KnoBo brings textbook knowledge priors to medical image AI

Researchers presented Knowledge Enhanced Bottlenecks (KnoBo) at NeurIPS 2024, a method that improves the robustness of medical image models to domain shifts. It extracts diagnostic concepts from medical texts such as PubMed and builds concept bottlenecks on them, making model decisions more interpretable.

• KnoBo builds concept bottlenecks from PubMed and medical textbook knowledge
• The method outperforms standard CNN and ViT models on medical image datasets under domain shifts
• Explicit medical concepts make model decisions more transparent to clinicians
• The paper's author is Yue Yang, a PhD student at the University of Pennsylvania