Learning Semantics-enriched Representation via Self-discovery, Self-classification, and Self-restoration.

Area of Science:

• Medical imaging analysis
• Deep learning for healthcare
• Computer vision in medicine

Background:

• Medical images contain rich anatomical semantics and recurring patterns.
• Harnessing these semantics for self-supervised learning in medical imaging is underexplored.

Purpose of the Study:

• To develop a self-supervised learning framework for extracting semantically enriched visual representations from medical images.
• To create a general-purpose, pre-trained 3D model for diverse medical applications.

Main Methods:

• Training deep models using self-discovery, self-classification, and self-restoration of anatomy.
• Developing a novel self-supervised learning framework named Semantic Genesis.
• Evaluating Semantic Genesis on six distinct classification and segmentation tasks across CT, MRI, and X-ray modalities.

Main Results:

• Semantic Genesis significantly outperforms all existing 3D self-supervised and fully supervised models.
• It also surpasses the standard ImageNet-based transfer learning in 2D for medical imaging tasks.
• The framework effectively learns semantic representations from anatomical patterns in medical images.

Conclusions:

• The proposed Semantic Genesis framework enables powerful semantic representation learning from medical images.
• This approach offers a superior alternative to existing methods for medical image analysis tasks.
• The model's effectiveness stems from leveraging inherent anatomical consistency through self-supervised learning.