The study, published in Frontiers in Urology, involved 413 patients with clear cell renal cell carcinoma, the most prevalent histological subtype of kidney cancer, accounting for roughly 70 to 80 percent of all cases. Renal cell carcinoma itself is the seventh most common malignancy worldwide and carries the highest mortality among urological cancers. Between June 2019 and April 2024, 312 of the patients in the cohort underwent partial nephrectomy while 101 underwent radical nephrectomy, with the final surgical decision made by attending urologic surgeons based on tumor anatomy, patient comorbidities, baseline kidney function, and surgeon expertise, reflecting real-world practice rather than a rigid algorithm. Notably, the R.E.N.A.L. Nephrometry Score, the study’s benchmark comparator, was calculated retrospectively for research purposes and was not formally used by the treating surgeons in their preoperative deliberations.

The R.E.N.A.L. Nephrometry Score, developed by Kutikov and Uzzo, classifies renal tumors into low, moderate, and high complexity groups based on cross-sectional imaging, assigning points for radius, endophytic or exophytic growth, nearness to the collecting system, anterior or posterior position, and location relative to the polar lines. More than ten competing scoring systems exist, including the PADUA score and the C-index, but the RNS was chosen for comparison because of its widespread adoption, its well-established correlation with perioperative outcomes, its extensive multi-center validation, and its intuitive component-based structure. Its fundamental limitation, however, is that it quantifies anatomy alone. Preoperative biopsy can probe tumor biology, but it is invasive, prone to sampling error, and carries procedural risks, leaving a clear gap for a non-invasive tool that captures quantitative tissue characteristics beyond what the eye can see on conventional images.

That gap is precisely where radiomics enters the picture. Radiomics is the high-throughput extraction of hundreds of quantitative features from medical images, converting what radiologists traditionally interpret qualitatively into reproducible numerical data. Subvisual texture patterns in tumors have been shown to correlate with pathological grade, genotype, and clinical outcomes across multiple cancer types, including kidney cancer. The research team extracted a total of 1,210 features from each tumor’s corticomedullary-phase CT scan, a contrast-enhanced imaging phase chosen because it best captures tumor vascularity and heterogeneity. The features spanned seven classes: first-order statistics describing voxel intensity distributions, three-dimensional shape descriptors, and five texture matrix families including the gray-level co-occurrence, run length, size zone, dependence, and neighboring gray tone difference matrices.

Rigorous standardization underpinned the entire pipeline. Images were resampled to isotropic 1-millimeter voxels, intensity histograms were normalized, and Laplacian of Gaussian filters and Daubechies-4 wavelet transforms were applied to capture texture information at multiple spatial scales, all in accordance with the Image Biomarker Standardization Initiative guidelines. Three blinded observers independently segmented tumors in 50 randomly selected cases to assess reproducibility, and features were retained only if their intraclass correlation coefficients exceeded 0.75 for both inter- and intra-observer agreement. That filter kept 851 features, or about 70 percent of the total, with a median ICC of 0.89. The team then applied LASSO regression, a machine-learning technique that penalizes less informative variables and shrinks their coefficients toward zero, performing feature selection exclusively within a training cohort of 289 patients to prevent data leakage. Six features survived the selection, and the resulting radiomic score was tested in an independent internal validation cohort of 124 patients.

The results were striking. The radiomic model achieved an area under the receiver operating characteristic curve of 0.886, with a 95 percent confidence interval of 0.81 to 0.91, compared with 0.793 for the nephrometry score model, a difference confirmed as statistically significant by the DeLong test with a P-value of 0.015. At the optimal threshold, the radiomic model delivered sensitivity of 83.9 percent versus 71.3 percent, specificity of 82.1 percent versus 74.6 percent, and accuracy of 83.2 percent versus 72.8 percent. Calibration, the agreement between predicted probabilities and actual outcomes, was also superior, with a Brier score of 0.11 against 0.16. Decision curve analysis, which quantifies the clinical net benefit of using a model across a range of decision thresholds, showed the radiomic model providing greater value across most clinically relevant probabilities, yielding a net benefit of 0.32 versus 0.21 for the anatomical score at a threshold of 0.6.

