Researchers have developed SAGE, a new post-hoc framework designed to provide semantic, language-grounded explanations for attention-based multiple instance learning (ABMIL) models used in computational pathology. Unlike existing methods that only offer local attention maps, SAGE extracts global explanations by scoring image patches against a dictionary of histological concepts and aggregating these scores based on the model's learned attention. This approach allows for the quantification of how specific concepts relate to prediction risk across patient cohorts, as demonstrated by its application to survival prediction in seven TCGA cancer datasets using three foundation models. AI
IMPACT Enhances interpretability of AI models in medical diagnostics, potentially aiding pathologists in biomarker discovery and clinical decision-making.
RANK_REASON The cluster describes a new research paper detailing a novel framework for explainability in AI models used for computational pathology. [lever_c_demoted from research: ic=1 ai=1.0]
- ABMIL
- Attention-based multiple instance learning
- foundation model
- Necrosis
- Pathologists' assistant
- renal cell carcinoma
- SAGE
- The Cancer Genome Atlas
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
