Deep Symmetric Adaptation Network for Cross-Modality Medical Image Segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Unsupervised domain adaptation (UDA) is crucial for cross-modality medical image segmentation.
• Existing UDA methods often use asymmetric translation networks, which struggle with large domain shifts.
• This limitation hinders effective domain gap elimination.

Purpose of the Study:

• To propose a novel deep symmetric architecture for UDA in medical image segmentation.
• To address the limitations of asymmetric structures in mitigating domain discrepancies.
• To enhance segmentation performance in cross-modality medical imaging.

Main Methods:

• Introduced a deep symmetric architecture with segmentation and two symmetric domain translation sub-networks.
• Implemented a bidirectional alignment scheme using a shared encoder and private decoders for feature alignment.
• Trained a pixel-level classifier using original and translated images from both source and target domains.

Main Results:

• The proposed symmetric architecture effectively mitigates the domain gap between source and target domains.
• Bidirectional feature alignment enhances the model's ability to handle domain discrepancies.
• The method demonstrated significant advantages over state-of-the-art methods in cardiac, BraTS, and abdominal multi-organ segmentation tasks.

Conclusions:

• The novel deep symmetric UDA architecture offers a robust solution for cross-modality medical image segmentation.
• Bidirectional alignment and comprehensive training strategies effectively leverage semantic information across domains.
• This approach shows strong potential for improving medical image segmentation accuracy in diverse clinical applications.