Abstract
Domain generalization aims to generalize models trained on source domains to unseen target domains, which usually have domain shift from the source domains. This challenging setting is common in clinical scenarios due to differences in scanning protocols, device manufacturers, and image modalities. In this paper, a novel style augmentation framework is proposed to improve out-of-domain performance for the model in the context of medical image segmentation. The proposed framework first investigates the style distribution between samples, for which a style clustering-based resampling is developed to generate evenly diversified styles, and then leverages a cross-structure style augmentation, which generates new styles by shuffling regional style information among structure within one sample. This structure style shuffling can greatly expand the style space while well respecting the anatomical structure of medical images. The method is validated through extensive experiments on 2 datasets, i.e., cross-modality abdominal multi-organ dataset, and cross-site cardiac dataset, with 2 settings, i.e., single-domain and multi-domain generalization. Experimental results show that our method yields significant improvement and outperforms state-of-the-art methods with significant margins.