Unsupervised Domain Adaptation for Cross-Modality Cardiac Image Segmentation Based on Contrastive Image Synthesis

Unsupervised domain adaptive(UDA) shows considerable performance in cross-modal medical image segmentation. The existing work usually uses the generation model to translate source domain images, then trains a segmentation model with the target domain images. However, some problems such as structural deviation are inevitable in the process of translation, and some asymmetric translations, even if very perfect, still cannot completely eliminate the differences between domains. In this paper, an unsupervised domain adaptive method is proposed for the cross-modal cardiac image segmentation, which includes a translation model and a segmentation model. In the translation model, we utilize the query attention module to consciously select the anchor points in contrast learning to minimize the structural bias in the translation process. Then, we employ the translated images and the target-domain images to train the segmentation model. Meanwhile, we leverage the global contrast learning and the output space adversarial learning to achieve domain adaptive. Different from the previous work, our global contrast learning method uses the spatial self-similarity mapping of the translated images and the target-domain images as input to capture their spatial relationship. The segmentation model trained in this way is better able to focus on domain-independent features. We evaluated our method using a cardiac dataset containing both CT and MRI modes. The experimental results have proven that our method is superior to all competing methods.