Cross-Domain Style Augmentation for Semi-Supervised Domain-Generalized Medical Image Segmentation
Jun Wang, Liping Wang · 2024
Medical image segmentation models based on deep learning often experience performance decline when applied to unseen domains, which restricts the use of deep learning models in real-world medical settings. Most existing methods require the use of multiple fully labeled source domain datasets for training, which can be challenging due to the difficulty of obtaining a substantial volume of annotated medical datasets in practical scenarios. In this paper, we focus on semi-supervised domain-generalized medical image segmentation, aiming to achieve gen-eralization to unknown domains by training with limited labeled data from multi-source domain datasets. We propose a novel cross-domain style augmentation method, which augments image style by applying nonlinear transformation functions to low level frequency information from images in different domains and further merging information from different domains. In addition, we propose a confidence-weighted consistency method between teacher and student predictions to better guide consistency regularization learning within the teacher-student framework. Extensive experiments show that our method achieves improved generalization effectiveness on two public benchmarks.