Learning Anatomically Consistent Embedding for Chest Radiography.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Self-supervised learning (SSL) excels at visual representation learning from unannotated data.
• Medical images possess inherent anatomical consistency, a property often underutilized in SSL.

Purpose of the Study:

• Introduce PEAC (patch embedding of anatomical consistency), a novel SSL approach for medical image analysis.
• Exploit anatomical consistency in medical images to improve representation learning.
• Enhance the interpretability of SSL methods in the medical domain.

Main Methods:

• Developed PEAC, a novel SSL approach utilizing patch embedding for anatomical consistency.
• Employed stable grid-based matching to learn global and local anatomical consistencies.
• Transferred pre-trained PEAC models to various downstream medical imaging tasks.

Main Results:

• PEAC significantly outperformed existing state-of-the-art fully and self-supervised methods.
• Demonstrated PEAC's ability to capture anatomical structure consistency across different views and patient demographics.
• Validated enhanced interpretability of the learned representations.

Conclusions:

• PEAC offers a powerful new SSL framework for medical image analysis by leveraging anatomical consistency.
• The method shows significant performance gains and improved interpretability compared to prior approaches.
• Pre-trained PEAC models and code are publicly available for further research and application.