The application of generative networks in MR image reconstruction

Abstract

In this thesis, the application of generative networks in MR image reconstruction is demonstrated from the perspective of Euclidean loss to the perspective of likelihood loss. The method trained by Euclidean loss consists of a generator and a discriminator. The generator serves as the proximal operator in the ADMM loop, and the discriminator is to define an MR image space when training the generator. The other method utilizes the autoregressive network as a prior model. The reconstruction is modeled by Bayesian theorem and achieved by maximizing the posterior. These two proposed methods are all generalizable for difference reconstruction settings and reserve the conventional way to enforce data consistency instead of embedding that into the network, that’s to say, the learned component used in the reconstruction is separated from the process of k-space data generation. These two methods are tested with knee and brain MRI database, and it shows considerable improvement over the former reconstruction methods such as parallel imaging and compressed sensing.