The research, led by Chih-Ying Huang and colleagues at Taipei Veterans General Hospital and National Yang Ming Chiao Tung University in Taiwan, set out to determine whether fully automated AI segmentation could extract meaningful three-dimensional volumetric features from routine pre-treatment CT scans and whether those features correlated with outcomes after cryoablation for renal cell carcinoma. Rather than asking radiologists to eyeball categorical scores, the team deployed two deep learning tools to do the measuring: TotalSegmentator, a robust open-source system capable of segmenting more than 100 anatomic structures in CT images, and a model developed for the 2023 Kidney and Tumor Segmentation Challenge (KiTS23). Both are built on the nnU-Net framework, a self-configuring deep learning architecture that has become a workhorse of modern biomedical image segmentation.

The technical workflow is deceptively simple. For each patient, the pre-treatment nephrographic-phase contrast-enhanced CT scan was fed into the segmentation pipeline, which generated masks for the kidneys, renal tumors, and renal cysts. Masks for the renal sinuses, the central fat-filled cavity of the kidney that houses the collecting system and major blood vessels, were produced through a geometry-based approach described in the study’s supplementary material. Once the masks were generated, the software extracted a series of quantitative 3D features: tumor volume, an automated RENAL nephrometry score, the minimum distance between tumor and renal sinus, and a novel metric called tumor-renal sinus contact area, which quantifies the surface area over which the tumor directly abuts the renal sinus. On a GPU-accelerated workstation, the entire process took approximately five minutes per case, with an additional five minutes when manual correction was needed.

That manual correction proved to be the exception rather than the rule. Of the 116 patients in the study, only 16 required any adjustment to the automated masks, and most of those involved tumors lying adjacent to renal cysts or cases in which the algorithm failed to identify the tumor at all. No case required complete re-segmentation. This level of automation matters because conventional nephrometry scores suffer from well-documented interobserver variability. The R component of the RENAL score, for example, is measured on orthogonal planes and may not reflect the true maximal three-dimensional extent of a tumor, while the exophytic and location components have shown low perfect agreement among human readers. Continuous volumetric features extracted by software sidestep both problems, offering objective, reproducible numbers that capture anatomy in its full spatial complexity.

The study cohort consisted of 116 patients who underwent CT-guided percutaneous cryoablation for renal lesions at a single center between October 2009 and December 2024. The group comprised 85 men and 31 women with a mean age of 70.7 years, and the mean follow-up duration was 2.8 years. The mean tumor volume was 15.1 milliliters and the mean tumor diameter was 3.6 centimeters, consistent with the small, early-stage lesions for which ablation is typically recommended. Clear cell renal cell carcinoma was the most common histologic subtype, accounting for 64.7 percent of tumors. Notably, all cryoablation procedures throughout the fifteen-year inclusion window were performed by a single operator, which the authors suggest may have reduced variability related to procedural technique even as practice patterns evolved.

The clinical outcomes were assessed using standardized definitions from the International Working Group on Image-Guided Tumor Ablation and the SIO and DATECAN consensus guidelines. At the first post-ablation follow-up, 94 percent of patients achieved complete ablation, while 6 percent had residual unablated tumor requiring repeat cryoablation. Among those with complete initial ablation, 8 patients, or 6.9 percent of the cohort, later developed local tumor progression at the ablation margin. One patient with persistent tumor after re-ablation was ultimately found to have suspected venous invasion and a new lesion in the same kidney, requiring systemic therapy. To capture both early failure and later progression, the researchers defined a composite endpoint of local tumor control failure, combining residual unablated tumor and local tumor progression.

