Meeting Banner
Abstract #2554

Low-Latency Radial GRAPPA Reconstruction Using Multi-Core CPUs and General Purpose GPU Programming

Haris Saybasili1, 2, Daniel A. Herzka2, Nicole Seiberlich3, Mark Griswold1, 3

1Department of Radiology, Case Western Reserve University, Cleveland, OH, United States; 2Biomedical Engineering, Johns Hopkins University School of Medicine, Baltimore, MD, United States; 3Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States

Real-time imaging using non-Cartesian trajectories permits very high acceleration rates with parallel imaging. However, image reconstruction from undersampled non-Cartesian datasets is computationally demanding, and leads to long reconstruction times .We present a hybrid (CPU- and GPU-based), fully auto-calibrated, fast reconstruction implementation for radial GRAPPA that uses RT-GROG to grid radial data prior to FFT. Radial GRAPPA and RT-GROG calibrations were performed on the CPU, while image reconstruction was performed asynchronously on the GPU. Our implementation was tested on healthy volunteer cardiac data for different parameters. Images from 12 coil, 144x256, R=9 data were reconstructed in 33ms/frame (acquisition time 42ms/frame).

Abstract PDF Poster PDF

Keywords

acceleration acceptable accuracy achieve achieved acquisition acquisitions application approval assumed asynchronous atlas axis better bigger biomedical black blue boxes broken calculation calculations calibrated calibration cardiac chose coil coils combination components computational computationally computed conditioning conjugate consuming continuous dataset datasets demanding depend depiction deployment developer diastolic driven either employed encoding engineering ensure estimation example execution fast frames full fully gating general generalized good gradient graphical grid grog guided hardware healthy hence implementation implemented improve includes intensive involve iterations latency leading library load long mark matching matrices matrix measured motivation negligible operations operator oversampling parallel partially penalties performance performances permit permitting pipeline prior procedures processes processing programming projections quality radial radiology readout real reconstruction reconstructions repetitions required reserve resolution resource robust sampled samples scanner schematic school sense sensitivity separately short significantly smaller software space system systolic take temporal thirty though threaded trajectories trans transfers transition unacceptable units unlike utilization varying western written yield