Disease Subtyping for Computational Pathology Benchmarking with TCGA dataset —White Paper

Advances in artificial intelligence (AI) and large language models (LLMs) are transforming computational pathology, yet the field still lacks robust public benchmarkings for disease subtyping. The Cancer Genome Atlas (TCGA) offers a rich repository of whole slide images (WSIs) and associated data for various types of human cancers; however, its primary diagnostic labels may pose challenges for AI models due to inconsistencies, overlapping categories, and high interobserver variability, limiting its effectiveness for robust model validation. This white paper addresses these limitations by proposing a rational and standardized classification framework for disease subtypes within the TCGA dataset, specifically designed for benchmarking pathology AI models and LLMs. We introduce two distinct classification schemes: one optimized for WSI classification and another for pathology report classification, acknowledging that morphologically similar findings may carry different nomenclature based on context. Our framework systematically re-evaluates the TCGA primary diagnoses across various organ systems, merging ambiguous or outdated categories and excluding rare entities. This refined approach aims to provide a more consistent and AI-ready dataset, thereby facilitating more accurate and reproducible benchmarking of foundation models and LLMs in computational pathology.