A Novel 3D Unsupervised Domain Adaptation Framework for Cross-Modality Medical Image Segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Unsupervised domain adaptation (UDA) for cross-modality medical image segmentation faces challenges with 2D analysis and insufficient target domain alignment.
• Existing methods often miss depth-level semantic information and struggle with domain shift due to one-to-one style transfer.

Purpose of the Study:

• To develop a 3D unsupervised domain adaptation framework for cross-modality medical image segmentation.
• To address limitations of 2D analysis and improve target domain alignment through multi-style image translation.

Main Methods:

• Introduced a novel framework for volumetric (3D) unsupervised domain adaptation (UDA).
• Employed multi-style image translation for complete image alignment to mitigate domain shift.
• Incorporated a quartet self-attention module to enhance feature relationships in higher dimensions.

Main Results:

• Achieved substantial improvements in segmentation accuracy in unlabeled target domains.
• Demonstrated superior performance over state-of-the-art methods in challenging cross-modality tasks (brain and abdominal segmentation).

Conclusions:

• The proposed 3D UDA framework with multi-style translation and quartet self-attention offers a significant advancement in medical image segmentation.
• The model shows potential as a benchmark for biomedical and health informatics research, improving segmentation accuracy across modalities.