FAUDA-Net: Frequency-aware unsupervised domain adaptation network for multimodal medical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) shows promise in medical image segmentation.
• Significant domain shifts between datasets remain a challenge for current UDA methods.
• Existing UDA approaches often use static alignment, leading to blurred boundaries and lost details.

Purpose of the Study:

• To develop a robust cross-domain solution for medical image segmentation.
• To address the limitations of static alignment in UDA.
• To enhance boundary delineation and segmentation accuracy in target domains.

Main Methods:

• Proposed Frequency-Aware Unsupervised Domain Adaptation Network (FAUDA-Net).
• Introduced dual-domain distribution disruption for domain-invariant representations.
• Utilized frequency constraints (phase and amplitude) for cross-domain adaptation.
• Employed frequency-aware contrastive learning with channel self-attention for feature alignment.

Main Results:

• FAUDA-Net dynamically adapts to feature distribution shifts.
• Enhanced boundary delineation and preserved structural details in target domains.
• Outperformed eight state-of-the-art methods on MMWHS17, BraTS21, and PROMISE12 datasets.
• Achieved superior performance in both overall segmentation accuracy (Dice) and boundary precision (ASD).

Conclusions:

• FAUDA-Net provides a reliable and effective solution for multi-modal and multi-center medical image segmentation.
• The frequency-aware approach improves robustness against domain shifts.
• The method offers significant improvements in segmentation accuracy and boundary delineation.