Structure preservation constraints for unsupervised domain adaptation intracranial vessel segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unsupervised domain adaptation (UDA) methods reduce data annotation needs but struggle with intracranial vessel segmentation due to structure mismatch during image synthesis.
• Existing UDA segmentation methods show performance degradation in fine intracranial vessel segmentation tasks.

Purpose of the Study:

• To propose a novel UDA segmentation method, StruP-Net, that improves image synthesis quality and segmentation performance by incorporating structure preservation approaches.
• To address the challenge of structure mismatch in UDA for intracranial vessel segmentation.

Main Methods:

• StruP-Net utilizes adversarial learning for image synthesis and two domain-specific segmentation networks for semantic consistency.
• Employs feature-level structure preservation (F-SP) using graph convolutional networks (GCNs) for structural similarity.
• Implements image-level structure preservation (I-SP) based on perceptual loss for structure similarity constraints.

Main Results:

• StruP-Net achieved superior segmentation performance compared to state-of-the-art methods in cross-modality experiments (MRA to CTA).
• The method demonstrated high inference efficiency, indicating potential for clinical applications.

Conclusions:

• StruP-Net effectively enhances image synthesis quality and segmentation performance in UDA for intracranial vessel segmentation.
• The proposed structure preservation techniques successfully alleviate structure mismatch issues, offering a promising solution for medical image analysis.