The statistical results pointed emphatically toward one feature. In univariable logistic regression, tumor volume, the RENAL nephrometry score, its R component, and tumor-renal sinus contact area were all significantly associated with local tumor control failure. But in a deliberately parsimonious multivariable model, limited by only 15 outcome events and bolstered by bootstrap resampling with 1,000 iterations to confirm coefficient stability, only tumor-renal sinus contact area remained independently predictive, with an odds ratio of 1.379 and a p-value of 0.002. The authors derived a composite risk score, calculated as 0.321 times the tumor-renal sinus contact area plus 0.023 times the tumor volume. In receiver operating characteristic analysis, this score achieved an area under the curve of 0.765 for predicting local tumor control failure, numerically higher than the 0.664 achieved by the RENAL nephrometry score, although the difference did not reach statistical significance on the DeLong test. With an optimal cutoff of 1.51, the score yielded a sensitivity of 67 percent and a specificity of 86 percent.

Two representative cases illustrate the score’s clinical texture. A 65-year-old man with biopsy-proven clear cell renal cell carcinoma had a 21.3-milliliter tumor with a tumor-renal sinus contact area of just 0.31 square centimeters, producing a risk score of 0.59, well below the cutoff. He underwent cryoablation and remained stable on follow-up imaging for more than four years. By contrast, a 91-year-old man with a similarly sized tumor of 25.2 milliliters but a contact area of 9.23 square centimeters scored 3.54, far above the threshold, and developed local tumor progression just nine months after treatment. In the Cox proportional hazards analysis, larger tumor-renal sinus contact area was also significantly associated with local tumor progression among patients who had achieved complete initial ablation, with a hazard ratio of 1.255 and a p-value of 0.026.

Why should contact with the renal sinus matter so much? The authors offer a biologically plausible but unproven explanation: the renal sinus contains major blood vessels whose continuous blood flow can dissipate cold energy from the ablation zone, a phenomenon known as the cold-sink effect. Tumors with extensive contact along the tumor-sinus interface may therefore be more susceptible to incomplete freezing, and indeed, viable tumor foci identified on follow-up imaging in this cohort were frequently located adjacent to the renal sinus. The researchers caution that this mechanism remains a hypothesis, since the study did not directly evaluate renal sinus vascular anatomy, tissue perfusion, intraprocedural temperature distribution, or ablation margin adequacy. Even so, the finding suggests that automated segmentation can capture an anatomically meaningful feature that is nearly impossible to quantify visually, information that existing categorical scoring systems simply do not encode.

The authors are careful to frame their conclusions as exploratory. The study was retrospective and single-center, the number of outcome events was small, the AI segmentation tools have not yet been formally validated across institutions and imaging protocols, and the composite risk score has not undergone external validation. Its discriminative performance was only moderate, and its incremental value over conventional predictors should be interpreted cautiously. Still, the implications are tantalizing. An objective, automated risk score generated in minutes from an existing CT scan could help clinicians identify anatomically challenging tumors, guide patient selection between ablation and surgery, and tailor surveillance intensity, with higher-risk patients potentially benefiting from closer imaging follow-up. Future work, the authors suggest, may extend the approach to segment renal vessels and ablation zones, register pre-treatment and intraprocedural images, and quantitatively assess ablation margins, bringing the field closer to a fully computational framework for predicting and ultimately preventing local treatment failure in kidney cancer.

Subject of Research: AI-based automated segmentation-derived 3D volumetric imaging features and outcomes after cryoablation for renal cell carcinoma

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

Article References: Huang, C.-Y., Hong, J.-A., Chang, N.-W., Li, C.-C., Liu, C.-A., & Shen, S.-H. (2026). Association between AI-based automated segmentation-derived 3D volumetric imaging features and outcomes after cryoablation for renal cell carcinoma. CVIR Oncology, 2(1), Article 24. https://doi.org/10.1007/s44343-026-00053-3

Image Credits: AI Generated

DOI: 10.1007/s44343-026-00053-3

Keywords: renal cell carcinoma, cryoablation, artificial intelligence, automated segmentation, 3D volumetric imaging, tumor-renal sinus contact area, RENAL nephrometry score, tumor ablation, risk score, kidney cancer, radiology, predictive modeling