Towards artificial intelligence‐assisted digital pathology: A systematic evaluation of multimodal generative artificial intelligence in clear cell renal cell carcinoma assessment

Abstract Clear cell renal cell carcinoma (ccRCC), the most common subtype of RCC, requires accurate pathological grading for effective prognosis. However, current grading methods rely heavily on subjective pathologist assessment, leading to variability. While generative artificial intelligence (GenAI) has shown promise in medical imaging, its application in digital pathology remains underexplored. This study evaluates the performance of three multimodal GenAI models—GPT‐4o, Claude‐3.5‐Sonnet, and Gemini‐1.5‐Pro—in ccRCC grading and prognosis prediction. A total of 499 ccRCC slides from The Cancer Genome Atlas and 349 external samples from two independent cohorts were analyzed. A standardized prompt repetition mechanism and variance‐based stability validation method guided GenAI models in extracting 17 pathological features. Feature stability was assessed using intraclass correlation coefficient (ICC). These features, combined with 3 clinical variables, were used to build grading and prognostic models via logistic regression and 113 machine learning algorithms. Performance was benchmarked against CellProfiler, ResNet‐50, DenseNet‐121, attention‐based multiple instance learning (MIL) and Pathology Language and Image Pre‐training, using the concordance index (C‐index) and area under the receiver operating characteristic curve (AUC). Claude‐3.5‐Sonnet outperformed the other two GenAI models (ICC = 0.76; micro‐average AUC = 0.87), exceeding ResNet‐50 (AUC = 0.78) and attention‐based MIL (AUC = 0.70). Its top prognostic models achieved an average C‐index of 0.739, effectively stratifying high‐ and low‐risk patients. Key predictors included stage, calcification, sarcomatoid differentiation, and vascular networks. GenAI, particularly Claude‐3.5‐Sonnet, enhances accuracy and consistency in ccRCC pathology, showing strong potential for clinical use, especially in resource‐limited settings.