UniAda: Domain Unifying and Adapting Network for Generalizable Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Generalizable medical image segmentation is challenging due to domain discrepancies (e.g., vendors, protocols).
• Existing domain generalization (DG) methods struggle to capture global domain characteristics during training or adapt to unseen domains during testing.

Purpose of the Study:

• To propose a novel "unifying while training, adapting while testing" paradigm for generalizable medical image segmentation.
• To develop a domain-aware base model that dynamically adapts to unseen target domains.

Main Methods:

• Introduced a domain Unifying and Adapting network (UniAda).
• Employed a feature statistics update mechanism to unify multi-source domains into a global inter-source domain.
• Utilized an uncertainty map to guide model adaptation to specific testing samples, even those outside the global inter-source domain.

Main Results:

• UniAda demonstrated strong generalization capacity on public and in-house cross-domain medical datasets.
• The proposed method outperformed state-of-the-art domain generalization techniques.
• The approach effectively addresses challenges in learning generalizable segmentation models for diverse medical imaging data.

Conclusions:

• UniAda offers a robust solution for cross-domain medical image segmentation.
• The "unifying while training, adapting while testing" strategy enhances model adaptability and performance on unseen domains.
• The method holds significant potential for improving the reliability of AI in medical imaging across different clinical settings.