Mind the Gap: Alleviating Local Imbalance for Unsupervised Cross-Modality Medical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised cross-modality medical image adaptation addresses domain gaps without target labels.
• Existing methods often focus on global alignment, neglecting local feature imbalances.
• Local alignment methods risk losing contextual information.

Purpose of the Study:

• To propose a novel strategy, Global-Local Union Alignment, for imbalanced domain gap adaptation in medical images.
• To improve the accuracy of cross-modality medical image segmentation.
• To tackle the limitations of existing global and local alignment techniques.

Main Methods:

• A feature-disentanglement style-transfer module synthesizes target-like source images to reduce global domain gaps.
• A local feature mask prioritizes discriminative features with larger domain gaps, reducing local imbalances.
• Combines global and local alignment for precise localization and semantic consistency.

Main Results:

• Achieved state-of-the-art performance on cardiac substructure segmentation.
• Demonstrated superior results in abdominal multi-organ segmentation.
• Validated the effectiveness of the proposed Global-Local Union Alignment strategy.

Conclusions:

• Global-Local Union Alignment effectively alleviates imbalanced domain gaps in cross-modality medical image adaptation.
• The proposed method enhances segmentation accuracy by precisely localizing crucial regions while preserving semantic consistency.
• This approach offers a significant advancement for unsupervised medical image segmentation tasks.