A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Deep learning models excel with natural images but often fail with medical scans due to domain shifts.
• Existing visual backbones lack architectural priors for reliable generalization in medical imaging.
• Domain shifts include variations from different hospitals or demographic confounders (sex, race).

Purpose of the Study:

• To investigate model sensitivity to domain shifts in medical imaging.
• To develop a novel deep learning approach that incorporates explicit medical knowledge for improved generalization.
• To enhance the reliability of AI models in analyzing chest X-rays and skin lesion images.

Main Methods:

• Introduction of Knowledge-enhanced Bottlenecks (KnoBo), a concept bottleneck model integrating medical knowledge priors.
• Utilizing retrieval-augmented language models to define concept spaces and train concept recognition.
• Evaluating KnoBo across 20 datasets with diverse domain shifts and two imaging modalities.

Main Results:

• KnoBo significantly outperforms fine-tuned models on confounded datasets, achieving an average improvement of 32.4%.
• PubMed emerged as a highly effective knowledge resource, surpassing others in information diversity and prediction performance.
• The proposed approach demonstrates reduced sensitivity to domain shifts in medical AI models.

Conclusions:

• Integrating explicit medical knowledge via KnoBo enhances deep learning model generalization in medical imaging.
• KnoBo provides a framework for building more robust and reliable AI systems for clinical applications.
• PubMed is a valuable resource for grounding AI models in clinical relevance and mitigating domain shift issues.