A Comparative Study of Unsupervised Deep Learning Methods in MRI Reconstruction
Recently, unsupervised deep learning methods have shown great potential in image processing. Compared with the large-scale demand on paired training data of supervised methods in specific tasks, unsupervised methods learn a universal and explicit prior information on the data distribution, and integrate it into the reconstruction process. Therefore, it can be used in various image reconstruction environments at the meantime of not degrading its performance. The importance of unsupervised learning in MRI reconstruction appears to be growing. Nevertheless, the construction of prior formulation in unsupervised deep learning varies a lot relying on mathematical approximation and network architectures.
It should be emphasized that only some of the competing algorithms are available for MRI reconstruction. We modified them to be suitable for reconstruction. Particularly, the code of DLMRI is originally from http://www.ifp.illinois.edu/~yoram/DLMRI-Lab/DLMRI.html. DC-CNN is originally from https://github.com/js3611/Deep-MRI-Reconstruction. DDP is originally from https://github.com/kctezcan/ddp_recon. PixelCNN is originally from https://github.com/openai/pixel-cnn and Glow is originally from https://github.com/openai/glow.