Uncertainty-aware AI for tumor subtyping with histology and immunohistochemistry: A multi-center study in Renal Cell Carcinoma

Background and objective: Accurate classification of Renal Cell Carcinoma (RCC) subtypes is essential for personalized therapy, as prognostic outcomes and treatment responses differ markedly among subtypes. Most artificial intelligence applications in RCC focus on histology, while immunohistochemistry applications mainly quantify biomarkers rather than perform subtype classification. In clinical diagnostics, immunohistochemistry remains a valuable complement to histology, though its implementation is limited by cost, labor intensity, and resource constraints. This study aims to develop an uncertainty-aware artificial intelligence framework that integrates histological and immunohistochemistry data to improve RCC subtype classification and optimize laboratory workflows. Methods: We designed a hierarchical, pathologist-guided artificial intelligence framework that integrates uncertainty estimation into RCC subtype classification. High-confidence histological predictions produced by deep learning models are accepted directly, while low-confidence cases automatically trigger targeted immunohistochemistry analysis automatically analyzed using dedicated machine learning algorithms. To address staining variability across institutions, a CycleGAN-based stain transfer module was employed to harmonize color domains and enhance generalizability. The framework was evaluated on multi-center datasets encompassing different staining protocols. Results: The proposed integrated framework demonstrated significant diagnostic performance improvements. Patient-level accuracy reached 97.5% on internal cross-validation and 95% on external cohorts when selective immunohistochemistry refinement was applied. The uncertainty-driven immunohistochemistry module reduced redundant biomarker testing to approximately one-fourth of all cases while maintaining or improving classification confidence. The stain transfer module also effectively mitigated inter-laboratory color discrepancies, supporting consistent model performance across different centers. Conclusions: Our framework combines histology and immunohistochemistry to deliver accurate and efficient subtype classification. By selectively activating immunohistochemistry analysis for low-confidence predictions on histology, the approach optimizes biomarker use and laboratory resources while maintaining high diagnostic reliability. The study highlights the potential of combining deep learning-based histology with targeted immunohistochemistry biomarker assessment to advance precision and reproducibility in cancer diagnostics workflows.