Clear cell renal cell carcinoma (ccRCC) represents the predominant histological variant of renal cell carcinoma (RCC) and necessitates precise pathological grading to inform accurate prognostic assessment. Nevertheless, conventional grading techniques depend primarily on the subjective evaluations of pathologists, which frequently result in considerable inconsistency. Although generative artificial intelligence (GenAI) has exhibited encouraging capabilities in medical imaging, its integration into digital pathology has received limited attention. This investigation examines the effectiveness of three leading multimodal GenAI systems—GPT-4o, Claude-3.5-Sonnet, and Gemini-1.5-Pro—in performing ccRCC grading and forecasting patient outcomes. The analysis encompassed 499 ccRCC whole-slide images sourced from The Cancer Genome Atlas along with 349 independent external cases drawn from two separate validation cohorts. Employing a standardized prompt repetition protocol and a variance-driven stability assessment procedure, the GenAI models were directed to identify 17 distinct pathological characteristics. Stability of these extracted features was quantified via the intraclass correlation coefficient (ICC). The derived features, integrated with 3 clinical parameters, facilitated the development of grading and survival prediction models through logistic regression as well as 113 diverse machine learning techniques. Comparative benchmarking was conducted against established approaches including CellProfiler, ResNet-50, DenseNet-121, attention-based multiple instance learning (MIL), and Pathology Language and Image Pre-training, utilizing the concordance index (C-index) and area under the receiver operating characteristic curve (AUC) as primary metrics. Among the GenAI systems, Claude-3.5-Sonnet demonstrated superior performance (ICC = 0.76; micro-average AUC = 0.87), surpassing both ResNet-50 (AUC = 0.78) and attention-based MIL (AUC = 0.70). The optimal prognostic models generated by this system yielded an average C-index of 0.739 and reliably differentiated patients into high- and low-risk categories. Prominent predictive variables included tumor stage, calcification, sarcomatoid features, and vascular architecture. These findings indicate that GenAI, with particular emphasis on Claude-3.5-Sonnet, can substantially improve the precision and reproducibility of ccRCC pathological evaluation, offering significant promise for clinical deployment, particularly in environments with restricted resources.
