Primal-dual net for multi-contrast CS-MRI reconstruction

Abstract

MRI is a tomography technique. This technique mainly obtains the corresponding electromagnetic signals from the human body through the magnetic resonance phenomenon, and then reconstructs the information contained in the human body through certain technical means. But the long scanning time is an important factor restricting the application of magnetic resonance imaging. Long-time scanning may cause motion artifacts in the image. In the field of magnetic resonance imaging reconstruction, compressed sensing theory can reduce the amount of data collected in k-space and reduce the scanning time, thereby achieving the purpose of accelerating imaging. Through various technical methods, we can recover high-quality medical images from these under-sampled data for medical diagnosis. To this end, this paper proposes a variety of neural network methods for under-sampled magnetic resonance image reconstruction, aiming to obtain as high-quality reconstructed images as possible. And this thesis applies the primal-dual algorithm to the multi-contrast MRI reconstruction, it combines the theoretical convergence guarantee with the powerful deep neural network.