A semi-supervised domain adaptive medical image segmentation method based on dual-level multi-scale alignment.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation often suffers from limited labeled data and domain shifts, impacting model generalizability.
• Domain shifts can occur between similar (homologous) or dissimilar (heterologous) medical image datasets.

Purpose of the Study:

• To propose a novel method for robust medical image segmentation under domain shift conditions with limited labeled data.
• To enhance model adaptation to unlabeled target domain data.

Main Methods:

• A model was trained using labeled source and target domain data to adapt to unlabeled data.
• Alignment was performed at the style level using multi-scale stylistic features to enhance unlabeled target image features.
• Inter-domain alignment maximized category centroid similarity between target and mixed image data.
• A fused supervised and alignment loss computation method was developed.

Main Results:

• Validation was conducted on constructed homologous and heterologous cross-domain medical image datasets.
• The proposed method demonstrated superior comprehensive performance compared to standard semi-supervised and domain adaptation techniques.
• The method effectively addressed challenges posed by domain shifts in medical image segmentation.

Conclusions:

• The novel method offers a significant advancement in medical image segmentation, particularly in scenarios with limited data and domain variability.
• The dual-level alignment strategy (style and inter-domain) is key to the method's effectiveness.
• This approach holds promise for improving the reliability and accuracy of automated medical image analysis.