Abstract
Domain generalization is a critical challenge in medical image segmentation, where models trained on one domain often fail to generalize to unseen domains. Traditional approaches typically require multiple source domains, which may not always be available in clinical settings. In this work, we propose a causality-inspired approach for single-source domain generalization in medical image segmentation.
Our method leverages causal inference principles to identify and preserve domain-invariant features while learning domain-specific representations. The framework consists of three main components: (1) a causal feature extraction module that identifies stable causal relationships, (2) a domain-invariant learning module that preserves causal features across domains, and (3) a domain-specific adaptation module that handles domain-specific variations.
We evaluate our approach on multiple medical image segmentation datasets with domain shifts, including cross-scanner and cross-institution scenarios. Experimental results demonstrate that our causality-inspired approach significantly outperforms existing domain generalization methods, particularly in single-source scenarios. The method shows robust performance across different types of domain shifts and maintains segmentation accuracy in unseen domains.
The proposed framework represents a significant advancement in domain generalization for medical imaging, providing a more principled approach that could improve the deployment of segmentation models across different clinical settings.
BibTeX
@article{ouyang2021causality,
title={Causality-inspired Single-source Domain Generalization for Medical Image Segmentation},
author={Ouyang, Cheng and Chen, Chen and Li, Surui and Li, Zeju and Qin, Chen and Bai, Wenjia and Rueckert, Daniel},
year={2022},
journal={IEEE Transactions on Medical Imaging},
doi={10.1109/TMI.2022.3224067}
}