Multi-category domain-dependent feature-based medical image translation

The challenge of inaccurate information containment in synthetic images during the process of cross-domain medical image translation could be resolved by using a common strategy of integrating the loss of the feature consistency of the real/synthetic image as a penalty factor into the loss function of the translator. However, the existing methods are capable of using only the "domain-independent" feature of the image when the aligned images are scarcity, which results in the under-utilization of the image information. In the present study, a novel feature consistency loss computing and integration method based on the "domain-dependent" features was proposed, and a multi-category feature consistency-cross-domain image translation (MFC-CIT) model was constructed. The present study is the first to utilize the image feature information related to image domain in the process of cross-domain medical image translation. In the proposed method, the MFC module was first trained on the basis of supervised learning on a limited number of paired real images. Next, cross-domain image translation training based on unsupervised learning was performed on unpaired datasets by the CIT module, and this process was constrained by the loss of feature consistency of the real/synthetic image obtained in the MFC module. The experimental results on two datasets demonstrate that the proposed method effectively improves the translation accuracy of synthetic images.