C3R: Category contrastive adaptation and consistency regularization for cross-modality medical image segmentation

Unsupervised Domain Adaptation (UDA) has been widely used in cross-modality medical image segmentation, where the segmentation network is trained using both labeled images from the source domain and unlabeled images from the target domain. A prominent research direction in UDA involves learning domain invariant features through image translation between the source and target domain images, typically utilizing cycle consistency loss and/or adversarial loss. However, we argue that the auxiliary cycle consistency loss and adversarial loss may interfere with the main segmentation loss during training, despite its utility learning domain-invariant features. Furthermore, existing approaches in this research direction overlook the differentiation of various categories in domain-invariant feature learning. In this paper, we propose novel UDA method, named C3R, for medical image segmentation. C3R mainly comprises four components: shared encoder for learning domain-invariant features, a segmenter for generating segmentation output, and two decoders for image translation between the source and target domains. C3R fully explores consistency regularization in training, including image-level consistency, feature-level consistency, and segmentation output-level consistency. Moreover, C3R employs a detached training strategy to alleviate conflict between the main segmentation loss and auxiliary cycle consistency loss and adversarial loss. Last, C3R applies contrastive learning to pull together pixels of the same category, while pushing apart pixels of different categories, thereby enhancing the final segmentation results. Experimental results show that C3R outperforms other state-of-the-art methods by a considerable margin in Dice: 2.25% in cardiac substructure segmentation, 7.98% in brain tumor segmentation, and 1.8% in abdominal multi-organ segmentation.