Transferable Visual Words: Exploiting the Semantics of Anatomical Patterns for Self-Supervised Learning.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical imaging generates similar anatomical patterns across patients, rich in semantic information.
• Deep learning models require extensive annotated data, posing a challenge in medical image analysis.
• Existing self-supervised methods can be enhanced with novel supervision signals.

Purpose of the Study:

• Introduce transferable visual words (TransVW) for annotation-efficient deep learning in medical imaging.
• Develop a self-discovery method to harvest visual words from anatomical consistency.
• Utilize self-discovered visual words as supervision for self-supervised learning.

Main Methods:

• Automated harvesting of visual words based on anatomical consistency.
• Self-supervised learning (self-classification and self-restoration) using visual words.
• Evaluating TransVW's performance in various medical image analysis applications.

Main Results:

• TransVW significantly improves annotation efficiency, leading to higher performance and faster convergence.
• Self-discovered visual words act as effective, free supervision signals for deep models.
• The learned, semantics-enriched image representations are robust and generalizable.

Conclusions:

• Transferable visual words offer a fully autodidactic scheme for self-supervised learning in medical imaging.
• TransVW complements existing self-supervised methods, enhancing their performance.
• The approach saves annotation efforts across diverse applications through transfer learning.