1Institute of Medical Engineering, Graz University of Technology, Graz, Styria, Austria; 2Center for Biomedical Imaging, NYU Langone Medical Center, New York, NY, United States; 3Department of Mathematics and Scientific Computing, University of Graz, Graz, Styria, Austria; 4Center for Biomedical Imaging, New York University School of Medicine, New York, NY, United States
Iterative parallel-imaging methods are highly promising for MR image reconstruction from undersampled data due to their flexibility to incorporate a priori knowledge using regularization. However, these methods are computationally very expensive and memory demanding. Consequently, most implementations so far used acquisition schemes that allow separating the reconstruction into smaller sub-problems, e.g. by reconstructing 3D volumes slice by slice. This comes at the expense of loosing acceleration capability in this direction, which limits the achievable overall scan efficiency. Furthermore, for certain imaging techniques like 3D radial ultra-short TE (UTE) imaging, separation of the reconstruction is not feasible. In this work, we present a method that treats the whole 3D imaging volume as single data set. This enables completely arbitrary 3D trajectories with acceleration in any dimension and incorporation of fully 3D regularization functionals.