Introduction:

Nephrometry scores, namely R.E.N.A.L and PADUA, provide an objective quantification of tumor complexity respective to renal anatomy and guide appropriate surgical decision making.  Its adoption in community practice has been modest, owing to the un-reimbursed time necessary to extrapolate such values, score ambiguity and inter-observer variability.  The use of Machine Learning (ML) and namely Deep Learning (DL) in predicting extent and complexity of tumors with as much accuracy as human experts remains intriguing.  We aimed to generate a R.E.N.A.L. nephrometry score using artificial intelligence (AI-Score) based off preoperative CT scans and compare it to expert human-generated nephrometry scores (H-score), in its ability to predict pathologic and surgical outcomes.

Methods:

300 patients with pre-operative CT scans were identified from a cohort of 544 consecutive patients undergoing surgical extirpation for suspected renal cancer at a single institution. A deep neural network approach was used to automatically segment kidneys and tumors and geometric algorithms were developed to estimate each component of  R.E.N.A.L. based on the segmented regions (AI-score)(figure 1). Tumors were manually scored by medical personnel blinded to the AI-score.  AI- and H-score agreement was assessed using Lin’s concordance correlation, and their predictive abilities for both oncologic (Figure 2) and perioperative outcomes were assessed using areas under the curve.

Results:

Median age was 60 years, 40% were female. Median tumor size was 4.2 cm, 91.3% had malignant tumors, including 27%, 37% and 24% with high-stage, high-grade and necrosis, respectively. There was significant agreement between H-scores and AI-scores (Lin’s ⍴=0.59). Both AI- and H-scores similarly predicted 1) presence of malignancy (AUC 0.67, p =0.0026 vs AUC 0.63, p =0.037), 2) presence of >pT2 disease (AUC 0.65 p<0.0005 vs AUC 0.71 p<0.0005) and 3) presence of high grade tumors (AUC 0.63 p<0.0005 vs AUC 0.65 p<0.0005).  Several other perioperative outcomes were also associated with AI Scores such as estimated blood loss, perioperative transfusion requirements and change in eGFR following surgery. Neither AI score nor H Scores predicted complications/readmission. 

Conclusion:

Fully automated, unambiguous AI-generated nephrometry scores are comparable to human generated nephrometry scores and robustly predict a wide variety of patient-centered outcomes. This AI-based scoring has the potential to help widen the adoption of R.E.N.A.L. and other nephrometry scores, optimize nephrometry scoring systems and improve individual predictions of a wide variety of perioperative outcomes.

Funding: no