Semi-supervised Medical Image Segmentation with Multiscale Contrastive Learning and Cross-Supervision

Area of Science:

• Medical image analysis
• Machine learning
• Computer vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Limited annotated data hinders the development of accurate segmentation models.
• Existing methods often require extensive manual annotation, increasing costs and time.

Purpose of the Study:

• To develop a semi-supervised segmentation method addressing the scarcity of annotated medical images.
• To enhance model generalizability through cross-supervision and multiscale contrastive learning.
• To improve the performance of medical image segmentation tasks with limited data.

Main Methods:

• A semi-supervised segmentation framework utilizing multiscale contrastive learning.
• Application of input image and feature perturbations with cross-supervision for consistency.
• Incorporation of patch-level and pixel-level contrastive learning to refine feature representations.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art techniques on three public datasets (brain tumor, left atrial, cellular nuclei segmentation).
• Achieved segmentation performance comparable to fully supervised methods despite using limited annotations.
• Validated enhanced intra-class compactness and inter-class separability of learned features.

Conclusions:

• The developed semi-supervised approach effectively overcomes annotation limitations in medical image segmentation.
• The method offers a viable solution for clinical applications requiring accurate segmentation with scarce data.
• The framework's adaptability suggests potential for extension to diverse clinical imaging tasks.