Boundary-Guided Contrastive Learning for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning reduces manual annotation needs in medical imaging.
• Accurate boundary segmentation is critical but challenging due to sparse boundary pixels.
• Existing methods often neglect boundary region performance.

Purpose of the Study:

• To develop a novel semi-supervised method for medical image segmentation focusing on boundary accuracy.
• To leverage unlabeled data more effectively for improved segmentation.
• To enhance the discriminative representation learning for boundary regions.

Main Methods:

• Introduced boundary-guided contrastive learning for semi-supervised medical image segmentation (BoCLIS).
• Employed conservative-to-radical teacher networks with uncertainty-weighted aggregation for pseudo-label generation.
• Utilized a boundary-guided patch sampling strategy and patch-based contrastive learning for representation learning.

Main Results:

• BoCLIS consistently outperformed existing methods across three public datasets.
• Significant Dice Similarity Coefficient (DSC) improvements were observed, particularly in boundary regions (20.47%, 16.75%, 17.18%).
• Demonstrated the effectiveness of the proposed uncertainty-weighted aggregation and boundary-guided sampling.

Conclusions:

• The proposed BoCLIS method effectively addresses the challenge of boundary segmentation in semi-supervised medical image analysis.
• The approach significantly improves segmentation performance, especially at object boundaries.
• The method offers a promising direction for efficient and accurate medical image segmentation using unlabeled data.