Predicting Renal Tumor Pathology From Intraoperative Gross Appearance: An AI-based Pilot Study.

OBJECTIVE

To evaluate the feasibility of using convolutional neural networks (CNNs) and vision transformers (ViTs) to predict renal tumor pathology intraoperatively based on gross appearance.

MATERIAL AND

METHODS

Intraoperative images were retrospectively extracted from surgical recordings of patients undergoing partial nephrectomy between 2008 and 2024. Static frames obtained prior to arterial clamping were curated and linked with final pathology. A ResNet50-based CNN and the General Surgery Vision Transformer (GSViT) were trained to classify 6 tumor types: clear cell RCC (ccRCC), papillary RCC (pRCC), chromophobe RCC (chRCC), hybrid oncocytic tumors, oncocytoma, and angiomyolipoma (AML). Models were trained with transfer learning, evaluated on held-out test data, and assessed using accuracy, AUC-ROC, and confusion matrices.

RESULTS

A total of 443 images from 118 patients (136 surgeries) were analyzed, including ccRCC (n = 149), pRCC (n = 97), chRCC (n = 42), hybrid tumors (n = 81), oncocytoma (n = 43), and AML (n = 31). In binary classification, the CNN achieved the highest AUCs for ccRCC (0.74), chRCC (0.70), hybrid tumors (0.73), and AML (0.70). Multi-class CNN performance was more variable, with notable AUCs for pRCC (0.70) and oncocytoma (0.71). The GSViT model underperformed across most categories, demonstrating prediction bias toward ccRCC. Attempts to unfreeze pretrained backbones led to rapid overfitting, underscoring dataset limitations.

CONCLUSION

CNN-based models demonstrate moderate ability to classify renal tumor pathology intraoperatively from gross appearance, providing proof of concept for AI-assisted surgical decision-making. Larger datasets and external validation are needed before clinical application.