Autodidactic dense anatomical models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Humans naturally interpret images using part-whole hierarchies.
• Deep learning models often struggle to explicitly encode these hierarchies, crucial for medical image analysis.
• Anatomical structures in medical imaging possess inherent hierarchical relationships.

Purpose of the Study:

• To introduce Adam-v2, a self-supervised learning framework designed to explicitly learn part-whole hierarchies in medical images.
• To develop a method that enhances anatomy understanding within medical imaging data.
• To create robust and generalizable representations for medical image analysis tasks.

Main Methods:

• Adam-v2 framework utilizes three branches: localizability for distinguishing structures, composability for parts-to-whole learning, and decomposability for whole-to-parts comprehension.
• The framework learns from unlabeled medical images, constructing explicit hierarchies for anatomical structures.
• Embeddings generated by Adam-v2, termed Eve-v2, are evaluated for their anatomical understanding capabilities.

Main Results:

• Eve-v2 embeddings demonstrate zero-shot anatomy understanding, preserving localizability and encoding part-whole relations.
• Emergent properties include understanding anatomical layouts, dense semantic embeddings, symmetry recognition, scale consistency, and cross-image matching.
• Adam-v2 representations show robustness and generalizability, excelling in few-shot learning, full-transfer learning, and anomaly detection.

Conclusions:

• Adam-v2 effectively learns explicit part-whole anatomical hierarchies from unlabeled medical images.
• The Eve-v2 embeddings offer advanced, zero-shot anatomy understanding with emergent properties beneficial for various medical imaging tasks.
• The framework provides robust and generalizable representations, advancing medical image analysis through a novel anatomy learning strategy.