LE-UDA: Label-Efficient Unsupervised Domain Adaptation for Medical Image Segmentation.

Area of Science:

• Medical imaging
• Deep learning
• Computer vision

Background:

• Deep learning excels in medical image segmentation but requires large labeled datasets, which are costly and time-consuming to create.
• Existing unsupervised domain adaptation (UDA) methods struggle with limited source data and significant domain shifts, particularly across different imaging modalities like MRI and CT.

Purpose of the Study:

• To address the challenge of label scarcity in unsupervised domain adaptation for medical image segmentation.
• To propose a novel framework, Label-Efficient Unsupervised Domain Adaptation (LE-UDA), for effective cross-modality segmentation with limited annotations.

Main Methods:

• Developed a generic framework (LE-UDA) incorporating self-ensembling consistency for knowledge transfer between domains.
• Implemented a self-ensembling adversarial learning module for enhanced feature alignment in UDA.
• Conducted experiments on cross-modality segmentation tasks involving MRI and CT images.

Main Results:

• The proposed LE-UDA framework demonstrated significant improvements in cross-domain segmentation performance.
• LE-UDA effectively leveraged limited source labels to overcome domain shifts and scarcity of annotations.
• The method outperformed state-of-the-art UDA approaches in the conducted experiments.

Conclusions:

• LE-UDA offers a viable solution for medical image segmentation in scenarios with limited labeled data and significant domain shifts.
• The framework shows promise for improving the generalizability of deep learning models across different medical imaging modalities.
• This work advances UDA techniques for practical clinical applications where data annotation is a bottleneck.