MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Robust medical image segmentation is essential for quantitative analysis but manual annotation is costly and time-consuming.
• Unsupervised domain adaptation (UDA) addresses label scarcity by transferring knowledge from labeled to unlabeled domains.
• Existing UDA methods face challenges with heterogeneous and volumetric medical data and diverse domain shift types.

Purpose of the Study:

• To introduce Masked Autoencoding and Pseudo-Labeling Segmentation (MAPSeg), a versatile and high-performing UDA framework for medical image segmentation.
• To systematically address four distinct domain shift challenges in medical image segmentation.
• To develop a framework applicable to centralized, federated, and test-time UDA.

Main Methods:

• Developed MAPSeg, a unified framework integrating masked autoencoding and pseudo-labeling for segmentation.
• Systematically evaluated MAPSeg across four different medical image domain shift scenarios.
• Compared MAPSeg against state-of-the-art methods on both private infant brain MRI and public cardiac CT-MRI datasets.

Main Results:

• MAPSeg demonstrated superior performance, achieving significant Dice coefficient improvements (10.5 on MRI, 5.7 on CT-MRI) over existing methods.
• The framework maintained comparable performance across centralized, federated, and test-time UDA settings.
• MAPSeg effectively tackled heterogeneous and volumetric medical image segmentation challenges.

Conclusions:

• MAPSeg offers a versatile and effective solution for unsupervised domain adaptation in medical image segmentation.
• The framework shows significant practical value for real-world applications requiring robust segmentation with limited labels.
• MAPSeg represents a significant advancement in addressing domain shifts in medical imaging analysis.