MUE-CoT: multi-scale uncertainty entropy-aware co-training framework for left atrial segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate left atrial segmentation is crucial for analyzing atrial fibrillation.
• Supervised learning methods face limitations due to high annotation costs.
• Semi-supervised learning offers a promising approach using limited labeled and abundant unlabeled data.

Purpose of the Study:

• To develop an efficient semi-supervised learning framework for left atrial segmentation.
• To address the challenge of limited labeled data in medical image analysis.
• To improve segmentation accuracy while reducing manual annotation efforts.

Main Methods:

• Proposed a collaborative training framework: multi-scale uncertain entropy perception (MUE-CoT).
• Utilized a pyramid feature network for learning from unlabeled data.
• Introduced novel loss constraints including diversity loss and multi-scale uncertainty entropy for co-training.
• Implemented a confidence-dependent empirical Gaussian function to weight pseudo-supervised loss.

Main Results:

• The MUE-CoT framework demonstrated superior performance compared to existing semi-supervised methods on public and in-house datasets.
• Achieved high Dice similarity coefficient (e.g., 84.94% ± 4.31) and Jaccard similarity coefficient (e.g., 74.00% ± 6.20) with only 5% labeled data.
• Reported favorable HD95 values (e.g., 4.63 mm ± 2.13), indicating precise segmentation.

Conclusions:

• The MUE-CoT model effectively overcomes data scarcity and high annotation costs in medical imaging.
• The proposed method significantly enhances left atrial segmentation accuracy.
• This framework holds potential for practical applications in clinical analysis and disease recognition.