Multi-Perturbation Consistency Learning for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Current semi-supervised learning (SSL) methods often rely on consistency learning.
• Existing SSL approaches typically validate consistency learning under single perturbations, which can fail with multiple perturbations, degrading performance.

Purpose of the Study:

• To propose a robust semi-supervised medical image segmentation method leveraging multi-perturbation consistency learning.
• To address the instability and performance degradation associated with multiple perturbations in SSL.

Main Methods:

• Developed a cross-teaching framework with 3D and 2D networks for network perturbations.
• Incorporated strong and weak data augmentation for input perturbations.
• Introduced uncertainty-aware correction algorithms for labeled and unlabeled data to enhance stability.

Main Results:

• The proposed method demonstrated superior performance across four medical imaging datasets (ProstateX, HPH55, ACDC, LA).
• Achieved strong generalization capabilities, outperforming existing methods.
• Effectively maintained performance with limited annotated data.

Conclusions:

• The multi-perturbation consistency learning approach enhances robustness and stability in medical image segmentation.
• This method offers an efficient solution for medical image segmentation with limited annotations.
• The developed algorithm shows significant potential for clinical applications.