Background: Previous studies demonstrated the independent association between infiltrative renal masses(IRMs) and the aggressiveness of the tumor. However, routine documentation of infiltrative features is not routinely done in clinical practice. The objective of this study was to evaluate the ability of Artificial Intelligence(A.I.) to automatically identify infiltrative features and compare the predictive outcomes with physician scores.
Methods: Three hundred patients with suspected renal cancer(2010-2018) were included. Preoperative contrast-CT-imaging was available for classification. Two radiologists, two urologists and one physician reviewed all images and determined whether each renal mass had infiltrative features. IRM was required to have a poorly defined interface with renal parenchyma and non-elliptical shape. Human-IRM(hIRM) was based on majority agreement. A.I.-generated IRM(AI-IRM) was developed by image segmentation and geometric algorithms using IRM criteria. Human raters were blinded to AI-IRM classification. Logistic regression assessed the association between IRM and oncologic outcomes.
Results: The median age was 60 years and 60% were male. The median tumor size was 4.1cm and 92% were malignant. Seventy-five patients(27%) had pT3 or above stage. Sixty-nine patients(25%) had tumor with necrosis and 27(10%) had lymphovascular invasion(LVI). Interobserver variability was high among the expert scorers. There were 37 hIRMs and 31 AI-IRMs with significant correlation between two scores. On univariate analysis, hIRM and AI-IRM were significantly associated with pathologic stage(pT3 or above), necrosis, LVI and high-grade. After adjusting for age, gender, and tumor size, hIRM remained a predictor for pathologic stage and LVI. In the multivariate models, AI-IRM was no longer statistically significant. Both hIRM and AI-IRM were not significantly associated with overall and cancer-specific survival(Figure 1).
Conclusions: Human-designated infiltrative features were associated with aggressive tumor biology; however, interobserver variability was high. The AI classification of infiltrative features is reproducible, fast, and automatic. AI-classification was comparable to the human-generated scores and more patients are needed to identify patient oncological outcomes.
