Objective: Chronic kidney disease (CKD) is a significant concern following renal tumor surgery, impacting long-term renal function and patient outcomes. This study investigates the potential of CT-based radiomics as a quantitative imaging approach to predict postoperative CKD in kidney tumor patients. Methods: We included adult patients with renal tumor surgery treated at our center between 2012 and 2022. Preoperative retrospective CT-imaging data were analyzed and radiomic features were extracted from tumor lesions and renal parenchyma. Machine learning models were trained to predict postoperative new-onset CKD based on clinical information and radiomics. Model performance was assessed using five-fold cross-validation on training-set (n=65) and on a separate test-set (n=17). Model performance was primarily evaluated using the receiver operating characteristic (ROC) curve, with the area under the curve (AUC) serving as the principal summary metric. Results: The study cohort comprised n=82 patients of which n=25; 30% developed postoperative new-onset CKD. Best models achieved a mean validation AUC of 0.74 95% CI 0.60-0.86 for solely radiomics, 0.83 0.73-0.93 with clinical information only, and 0.80 0.67-0.91 on radiomics and clinical parameters, respectively (p > 0.05). For the test dataset, AUCs were 0.62 95% CI 0.29-0.92, 0.77 0.50-0.98, and 0.80 0.52-1.00, respectively (p > 0.05). Conclusion: Preoperative CT-based radiomic features in combination with clinical information can serve as a non-invasive predictor of postoperative CKD in renal tumor patients undergoing surgical resection. While prospective and external validation is needed, this approach facilitated clinical decision-making and enables personalized treatment strategies in patients with renal tumors.
Holzschuh et al. (Mon,) studied this question.