91.75%) and RFC (AUC: 0.90, sensitivity: 91.3%, specificity: 88.2%) also performed well. No significant age or gender differences were noted be-tween cohorts (p 0.05).Conclusion: Integrating CT-based radiomics with machine learning, particularly XGBoost, offers a highly accurate, non-invasive paradigm for preoperative ChRCC and RO differentiation. This approach holds substantial promise for optimizing clinical decision-making, supporting nephron-sparing interventions and potentially reducing overtreatment of benign renal oncocytomas.
Mendi et al. (Mon,) studied this question.