Deep-learning-based artificial intelligence approaches for grading and progression prediction of clear cell renal cell carcinoma

The grading classification of clear cell renal cell carcinoma (CCRCC) is a crucial prognostic factor. However, manual observation can lead to subjective and inconsistent diagnoses. In this study, we aimed to develop a deep-learning-based grading system for CCRCC using whole-slide image (WSI) and to predict disease progression. A total of 305 CCRCC cases were used to develop the renal cell carcinoma artificial intelligence diagnosis (RCCAID) model for grading (188 training, 47 verification, and 70 testing cases) and the renal cell carcinoma artificial intelligence prognostic (RCCAIP) progression prediction model (288 cases). Progression was defined as the presence of distant and/or lymph node metastasis. Xception, ResNet50, and Inception V3 were employed. Model performance was assessed via confusion matrix, receiver operator characteristic curve (ROC) curve, and classification map. RCCAID achieved area under the curve (AUC) values of 0.987 (internal), 0.996 (in-house), and 0.967 (external). The quadratic weighted Cohen's kappa was 0.864, comparable to senior pathologists (0.871). RCCAIP achieved an AUC of 0.94 for progression prediction. Higher grades correlated with worse prognosis. RCCAID can accurately grade CCRCC in whole-slide images, and RCCAIP is effective in progression prediction, even for low-grade tumors.