GPU-Accelerated Radial Image Reconstruction with an Improved Parallel Gridding Method

Introduction: Radial acquisition is becoming increasingly popular in MRI due in part to its reduced sensitivity to motion and its capability for high temporal resolution imaging via azimuthal undersampling. In some applications such as DCE-MRI, a continuous scan on the order of ~10 min is typically performed and yields a large dynamic series data set. For example, in a typical 3D DCE-MRI experiment with 32 slices, five channels, 6000 views and 384 readout samples, the size of acquired data set is ~3.0GB. It takes ~30 min on a modern desktop PC to reconstruct this large volume data set. Recently, graphic processing unit (GPU) has been utilized to accelerate the radial image reconstruction. One of the challenges of gridding using massively parallel processing is the issue of synchronization, in which multiple acquired points are gridded onto the same Cartesian grid location. Several approaches were reported to address the synchronization problem in GPU-accelerated gridding reconstruction [1-3]. In this work, a simple and effective parallel gridding algorithm is proposed and tested on two GPU systems for the reconstruction of DCE-MRI dynamic series.