Segmentation-synthesis co-training for semi-supervised domain generalizable medical image segmentation

Semi-supervised domain generalization (SSDG) faces two fundamental challenges that hinder model generalizability: label scarcity and domain shifts. Recent studies tackle this challenging task by building on a scheme that integrates the strong-weak pseudo supervision paradigm with specific data augmentation strategies. However, semantic inconsistencies between style-augmented unlabeled images and their pseudo labels limit the effectiveness of this scheme and impair the generalizability of trained models. One critical question arises: How to ensure semantic consistency and style diversity of unlabeled-image and pseudo-label pairs for training a well-generalized model? To this end, we introduce ReMatch, a segmentation-synthesis co-training framework for semi-supervised domain generalization in medical image segmentation. The core of ReMatch lies in the SynTS algorithm that Synthesizes unlabeled images with both high semantic consistency and style diversity by leveraging Texture and Shape features derived from the segmentation process. Extensive experiments on single-source single-target and single-source multi-target cross-domain settings with various image modalities demonstrate that ReMatch offers an effective solution for SSDG, achieving compelling performance. For example, compared with the state-of-the-art based on the aforementioned scheme, ReMatch achieves average improvements of 2.31% and 1.68% in Dice similarity coefficients under the two cross-domain settings with 10% labeled data, respectively. Code is available at https://github.com/Senyh/ReMatch.