Abdomen MRI Synthesis Based on Conditional GAN

Obtaining different contrast images of the same anatomical location of the same subject can increase the diversity of diagnostic information. However, due to the cost constraints, it is often not possible to obtain sufficient different contrast images, and images in some contrasts may be corrupted by noise and artifacts. In this case, if these images with different contrasts can be generated from existing images, the diagnostic accuracy will be greatly improved. Here, we explore an end-to-end deep convolutional neural networks for multi-contrast MRI synthesis based on conditional generative adversarial networks (cGAN). The proposed approach preserves high-frequency details via an adversarial loss and a pixel-wise loss, to transform abdomen CT into abdomen MR. Paired MR-CT volumes of 10 patients are analyzed. A quantitative evaluation showed that the approach is capable to synthesize MR images that closely similar to reference MR images.