AI Risk Scoring Cuts Unnecessary Kidney Surgery in Large Randomized Trial

An artificial intelligence system that reads CT scans of cystic kidney masses helped radiologists cut the rate of unnecessary surgery in a large randomized trial, according to a study published in eClinicalMedicine. Among patients whose treated kidney lesions turned out to be pathologically benign, the proportion was 16.4 percent with AI assistance versus 40.3 percent under standard care, a difference of 23.9 percentage points. The system, called KITE-CRM, was evaluated across retrospective model development, multicenter external validation, a six-reader crossover study, and a pragmatic randomized implementation trial embedded in routine hospital workflow.

The clinical problem the researchers targeted is a familiar one in oncologic imaging. Cystic renal masses are increasingly encountered as cross-sectional imaging expands, and the central difficulty is not simply spotting cancer but balancing the risk of delayed treatment for clinically significant malignancy against the risk of operating on benign or indolent lesions. The Bosniak classification, the foundation of imaging-based evaluation for these masses, assigns lesions to ordinal categories that compress biologic heterogeneity into a limited number of groups. Interobserver variability remains important, especially for Bosniak IIF and III lesions, so lesions placed in the same category can carry meaningfully different malignancy risks and may not warrant identical management.

KITE-CRM was built as a cascaded CT-based decision-support pathway with four components: kidney-level abnormality detection, cyst-solid gating, lesion-level malignancy discrimination, and downstream characterization of malignant lesions, including clear-cell renal cell carcinoma identification and aggressiveness stratification. All AI modules used the corticomedullary-phase volume from a clinically indicated contrast-enhanced CT examination as their sole imaging input. Kidney and lesion segmentation used LightM-UNet with three-dimensional inference, abnormality detection used a three-dimensional DenseNet121, and cyst-solid routing applied a deterministic lesion-volume rule in which voxels at or below 20 Hounsfield units were classified as cystic-density and a lesion was routed as cystic when such voxels comprised at least half of the segmented volume. The deployed architecture, weights, preprocessing steps, and operating thresholds were locked before external testing.

The malignancy discrimination task operated at a fixed threshold of 0.586442888, with probabilities at or above that value classified as positive. In internal validation, the lesion-level malignancy discrimination AUROC was 0.996, with sensitivity of 97.1 percent and specificity of 94.9 percent. Across external test cohorts, AUROCs for malignancy discrimination ranged from 0.913 to 0.999 in cohorts with adequate malignant case numbers. Operating-point performance varied: external-cohort sensitivity ranged from 79.2 percent to 100.0 percent and specificity from 86.2 percent to 100.0 percent, with the 100 percent sensitivity in one test based on only three malignant lesions. In that same sparse test, low malignant prevalence and 16 false-positive classifications produced a positive predictive value of just 15.8 percent, underscoring that positive outputs also require clinical review rather than automatic intervention.

The study also examined whether the AI could refine risk within Bosniak categories rather than replace them. In a non-training subset of 128 Bosniak IIF-III records, the system achieved an AUROC of 0.932 (95 percent CI, 0.879 to 0.978) and a Brier score of 0.084, supporting complementary within-category risk refinement. Overall, among 1082 abnormal cases with available Bosniak category and AI probability, the system achieved an AUROC of 0.990, though this full evaluable set included development cases and was reported as supportive. The authors emphasize that these analyses do not establish replacement of Bosniak 2019 or prospective CRM-specific treatment superiority.

The multi-reader multi-case study tested how AI assistance changed radiologist performance. Six genitourinary radiologists interpreted 117 cases from an external test cohort enriched for challenging Bosniak IIF-III lesions, with and without AI assistance in a randomized crossover design separated by a washout of at least four weeks. Reader-averaged AUROC increased from 0.825 to 0.973, a difference of 0.149 (95 percent CI, 0.091 to 0.206), and mean interpretation time fell from 9.7 to 5.5 minutes per case. Reader-averaged sensitivity rose from 66.1 percent to 87.3 percent and specificity from 79.8 percent to 96.0 percent. The relative benefit was greatest among less experienced radiologists, a pattern the authors say suggests the system can function as a practical interpretive aid rather than merely a retrospective analytic model.

