Meeting Banner
Abstract #3838

A Generic, Multi-Node, Multi-GPU Reconstruction Framework for Online, Real-Time, Low-Latency MRI

Haris Saybasili1, Daniel A. Herzka2, Kestutis Barkauskas3, Nicole Seiberlich3, Mark A. Griswold1

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

In the recent years, many research oriented, customizable, external MR image reconstruction frameworks have been presented. To the best of our knowledge, none of these frameworks provided fully automated, remotely and locally distributed (multi-node, and multi-GPU) image reconstruction capabilities. Additionally, these frameworks may depend on high-level software libraries that make it difficult to maintain, debug and update the existing code. In this work, we present a highly customizable, automatically distributed, multi-threaded image reconstruction environment, built using only low-level system libraries for improved performance and portability. Our framework utilizes multiple GPUs and multiple workstations (nodes) by transparently distributing reconstruction tasks.

Abstract PDF Presentation Video

Keywords

acceleration accomplished acquisition adapt addition additionally affect approval assigned asynchronous audience automatically axis beneficial benefit biomedical breathing calibration capable cardiac chain challenging code coil coils command communicate communication communications complicates computational computer configuration configurations configure connections convolution cropping customizable datasets decreased deploy deploying detected developed development distribute distributed distributes distributing distribution dynamic enable engineering environment environments evaluate even executed executes execution external extra final flexible forwards framework frameworks freedom funded gated generic great ideas implemented improved informed inter language latency libraries lightweight local locally manager master matrix medicine memory network next node nodes object output partial people performance performances pipeline portability portions process programming protocol providing radial radiology readily real reconstructed reconstruction reduced remote required requirements reserve resolution scanner school separate service services short since software subset support system systems systolic table target task tasks temporal though thread threading threads toolkit tools totally transfers transparent transparently user various volunteer western workstation