Predicting clear cell subtype for kidney tumors from cross-sectional imaging using artificial intelligence.

Abstract

429 Background: Clear cell renal cell carcinoma (ccRCC) is the most common subtype of kidney cancer in adults. The standard of care for localized RCC is partial or radical nephrectomy, or active surveillance for small tumors. The histologic tumor diagnosis may assist in shared decision making, but it is typically unknown prior to nephrectomy in the absence of a renal mass biopsy. Additionally, biopsies have imperfect diagnostic performance due to tumor heterogeneity. We sought to explore whether histopathologic diagnoses could be predicted based on radiomic data. In this study, we present the application of artificial intelligence (AI) to predict the clear cell subtype directly from imaging. Methods: We retrospectively reviewed patients who underwent nephrectomy and had a preoperative contrast-enhanced abdominal CT imaging available from 10/2009 to 7/2024 at a single large health system. A ResNet-50 architecture was fine-tuned to predict whether the tumor was ccRCC. 5-Fold cross-validation was used to obtain predictions for all patients. A second set of models was trained on the subset of patients with tumor sizes of 3cm to 7cm. The ensemble of 5 models was used to perform external validation on an independent dataset. DeLong’s tests were performed to compare AUC of the AI model and tumor size prediction model. Results: A total of 1,642 nephrectomy patients at the primary institution had available imaging and subtype classifications, of which 822 fell into the 3-7 cm cohort. For the whole cohort, the model achieved an AUC of 0.71, significantly outperforming the tumor size AUC of 0.54 (p = 6.6e-21). For the 3-7cm subset, the model achieved an AUC of 0.81, again significantly outperforming the tumor size AUC of 0.49 (p = 2.1e-26). For the external validation cohort, the model did significantly outperform tumor size with AUCs of 0.59 and 0.44 respectively (p = 0.009) on the 3-7 cm cohort. However, on the whole external validation cohort, the model did not significantly outperform tumor size. Conclusions: This study demonstrates that a computer-vision-based AI model is capable of predicting clear cell renal cell carcinoma from CT images in kidney cancer patients better than tumor size. This presents an alternative to biopsies for pre-surgical decision-making in patients whose treatment decision may be influenced by knowing their renal tumor subtype.