CiSeg: Unsupervised Cross-Modality Adaptation for 3D Medical Image Segmentation via Causal Intervention.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Unsupervised domain adaptation (UDA) transfers knowledge from labeled source domains to unlabeled target domains, crucial for medical imaging where labeled data is scarce.
• Current UDA methods struggle with generalization due to reliance on spurious correlations, limiting performance across different imaging modalities (e.g., CT to MRI).

Purpose of the Study:

• To develop a novel framework, the Causal Intervention Segmentation Network (CiSeg), that integrates causal inference into UDA to improve cross-domain generalization.
• To disentangle causal variables from bias variables using a Structural Causal Model (SCM) and a Counterfactual Disentanglement (CD) module.

Main Methods:

• Constructing an SCM for the source domain to identify and separate causal and bias variables.
• Implementing a CD module to decompose latent features into independent causal and bias components, mitigating spurious correlations.
• Utilizing Prototype-guided Contrastive Learning (PCL) for pixel-level alignment and Causal-bias Residual Alignment (CBRA) for feature-level invariance across domains.

Main Results:

• CiSeg demonstrated superior segmentation performance compared to state-of-the-art methods on cardiac, abdominal multi-organ, and BraTS18 datasets.
• The proposed framework achieved robust cross-domain generalization, effectively addressing the domain shift problem.
• Experiments confirmed the efficacy of causal inference in alleviating the impact of spurious correlations in UDA.

Conclusions:

• CiSeg offers a novel approach to UDA in medical image segmentation by incorporating causal inference.
• The framework successfully disentangles causal and bias factors, leading to improved generalization and segmentation accuracy across domains.
• The integration of PCL and CBRA further enhances cross-domain consistency and robustness.