Multi-phase Liver-Specific DCE-MRI Translation via A Registration-Guided GAN

In the diagnosis of liver lesions, Gd-EOB-DTPA-enhanced magnetic resonance imaging (MRI) at the hepatobiliary phase (GED-HBP) is particularly valuable. However, the acquisition of GED-HBP is more costly than that of a conventional dynamic contrast-enhanced MRI (DCE-MRI). This paper introduces a new dataset and a novel application of image translation from multi-phase DCE-MRIs into a virtual GED-HBP image (v-HBP) that could be used as a substitute for GED-HBP in clinical liver diagnosis. This is achieved by a generative adversarial network (GAN) with an auxiliary registration network, referred to as MrGAN. MrGAN bypasses the challenges from intra-sequence misalignments as well as inter-sequence misalignments. Additionally, MrGAN incorporates a pre-trained shape consistency network to promote local generation in the liver region. Extensive experiments demonstrated the superiority of our MrGAN over other state-of-the-art methods in terms of quantitative, qualitative, and clinical evaluations. We outlook the utility of our new dataset will extend to other problems beyond lesion detection due to the improved quality of the generated image. Code can be found at https://github.com/Jy-stdio/MrGAN.git .