Preoperative prediction of preserved renal parenchymal volume via multilevel CT feature fusion: a proof-of-concept study.

PURPOSE: Accurate preoperative prediction of preserved renal parenchymal volume (RPV) following partial nephrectomy (PN) is critical for individualized postoperative management. However, current assessment approaches remain limited in precision and generalizability. This study aimed to develop and validate a CT-based multilevel feature model for accurate preoperative prediction of postoperative preserved RPV after PN. METHODS: In this retrospective study, 185 patients who underwent PN at Sun Yat-sen University Cancer Center between 2019 and 2023 were included. A nnUNetV2-based segmentation model was used to automatically delineate renal parenchyma and tumors, generating reference postoperative RPVs. Regions of interest were constructed, and three hierarchical feature sets were extracted: (1) radiomic features using PyRadiomics after variance and correlation filtering, (2) handcrafted features derived from R.E.N.A.L. nephrometry score-weighted image dilation, and (3) deep learning (DL) features obtained from a dedicated network with principal component analysis (PCA) dimensionality reduction. Ten machine and deep learning regressors were trained using five-fold cross-validation and compared against clinician-derived RPV estimates. Feature importance was evaluated using Shapley additive explanations (SHAP). RESULTS: The segmentation model achieved a mean Dice similarity coefficient of 0.93, indicating high delineation accuracy. Among all regressors, the TabPFN model yielded the best predictive performance. Models using radiomic features alone achieved an R² of 0.873 and a mean absolute percentage error (MAPE) of 6.8%; performance improved with the inclusion of handcrafted features (R²=0.883, MAPE = 6.7%) and further with DL features (R² = 0.900, MAPE = 6.4%), significantly outperforming manual clinical estimates (R²=0.780, MAPE = 9.7%). SHAP analysis demonstrated that all three feature levels contributed substantially to the overall prediction accuracy. CONCLUSION: The proposed CT-based multilevel feature model enables highly accurate preoperative prediction of postoperative preserved RPV following PN, significantly outperforming conventional clinical assessments. This model offers a robust, interpretable framework to support precision surgical planning and personalized renal function preservation strategies.