Liver Segmentation via Learning Cross-Modality Content-Aware Representation

Liver segmentation has important clinical implications using computed tomography (CT) and magnetic resonance imaging (MRI). MRI has complementary characteristics to improve the accuracy of medical analysis tasks. Compared with MRI, CT images of the liver are more abundant and readily available. Ideally, it is promising to transfer learned knowledge from the CT images with labels to the target domain MR images by unsupervised domain adaptation. In this paper, we propose a novel framework, i.e. cross-modality content-aware representation (CMCAR), to alleviate domain shifts for cross-modality semantic segmentation. The proposed framework mainly consists of two modules in an end-to-end manner. One module is an image-to-image translation network based on the generative adversarial method and representation disentanglement. The other module is a mutual learning model to reduce further the semantic gap between synthesis images and real images. Our model is validated on two cross-modality semantic segmentation datasets. Experimental results demonstrate that the proposed model outperforms state-of-the-art methods.