Test-time adaptation in pathology models can destabilize explanations

A new benchmark study on test-time adaptation (TTA) in computational pathology reveals that while TTA methods improve model accuracy, they can significantly alter the model's explanations. Researchers found that frozen-backbone TTA methods cause minimal drift in explanations, whereas continual methods like CoTTA and RoTTA lead to the largest shifts. The study also highlights that convolutional networks are more sensitive to explanation drift than transformer and foundation models, and that explanation stability is only weakly correlated with adaptation quality, potentially leading to silent failures in clinical applications. AI

IMPACT Highlights a critical reliability issue for AI models in clinical settings, suggesting a need for new evaluation metrics beyond accuracy.

RANK_REASON The item is a research paper detailing a benchmark study on explanation stability in computational pathology. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• CAMELYON17
• CatalyzeX
• CoTTA
• DagsHub
• Gotit.pub
• Hugging Face
• NCT CRC-HE
• RoTTA
• ScienceCast
• Test-Time Adaptation

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →