A Structure-Aware Framework of Unsupervised Cross-Modality Domain Adaptation via Frequency and Spatial Knowledge...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) is crucial for medical image segmentation, enabling models trained on one data domain to generalize to others.
• Existing UDA methods often rely on complex and inefficient adversarial learning to bridge domain gaps between medical image modalities.
• A simpler, more efficient approach is needed to address the challenges of cross-modality medical image segmentation.

Purpose of the Study:

• To develop a novel and efficient unsupervised domain adaptation method for cross-modality medical image segmentation.
• To reduce the complexity and inefficiency associated with traditional adversarial learning approaches in UDA.
• To enhance the performance of medical image segmentation models on unlabeled target domains.

Main Methods:

• Proposed a multi-teacher distillation framework incorporating both frequency and spatial domain transfer.
• Utilized non-subsampled contourlet transform to identify and adapt domain-invariant and domain-variant frequency components.
• Implemented a batch momentum update-based histogram matching for spatial domain style bias minimization and dual contrastive learning for structure-related information.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art techniques.
• Effective domain gap reduction was achieved through frequency and spatial domain transfer.
• Dual contrastive learning enhanced the model's ability to capture structure-related information.

Conclusions:

• The developed UDA method offers a simple yet effective solution for cross-modality medical image segmentation.
• The combined frequency and spatial domain transfer approach significantly improves adaptation efficiency and performance.
• This work provides a promising direction for advancing unsupervised domain adaptation in medical imaging.