Eliminating domain-related confounding factors in cross-domain one-shot medical image segmentation via causal...

Cross-domain few-shot medical image segmentation (CD-FSMIS) is important in clinical practice, especially when annotations are limited and domain differences are large. However, existing methods often fail to handle two main challenges at the same time: domain shifts caused by scanner differences and protocol variations, and the lack of labeled data. From a causal inference perspective, these problems make the model rely more on spurious correlations related to domain-specific factors, which harms its ability to generalize to new domains. To solve this, we propose a method called Domain Feature Correction Network (DFCN). It simulates causal interventions to learn domain-invariant anatomical features by blocking non-causal confounding pathways. DFCN includes three main modules: (1) Multi-Band Perturbation Rectification Module (MPRM), which adjusts cross-domain features using frequency-based perturbation and correction. (2) Domain Prompt Prototype Contrastive Learning (DPPCL), which uses learnable domain prompts and contrastive learning to improve prototype generalization. (3) Prototype Relationship Optimization (PRO), which models hidden relationships between regions to refine foreground prototypes and improve consistency. These modules help the model learn stable, domain-invariant features and improve performance on unseen domains. Experiments on several CD-FSMIS benchmarks show that DFCN performs better than existing methods like IFA and FAMNet in both segmentation accuracy and cross-domain robustness. Code is available at https://github.com/394481125/DFCN.