Researchers unveil SAGE for AI explainability in pathology

Researchers from institutions including [names] have developed **SAGE (Semantic Attention Global Explanations)**, a breakthrough framework for interpretability in computational pathology. Published on arXiv (arXiv:2608.02803), SAGE addresses a critical gap in attention-based multiple instance learning (ABMIL) models, which dominate slide-level prediction but only provide local, pixel-level attention maps without semantic context.

SAGE operates as a post-hoc explainability layer that freezes existing ABMIL models and extracts global, language-grounded explanations using a pathology vision-language model. It scores image patches against a dictionary of 25 predefined histological concepts (e.g., necrosis, angiogenesis) and aggregates these scores based on the model's learned attention weights. When tested on seven TCGA cancer cohorts and three foundation models, SAGE successfully recovered known prognostic features—such as necrosis’s adverse association—and uncovered cancer-specific biology, including a favorable angiogenic signature in renal cell carcinoma that aligns with established molecular subtypes. Ablation studies confirmed that these associations depend on the model’s attention mechanisms rather than simple feature prevalence.