Deep Regularization For Scale-Agnostic Superresolution of MR Images

Magnetic Resonance Imaging (MRI) is the preferred approach for soft-tissue imaging due to its good contrast and non-invasiveness. While traditional MRI yields high-quality images, low-field scanners, though affordable and portable, produce lower-resolution images due to time and hardware constraints. The resolution can be enhanced using deep learning; however, the end-to-end nature of such models necessitates retraining the network with changes in the measurement parameters, such as the downsampling factor. Moreover, conventional superresolution (SR) models are not ideally suited for MRI, where the measurements are obtained in the k space. To address these challenges, we propose to decouple the forward model from the network using the Plug-and-Play framework. Specifically, we use a trained denoiser for scale-agnostic regularization, i.e., the input to the network is independent of the downsampling factor. Consequently, our method can be used with different resolution scanners, which is not possible with end-to-end networks. The innovation of our approach is that (i) we use a loss function derived from an MRspecific forward model, and (ii) instead of a standard off-the-shelf Gaussian denoiser, we train a U-Net denoiser to remove “artifacts” from the intermediate reconstructions. Our method achieves stateof-the-art reconstructions and is robust to acquisition and resolution settings.