While undersampled MRI data is easy to obtain, lack of high-quality labels for dynamic organs impedes the common supervised training of deep neural nets for MRI reconstruction. We propose an unpaired training super-resolution model with pure GAN loss to use a minimal amount of labels but all available low-quality data for training. Leveraging Wasserstein-GANs with gradient penalty followed by a data-consistency refinement high-quality Knee MR images are recovered from 3-fold undersampled single coil measurements using 20% of the labels compared with a paired training model.
This abstract and the presentation materials are available to members only; a login is required.