For patients with localized renal cell carcinoma, the choice between partial nephrectomy (removing only the tumor) and radical nephrectomy (removing the entire kidney) has major long-term consequences. Partial nephrectomy preserves kidney function but is more technically complex and carries higher surgical risk for certain tumor configurations.
Current decision-making relies on anatomical scoring systems such as RENAL and PADUA, which rate tumor complexity based on features like location, proximity to the collecting system, and relationship to the renal sinus. However, these scores require manual measurement and involve subjective assessments.
Advanced RCC with complex tumor anatomy is increasingly managed with robot-assisted surgery, but the surgeon still decides preoperatively which procedure is appropriate. An automated system that objectively quantifies relevant anatomical features from CT could standardize and potentially improve this decision.
This study developed a fully automated AI framework that segments kidney anatomy from 3D CT scans and extracts multi-level features to predict which surgical approach (partial or radical nephrectomy) was actually chosen by experienced urological surgeons.
The framework begins with 3D-UNet, a deep learning segmentation architecture specifically designed for volumetric medical images. Applied to 3D CT volumes, 3D-UNet automatically delineates the tumor, kidney parenchyma, collecting system, and vasculature, producing precise anatomical masks without manual annotation during deployment.
From the segmented 3D structures, three levels of features were extracted: radiomics features (200 features capturing shape, texture, and intensity statistics of the tumor and kidney regions), deep learning features (128 features from intermediate layers of a CNN trained to predict tumor stage and grade), and dimensionality-reduced features (320 features derived via Principal Component Analysis and Singular Value Decomposition applied to the full deep feature space).
Combining all three feature levels yielded 648 multi-level features per patient. This feature fusion strategy was designed to capture complementary information: radiomics reflect handcrafted clinical intuitions about relevant anatomy, while deep learning features capture patterns that may not have explicit clinical names but are predictive of surgical complexity.
XGBoost (Extreme Gradient Boosting) was selected as the final classifier because of its strong performance on tabular feature sets, robustness to irrelevant features through its gradient boosting mechanism, and compatibility with SHAP (SHapley Additive exPlanations) for post-hoc interpretability.
XGBoost builds an ensemble of decision trees sequentially, each tree correcting errors made by the previous ensemble. Its regularization terms control overfitting, making it reliable on moderate-sized medical datasets where neural networks might overfit.
Gradient boosting with XGBoost naturally handles missing features, mixed feature types, and nonlinear interactions between anatomical measurements, all of which characterize real-world radiological data from heterogeneous patient cohorts.
SHAP values were computed for each patient to explain which features drove the model's prediction. SHAP decomposes the model's output into contributions from each feature, providing feature importance that is both locally accurate (for each individual prediction) and globally consistent (across the dataset).
Feature importance via SHAP identified tumor shape Maximum 3D Diameter as the single most influential feature, consistent with clinical guidelines that use tumor size as a primary criterion for surgical approach selection. This alignment with clinical knowledge validates the model's interpretability.
The dataset comprised 473 participants: 412 used for internal training and validation, and 61 from a separate external institution used for testing. The external cohort originated from a different center with different imaging protocols and patient demographics.
Internal validation achieved an AUROC of 0.94, indicating very strong discrimination between partial and radical nephrectomy cases based on the 3D CT multi-level features. This performance surpassed existing scoring systems like RENAL and PADUA when evaluated on the same cohort.
External validation on the 61-patient independent cohort achieved AUROC of 0.82, demonstrating that performance holds across institutions despite the differences in CT acquisition parameters and patient populations. A drop from 0.94 to 0.82 is typical for cross-institutional generalization of imaging AI models.
The model was also calibrated, meaning the predicted probabilities of requiring radical nephrectomy matched actual observed rates across probability bins, which is important for clinical use where the probability estimate itself informs pre-operative counseling.
The clinical workflow implications are significant: currently, the decision between partial and radical nephrectomy involves review of CT images by an experienced urologist, often using manual anatomical scoring and subjective judgment about feasibility. Automated prediction could standardize this assessment.
In settings where experienced urological oncologists are not readily available, an AI decision support tool could help less specialized surgeons assess tumor complexity and identify cases that should be referred to high-volume centers for partial nephrectomy attempts.
The SHAP-based explainability layer means the system does not operate as a black box: surgeons can see which specific anatomical features drove the prediction, compare them to their own clinical reading, and decide whether to accept or override the AI recommendation.
For patient counseling, knowing preoperatively that tumor anatomy strongly predicts radical rather than partial nephrectomy allows informed discussions about kidney function preservation, dialysis risk, and potential oncologic trade-offs before the patient undergoes anesthesia.
This study demonstrates that fully automated 3D CT analysis using deep learning segmentation and multi-level feature extraction can objectively predict surgical approach selection with accuracy exceeding manual anatomical scoring systems.
The integration of radiomics, deep learning features, and dimensionality reduction into a single XGBoost model illustrates the value of feature-level fusion over relying on any single feature type, capturing complementary anatomical information from each level.
Cross-institutional validation with AUROC 0.82 provides a meaningful demonstration that the framework generalizes beyond its training environment, a prerequisite for real-world deployment across diverse clinical settings.
Future work should prospectively validate the system in a multi-center trial, assess whether AI-assisted decisions improve patient outcomes compared to standard care, and extend the framework to predict specific surgical difficulty metrics and complication risk beyond the binary procedure choice.