Synergistic Label-Stability Learning for Semi-supervised Left Atrium Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Semi-supervised learning reduces annotation needs in medical image segmentation.
• Existing methods struggle with intra-class variations and non-selective consistency regularization.

Purpose of the Study:

• To propose a novel synergistic label-stability learning (SLSL) framework for semi-supervised medical image segmentation.
• To address limitations of existing methods, specifically intra-class variation and non-selective stability learning.

Main Methods:

• The SLSL framework utilizes a teacher-student model.
• Incorporates pseudo-label learning for easy regions and cyclic real-label learning with class prototypes for intra-class feature regularization.
• Employs difficulty-selective stability learning, focusing regularization on high-entropy (difficult) regions.

Main Results:

• The proposed method effectively leverages unlabeled data for improved segmentation.
• Demonstrated superior performance compared to other semi-supervised methods in left atrium segmentation from MRI.
• The framework successfully handles intra-class variations and focuses learning on challenging areas.

Conclusions:

• The SLSL framework offers a robust solution for semi-supervised medical image segmentation.
• It enhances the exploitation of unlabeled data, outperforming existing approaches.
• This method can facilitate the development of high-performance automatic segmentation models for clinical applications, such as atrial fibrillation treatment, under annotation constraints.