Causality-Inspired Single-Source Domain Generalization for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning models often fail to generalize to new domains due to domain shift.
• This is a significant challenge in medical imaging where data acquisition varies.
• Single-source domain generalization aims to train models robust to unseen domains using data from only one source.

Purpose of the Study:

• To develop a data augmentation technique for single-source domain generalization in cross-domain medical image segmentation.
• To improve the robustness of deep learning models against domain shifts caused by different acquisition processes.

Main Methods:

• A causality-inspired data augmentation approach using randomly-weighted shallow networks to introduce appearance transformations.
• Causal intervention by resampling object appearances independently to remove spurious correlations.
• Validation on cross-modality (CT-MRI), cross-sequence (cardiac MRI), and cross-site (prostate MRI) segmentation tasks.

Main Results:

• The proposed method consistently improved segmentation performance on unseen domains across all tested scenarios.
• The approach demonstrated superior robustness compared to existing methods.
• Synthesized domain-shifted examples effectively exposed the model to variations.

Conclusions:

• The causality-inspired data augmentation is effective for single-source domain generalization in medical image segmentation.
• Removing spurious correlations enhances model robustness to domain shifts.
• The method offers a promising solution for deploying deep learning models in diverse clinical settings.