The R.E.N.A.L. nephrometry score is a standardized system used to describe the complexity of renal masses on CT imaging, incorporating factors such as tumor radius, exophytic proportion, proximity to the collecting system, anterior or posterior location, and relationship to the polar lines. It was designed to guide surgical planning and predict pathologic outcomes.
Despite its widespread use, the conventional R.E.N.A.L. score has important limitations. Each component is discretized into categorical bins, which causes information loss. For example, a tumor measuring 3.9 cm and one measuring 4.1 cm are scored identically for the radius component despite being nearly the same size, while a tumor measuring 4.1 cm and one measuring 7.0 cm also receive the same score.
This study proposed an AI-enhanced approach called R.E.N.A.L.+ that uses continuous measurements derived from deep learning CT segmentation rather than discretized categories. By retaining the full quantitative information from each nephrometry component, the R.E.N.A.L.+ score aims to improve prediction of malignancy, pathologic stage, tumor necrosis, and surgical approach.
CT scans from 300 patients with renal masses were processed using nnU-Net, a self-configuring deep learning framework for medical image segmentation. nnU-Net automatically adapts its architecture, preprocessing, and training strategy based on the characteristics of the input dataset, making it highly practical for medical imaging applications.
The nnU-Net model segmented the renal mass and surrounding kidney anatomy, enabling automated extraction of continuous measurements corresponding to each R.E.N.A.L. component. Rather than a radiologist manually assigning a score of 1, 2, or 3 to each component, the AI system generated precise numerical values such as exact tumor diameter in centimeters or exact distance to the collecting system in millimeters.
These continuous measurements were then used as inputs to a multivariate logistic regression model, which assigned weights to each component based on its actual predictive value for each clinical outcome. This weighting reflects the empirical importance of each feature rather than the equal weighting implicit in simple summation of categorical scores.
For malignancy prediction, R.E.N.A.L.+ achieved an AUC of 0.69, compared to 0.67 for the standard AI-generated R.E.N.A.L. score and 0.64 for the H-score (a Hounsfield unit-based metric). While the absolute AUC difference for malignancy was modest, the comparison across all outcomes consistently favored R.E.N.A.L.+.
The largest performance gains were seen for predicting high pathologic stage and tumor necrosis. For high-stage disease, R.E.N.A.L.+ achieved AUC 0.82 versus 0.65 for the AI-score and 0.71 for the H-score. For tumor necrosis prediction, AUC was 0.81 versus 0.72 and 0.74 respectively. These are clinically significant improvements in discriminative ability.
For predicting whether a patient could undergo partial nephrectomy (kidney-sparing surgery) versus radical nephrectomy, R.E.N.A.L.+ achieved AUC 0.88 versus 0.74 for the AI-score and 0.79 for the H-score. This finding has direct surgical planning implications, as partial nephrectomy preserves renal function and is the preferred approach when technically feasible.
Feature importance analysis revealed that tumor diameter, the R component of the R.E.N.A.L. score, was the single most important predictor across multiple outcome categories. This finding validates the biological primacy of tumor size as a prognostic factor in renal cell carcinoma and supports existing staging systems that use size as a major criterion.
However, the key advantage of the R.E.N.A.L.+ approach is that it captures the full gradient of tumor diameter rather than grouping all tumors above 7 cm into a single category. A tumor measuring 9 cm carries meaningfully different risk than one measuring 7.5 cm, a distinction that conventional scoring erases.
Other components such as proximity to the collecting system and endophytic proportion also contributed meaningfully to specific outcome predictions, confirming that multiple anatomic features together provide complementary prognostic information beyond what any single feature can offer.
A key advantage of the nnU-Net-based segmentation approach is full automation of measurement extraction. Once the segmentation model is deployed, it can process CT scans and generate R.E.N.A.L.+ components without any manual radiologist input, enabling high-throughput application across large patient volumes.
Automated segmentation also reduces measurement variability between radiologists. Inter-observer variability in nephrometry scoring is a recognized limitation of conventional scoring, with studies showing meaningful disagreement particularly for components such as the nearness to collecting system and polar line relationships. AI-based measurement standardizes this process.
The pipeline from CT input to R.E.N.A.L.+ score output is fully reproducible, a critical requirement for a clinical decision support tool. The team reported good agreement between automated measurements and manual reference measurements, supporting the reliability of the segmentation approach.
The ability to predict partial nephrectomy eligibility preoperatively with AUC 0.88 could assist urologists in surgical planning and patient counseling. Patients predicted to require radical nephrectomy could be prepared accordingly, while those with high R.E.N.A.L.+ partial nephrectomy scores might be prioritized for referral to high-volume centers with expertise in complex kidney-sparing procedures.
Predicting tumor necrosis and high pathologic stage preoperatively has implications for risk stratification and adjuvant therapy discussions. High-stage disease and tumor necrosis are associated with worse prognosis, and early identification of these features could prompt closer surveillance or earlier initiation of systemic therapy discussions after surgery.
The broader implication of this work is that AI-based measurement of continuous imaging features, when combined with appropriate statistical modeling, can extract significantly more prognostic information from CT scans than traditional categorical scoring systems. This principle extends beyond nephrometry to other imaging-based scoring systems used across oncology.