The randomized implementation study enrolled 5869 consecutive adults undergoing contrast-enhanced abdominal CT at a single tertiary center between November 2025 and February 2026, randomizing them 1:1 to AI-assisted or standard-care reporting. Because enrollment was not restricted to patients with suspected cystic renal masses, most randomized examinations did not contain one, preserving real-world workflow context. In the treated-lesion pathology analysis, the primary endpoint, benign pathology among treated lesions was significantly lower with AI assistance. In a post hoc intention-to-treat analysis of all randomized participants, benign intervention occurred in 0.95 percent of the AI-assisted arm versus 3.54 percent of standard care, a risk difference of minus 2.59 percentage points, corresponding to roughly 26 fewer benign interventions per 1000 participants. Malignant intervention rates were statistically inconclusive, and the authors caution that a non-significant contrast is not evidence of equivalence or absence of undertreatment.

Case-level radiologist reporting time was substantially shorter with AI assistance, with a median of 7.6 minutes versus 18.5 minutes. The abnormality detection module, evaluated in the AI-assisted arm of the prospective study, achieved an AUROC of 0.991 with sensitivity of 98.5 percent and specificity of 94.6 percent. Interaction data show that radiologists fully agreed with AI outputs in 65.9 percent of AI-assisted examinations, modified them in 28.4 percent, and explicitly overrode them in 5.7 percent, indicating that final reports reflected human-AI decision support rather than autonomous AI classification.

The system incorporated explicit uncertainty handling. A prespecified gray-zone pathway flagged examinations with intermediate malignancy probability between 40 and 60 percent, marked discordance between AI risk and Bosniak category, or high-risk clinical context, routing such cases preferentially to multidisciplinary team review. An abstention safeguard instructed radiologists to disregard AI risk estimates when image quality or morphology, such as severe motion artifact, extensive calcification, or atypical appearance, was likely to compromise model reliability. The authors stress that the corticomedullary-phase input came from complete multiphase examinations and does not justify omitting indicated CT phases or MRI, or replacing Bosniak assessment, since a septum or mural nodule inconspicuous in that phase may be better depicted on other phases or MRI.

Several limitations frame the findings. The prospective study was conducted at a single tertiary referral center and randomized all-comer examinations rather than only patients with cystic renal masses, so the randomized findings concern an all-comer reporting pathway rather than proof of CRM-specific clinical superiority. Treated-lesion comparisons condition on intervention after randomization, and evaluability was assessed differently between arms. Race and ethnicity were not collected as standardized self-identified variables, limiting assessment of equity and transportability. Follow-up has not matured enough to assess delayed diagnosis, oncologic outcomes, renal function, or quality of life, and these data do not establish active-surveillance eligibility or long-term oncologic safety. In community or non-tertiary settings without access to specialist renal imaging, the authors recommend that an elevated AI malignancy score prompt referral for tertiary multidisciplinary assessment rather than direct surgical referral based on the AI score alone.

Subject of Research: Multicenter validation and randomized implementation of a CT-based AI risk-support system for cystic renal masses

Article Title: CT-based AI risk support for cystic renal masses: multicenter validation, reader study, and a randomized implementation study

Article References: Chen, C., Kang, Q., Yang, Y., Li, X., Li, N., Zhang, Q., Yang, X., Yang, X., Zhang, Y., Lu, J., Zhang, S., Wang, L., Shao, C., & Bian, Y. (2026). CT-based AI risk support for cystic renal masses: multicenter validation, reader study, and a randomized implementation study. eClinicalMedicine, 101, Article 104279. https://doi.org/10.1016/j.eclinm.2026.104279

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104279

Keywords: artificial intelligence, cystic renal masses, Bosniak classification, CT imaging, radiology, randomized controlled trial, kidney cancer, deep learning, clinical decision support, overtreatment, reader study, external validation