Fast Regularized Reconstruction for PatLoc MR Imaging using Total Generalized Variation and GPUs

PatLoc MR Imaging uses two nonlinear, nonbijective encoding fields in addition to the conventional three linear gradients for image encoding. This leads to new ways to perform image encoding. However, iterative reconstruction of PatLoc data is a computationally challenging task due to the fact that Fourier encoding does not apply anymore. This work aims at implementing a GPU-accelerated reconstruction framework for PatLoc MR imaging based on two discretization schemes of the forward model. Further, TGV regularization is performed to improve image quality. To improve convergence, a new method for numerically solving the TGV method is proposed, called TGV-CG. The reconstruction is evaluated on in-vivo and phantom data acquired with the PatLoc hardware. It is shown that GPU-acceleration leads to significantly improved performance which renders the investigated methods practical. In addition, TGV improves image quality even for undersampled data. TGV-CG leads to faster convergence in some cases which further decreases reconstruction time.