Enhanced Domain Adaptation for Foot Ulcer Segmentation Through Mixing Self-Trained Weak Labels.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Accurate wound parameter measurement is crucial for effective wound management.
• Computerized wound analysis faces challenges due to inexact image segmentation, often caused by limited or inaccurate labels.
• A common issue is the disparity between abundant labeled data in source domains and limited labels in target domains.

Purpose of the Study:

• To develop a novel approach for improving wound image segmentation accuracy and robustness.
• To address the challenge of limited labels in target domains for wound image analysis.
• To enhance the generalization capability of neural networks across diverse wound datasets.

Main Methods:

• A novel approach combining self-training learning and mixup augmentation was proposed.
• A neural network was trained on a source domain to generate weak labels for the target domain via self-training.
• Generated labels were then mixed with source domain labels to retrain the network, improving generalization.

Main Results:

• The proposed method demonstrated substantial improvements in segmentation accuracy and robustness across different data distributions.
• Single-domain experiments showed dice scores of 0.711 on the DFUC 2022 dataset and 0.859 on the FUSeg dataset.
• For domain adaptation, dice scores of 0.714 on DFUC 2022 and 0.561 on FUSeg were achieved when used as target datasets.

Conclusions:

• The combined self-training and mixup augmentation approach effectively enhances wound image segmentation.
• The method shows significant promise for improving automated wound analysis, particularly in scenarios with limited labeled data.
• The study highlights the potential for robust domain adaptation in medical image segmentation tasks.