Introduction:

Partial nephrectomy (PN) is the standard of care for localized kidney tumors when technically feasible, while radical nephrectomy (RN) remains necessary for larger or anatomically complex tumors. PN is preferred due to renal function preservation but carries a slightly higher risk of postoperative complications and longer operative times. Additional perioperative factors, such as extended warm ischemia time and significant intraoperative blood loss, can contribute to long-term renal impairment and worse outcomes. Accurately predicting these perioperative risks is crucial for selecting the most suitable surgical approach. Nephrometry scores such as RENAL. can quantify PN surgical complexity, but rely on human-defined heuristics about morphology and achieve only modest correlation with perioperative outcomes. In this study, we present an artificial intelligence (AI) based image analysis approach to predict the likelihood of PN versus RN, as well as expected ischemia time and estimated blood loss, from preoperative imaging.

Methods:

We retrospectively reviewed patients who underwent partial or radical nephrectomy and had a preoperative contrast-enhanced abdominal CT available from 10/2009 to 7/2024 at a single large health system. A ResNet-50 architecture was fine-tuned to predict whether the patient underwent a radical nephrectomy vs partial nephrectomy. 5-Fold cross-validation was used to obtain predictions for all patients. To target patients with the greatest diagnostic uncertainty, a subanalysis of the patients with tumor sizes from 3cm to 7cm was performed. RENAL scores on the same cohort were estimated by an AI system published previously. The predictive utility of our model was compared against RENAL score and tumor size via AUC-ROC, and AUC values were compared using DeLong’s test. Regression models were fit to predict estimated blood loss and ischemia time from the radical nephrectomy prediction probabilities and RENAL scores and were compared with Pearson correlation coefficient and Steiger’s test.

Results:

A total of 1,641 nephrectomy patients had available imaging and perioperative information. Prediction of radical nephrectomy had an AUC of 0.88 and showed greater predictive capability than tumor size (AUC: 0.86, DeLong’s p=0.008) and RENAL score (AUC: 0.80, p=3.2e-12). In the 3cm to 7cm cohort subanalysis, prediction of radical nephrectomy had an AUC of 0.78 and showed greater predictive capability than tumor size (AUC: 0.73, p = 0.001) and RENAL scores (AUC: 0.72, p = 4.0e-4). The coefficients for predicted ischemia time and estimated blood loss were calculated and compared using both the RENAL nephrometry score and a predictive model (Table 1). Both the RENAL score and the model were associated with outcomes, and the AI-based predictions performed significantly better for estimating blood loss.

Conclusion:

This study highlights the capability of AI-based image analysis to more accurately predict the likelihood of radical versus partial nephrectomy compared to a regression model of tumor size or RENAL score. Additionally, it demonstrates the model’s ability to estimate total ischemia time and intraoperative blood loss. This provides a more personalized prediction of perioperative outcomes, supporting improved surgical planning, patient counseling, and risk stratification for kidney cancer patients. Prospective validation and external testing will be key next steps toward clinical deployment and potential integration of AI-based imaging tools into preoperative planning workflows.

Funding: This work was supported in part by the Department of Defense under Award Number HT94252310918. Additional funding was provided by Climb 4 Kidney Cancer, a nonprofit organization dedicated to advancing research, education, and advocacy for kidney cancer.