Collaborative Learning of Augmentation and Disentanglement for Semi-Supervised Domain Generalized Medical Image Segmentation

This paper explores a challenging yet realistic scenario: semi-supervised domain generalization (SSDG) that includes label scarcity and domain shift problems. We pinpoint that the limitations of previous SSDG methods lie in 1) neglecting the difference between domain shifts existing within a training dataset (intra-domain shift, IDS) and those occurring between training and testing datasets (cross-domain shift, CDS) and 2) overlooking the interplay between label scarcity and domain shifts, resulting in these methods merely stitching together semi-supervised learning (SSL) and domain generalization (DG) techniques. Considering these limitations, we propose a novel perspective to decompose SSDG into the combination of unsupervised domain adaptation (UDA) and DG problems. To this end, we design a causal augmentation and disentanglement framework (CausalAD) for semi-supervised domain generalized medical image segmentation. Concretely, CausalAD involves two collaborative processes: an augmentation process, which utilizes disentangled style factors to perform style augmentation for UDA, and a disentanglement process, which decouples domain-invariant (content) and domain-variant (noise and style) features for DG. Furthermore, we propose a proxy-based self-paced training strategy (ProSPT) to guide the training of CausalAD by gradually selecting unlabeled image pixels with high-quality pseudo labels in a self-paced training manner. Finally, we introduce a hierarchical structural causal model (HSCM) to explain the intuition and concept behind our method. Extensive experiments in the cross-sequence, cross-site, and cross-modality semi-supervised domain generalized medical image segmentation settings show the effectiveness of CausalAD and its superiority over the state-of-the-art. The code is available at https://github.com/Senyh/CausalAD.