Disentangled representation and cross-modality image translation based unsupervised domain adaptation method for...

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation models struggle with cross-modality data due to domain shift.
• This performance degradation hinders real-world deployment of AI in healthcare.

Purpose of the Study:

• To develop an unsupervised domain adaptation framework for cross-modality medical image segmentation.
• To mitigate performance loss caused by domain shift between different imaging modalities.

Main Methods:

• Utilized a multimodal image translation framework decomposing latent space into content and style.
• Employed encoders to separate content and style codes, recombining them for cross-modality image generation.
• Implemented content and style reconstruction losses and content discriminators for domain alignment.

Main Results:

• Achieved significant performance improvements in bidirectional adaptation experiments on MRI and CT abdominal organ segmentation.
• Demonstrated a substantial increase in Dice Similarity Coefficient (DSC) by approximately 30% and 25%.
• Showed a notable reduction in Average Symmetric Surface Distance (ASSD) by 13.3 and 12.2.

Conclusions:

• The proposed unsupervised domain adaptation framework effectively enhances cross-modality segmentation performance.
• The method successfully preserves semantic information and anatomical structures in translated images.
• Outperformed several competing methods in cross-modality segmentation tasks.