基于对比学习和特征解耦的医学图像泛化分割

omain shift refers to the distribution difference between the training data (source domain) and the data in practice (target domain), which may originate from different imaging devices, imaging parameters, patient populations, or medical institutions. Due to the domain shift, segmentation models trained on the source domain often perform poorly on the unseen target domain, which seriously limits the generalization ability and practical application value of the model. To solve the domain shift, a new domain generalization medical image segmentation method is proposed in this paper, which combines feature decoupling and contrast learning, so that the model can obtain more general domain invariant features in the training process and improve the segmentation performance of the unseen domain. In the network framework, the joint training of domain-specific normalized layers is adopted to train different normalized layers for different source domains, map features to the common space. Meanwhile, contrast learning is introduced to alleviate the inter-domain differences between different source domains. For the features generated at the encoder end, the domain-specific decoupling is carried out, the domain-specific channels are selected by the domain discriminator, and the domain-specific channels are suppressed. At the same time, the dual attention enhancement module is used to enhance the features from both channel and space perspectives to improve the model performance. The network model is compared with other methods on two common medical image domain generalization segmentation datasets, and the results show that the model is better than other models