Meeting Banner
Abstract #0185

Multi-Node, Multi-GPU Radial GRAPPA Reconstruction for Online, Real-Time, Low-Latency MRI

Haris Saybasili1, 2, Daniel A. Herzka3, Kestutis Barkauskas4, Nicole Seiberlich4, Mark A. Griswold1, 4

1Radiology, Case Western Reserve University, Cleveland, OH, United States; 2Biomedical Engineering, Johns Hopkins University School of Medicine, Baltimore, MD, United States; 3Biomedical Engineering, Johns Hopkins University, Baltimore, MD, United States; 4Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States

A hybrid (CPU- and GPU-based), faster-than-acquisition through-time radial GRAPPA reconstruction was previously demonstrated for 15 coil, rate 8 (16 projections, 128x128 matrix) radial datasets. However, because of the increased number of acquisition coils on modern scanners, single-GPU radial GRAPPA reconstructions were challenging for low-latency, real-time MRI with high number of acquisition coils. We present a completely automated, multi-node (group of workstations connected via network), multi-GPU radial GRAPPA implementation that can reconstruct 32-coil 16 projection radial datasets much faster than acquisition. Images from 32 coil, 16x256 data (acquisition time 42ms/frame) were reconstructed in 11.2 ms/frame using four nodes (two GPUs on each).

Abstract PDF Presentation Video

Keywords

accelerated acceleration accomplished acquisition adapt added approval array atlas audience automatically axis biomedical calculation calibration capability cardiac cardiovascular challenging coil coils combined command communications completely computational computer configurations conjugate connections convolution copy datasets dedicated developer diastolic display displayed distributed distributes distribution dual enable engineering excellent execution executions faster final full funding generalized graphical hardware healthy highly hybrid implementation implemented included informed inter johns larger latency library local mark master math matrix medicine modern module network node nodes operations parallel partial partially performance performances pipeline prior process processing product programming projections protocol radial radiology reached real received receiver reconstructed reconstructing reconstruction reconstructions repetitions requires reserve resolution resolutions scanner scanners scenarios school segment selective sense sensitivity sent sets short since sockets software spatial started stopped studies subsequently subset suggested synchronization systems systolic target task temporal thread threaded threads timer took toolkit trajectories trans transferred transfers transparent units user utilizing varying volunteers western workstation