To open the black box of machine-learning prediction, the researchers applied SHAP analysis, a game-theoretic interpretation method that quantifies each feature’s contribution to individual predictions. Tumor size, captured by the maximum two-dimensional diameter, emerged as the single most influential predictor, which aligns with the well-known relationship between tumor dimensions and surgical complexity. But texture features ranked close behind, particularly a gray-level co-occurrence measure and a texture strength descriptor. Lower values of the co-occurrence feature, indicating less complex and more predictable internal texture, were associated with a higher likelihood of partial nephrectomy. This makes biological sense: increased texture complexity in clear cell renal cell carcinoma has been linked to higher Fuhrman grade, greater microvascular density, and increased intratumoral heterogeneity, all of which can produce more infiltrative growth patterns, less distinct pseudocapsules, and more necrotic areas that complicate dissection and reconstruction.

Perhaps the most intriguing finding concerned the limits of the traditional score itself. The relationship between nephrometry complexity and surgical approach was non-monotonic in this cohort: radical nephrectomy rates were 15.2 percent for low-complexity tumors, 34.7 percent for moderate-complexity tumors, and just 15.2 percent for high-complexity tumors. The authors caution that this counterintuitive pattern may reflect selection bias, the small number of high-complexity cases, or institutional practice patterns favoring attempted nephron-sparing surgery even in difficult cases. Interestingly, larger maximum diameter predicted a higher probability of partial nephrectomy in this dataset, likely because surgeons were more willing to attempt nephron-sparing surgery for larger but anatomically favorable peripheral tumors, while small but deeply embedded or central lesions, technically challenging despite their size, more often resulted in radical nephrectomy. The combined model integrating both the radiomic score and the anatomical score reached an AUC of 0.888, statistically indistinguishable from radiomics alone, suggesting the quantitative imaging signature subsumed most of the anatomical information.

The authors are appropriately candid about the caveats. The endpoint predicted was the actual surgical procedure performed, not an objective measure of tumor resectability, meaning the model captures institutional decision-making patterns that integrate surgeon judgment and patient factors. There is also a degree of circularity in the comparison: because anatomical assessment informed the real surgical decisions, the RNS model’s performance was likely biased upward, making the radiomic model’s superiority all the more noteworthy. Fuhrman grade differed between the surgical groups but is unknown preoperatively; a post-hoc sensitivity analysis adjusting for grade confirmed the radiomic model’s performance remained strong at an AUC of 0.875, indicating the imaging signal is not merely a surrogate for tumor aggressiveness. The study was single-center and restricted to pathologically confirmed clear cell histology, which is rarely known before surgery, and radiomic features remain sensitive to scanner manufacturers and acquisition protocols.

Those limitations explain why the researchers emphasize that prospective, multi-center validation is essential before any routine clinical adoption, with calibration to local partial nephrectomy rates and eventual linkage to long-term oncological outcomes such as overall and progression-free survival. Still, the proof of concept is compelling: a standard preoperative CT scan, already acquired in virtually every patient with a kidney mass, contains quantitative information about tumor size and internal heterogeneity that predicts surgical decision-making more accurately than the anatomical score that has dominated the field for over a decade. If validated externally, such radiomic decision-support tools could help surgeons and patients navigate the choice between nephron-sparing surgery and radical nephrectomy with greater confidence, preserving kidneys where possible while avoiding futile attempts at complex reconstructions, all without a single additional test, needle, or scan.

Subject of Research: CT-based radiomics for predicting partial nephrectomy likelihood in clear cell renal cell carcinoma

Article Title: Radiomic model versus R.E.N.A.L. nephrometry score for predicting the likelihood of undergoing partial nephrectomy in clear cell renal cell carcinoma

Article References: Liu, F.-D., Yu, H.-L., Wang, D.-J., Ju, X.-B., & Yang, Y.-H. (2026). Radiomic model versus R.E.N.A.L. nephrometry score for predicting the likelihood of undergoing partial nephrectomy in clear cell renal cell carcinoma. Frontiers in Urology, 6, Article 1928982. https://doi.org/10.3389/fruro.2026.1928982

Image Credits: AI Generated

DOI: 10.3389/fruro.2026.1928982

Keywords: radiomics, clear cell renal cell carcinoma, partial nephrectomy, R.E.N.A.L. nephrometry score, CT imaging, machine learning, LASSO regression, SHAP analysis, tumor heterogeneity, nephron-sparing surgery, urologic oncology, decision curve analysis