Meeting Banner
Abstract #2553

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

Jiangsheng Yu1, Yiqun Xue1, Xia Zhao1, Ping Wang1, Walter Witschey1, Hee Kwon Song1

1University of Pennsylvania, Philadelphia, PA, United States

Graphic processing units (GPU) have recently been utilized to accelerate radial MR image reconstruction. One of the challenges of gridding using GPU is the issue of synchronization, and several approaches have been reported to address this issue. This work presents a simple and effective solution for avoiding this synchronization problem by utilizing a divide-and-conquer strategy that ensures that the gridded points are non-overlapping. Performance tests on the reconstruction of a DCE-MRI data set demonstrates that an acceleration factor of 67 can be achieved on an Nvida Tesla C2050 GPU system.

Keywords

accelerated acceleration access according accumulation achieved acknowledgments acquisition address advantages although applied approaches assigned assigns atomic available avoids azimuthal becoming binning cancer capability capacity challenges channel channels circles collecting commonly complexity computation computer computing condition conference conquer consuming continuous contributions convolution core cores dataset density desktop distance distributing divide divided donor donors driven dynamic effective example experiment final five frame frontiers full generate graphic grid gridded highly illustrated improved increasingly indexed interleave interleaved involve issue kernel laptop location massively medical memory modern motion neighboring neighbors noted occur occurs offsets overhead overlapped overlapping oversampled parallel parallelism part particularly performance ping popular problem proceedings process processing proposed race radial readout recently reconstruct reconstruction reduced reported represent resolution respectively samples sampling scenario searches sensitivity separate series several simple slice slices society solid sorting space synchronization system systems table takes temporal thread threads transactions typical typically uniformly unit utilize utilized view views volume whose yields