Optimal Transport and Central Moment Consistency Regularization for Semi-Supervised Medical Image Segmentation

Area of Science:

• Medical image analysis
• Machine learning
• Computer vision

Background:

• Semi-supervised learning enhances model generalizability by utilizing unlabeled data.
• Existing methods often focus on local features, neglecting global structures in medical image segmentation.
• Label propagation techniques can be limited by localized information and prone to local optima.

Purpose of the Study:

• To develop a semi-supervised medical image segmentation framework that integrates global perspectives.
• To improve label propagation and feature learning by considering the entire data distribution and geometric structure.
• To overcome limitations of existing methods that focus on local representations.

Main Methods:

• Introduced a framework integrating optimal transport (OT) and central moment consistency regularization (OTCMC).
• Utilized OT for comprehensive label propagation from labeled to unlabeled data.
• Incorporated central moment consistency regularization to focus on image geometric structures.

Main Results:

• Achieved state-of-the-art (SOTA) performance on multiple medical image segmentation datasets.
• Demonstrated improved generalizability and global structure learning compared to existing methods.
• Successfully applied to NIH pancreas, left atrium, brain tumor, and skin lesion dermoscopy datasets.

Conclusions:

• The proposed OTCMC framework effectively enhances semi-supervised medical image segmentation by leveraging global information.
• Integrating optimal transport and central moment consistency offers a more robust approach to label propagation and geometric learning.
• The method shows significant potential for improving segmentation accuracy in various clinical applications.