Association between AI-based automated segmentation-derived 3D volumetric imaging features and outcomes after cryoablation for renal cell carcinoma

Abstract Purpose To evaluate the feasibility of artificial intelligence (AI)-based automated segmentation of renal structures for extracting 3D volumetric imaging features and to evaluate their association with outcomes after cryoablation for renal cell carcinoma (RCC). Methods This retrospective study included 116 patients with RCC who underwent cryoablation. Pre-treatment contrast-enhanced CT images were processed using AI-based automated segmentation (TotalSegmentator and an open-source KiTS23 model). 3D volumetric imaging features were extracted. Logistic regression and Cox proportional hazards models were used to assess associations between these features and outcomes. A composite risk score for local tumor control failure was derived from multivariable logistic regression coefficients, and its discriminative performance was assessed using receiver operating characteristic (ROC) curve analysis. Results In multivariable logistic regression, larger tumor-renal sinus contact area was independently associated with local tumor control failure (odds ratio [OR], 1.379; P = 0.002). A composite risk score incorporating tumor-renal sinus contact area and tumor volume achieved an area under the ROC curve of 0.765 for predicting local tumor control failure, which was numerically higher than that of the RENAL nephrometry score (0.664), although the difference was not statistically significant. In Cox analysis, larger tumor-renal sinus contact area (hazard ratio [HR], 1.255; P = 0.026) was significantly associated with local tumor progression following complete ablation. Conclusion AI-based automated segmentation enables extraction of imaging features associated with cryoablation outcomes in RCC. A composite risk score based on tumor volume and tumor-renal sinus contact area showed moderate discrimination for local tumor control failure. This score may help identify anatomically challenging tumors for patient selection and closer surveillance; however, its clinical use remains preliminary. Further validation in larger, prospective cohorts may support its use for risk stratification and treatment planning.