Meeting Banner
Abstract #0313

GPU Based Fast Inverse Gauss-Newton Motion Correction Method for High Throughput ofMRI

Zhongnan Fang1, Jin Hyung Lee1, 2

1Electrical Engineering, University of California, Los Angeles, Los Angeles, CA, United States; 2Department of Neurology and Neurological Sciences, Department of Bioengineering, Stanford University, Stanford, CA, United States

A novel GPU based parallel motion correction method for high throughput ofMRI study is presented. With ultra-high processing speed, this method enables real-time motion correction while allowing future integration of computationally intense processing steps including iterative reconstruction and automatic segmentation for high-throughput interactive brain circuit analysis. The algorithm utilizes an iterative inverse update strategy, which dramatically reduces computational cost. GPU specific features such as texture caching and hard-wired interpolation are also utilized for the highest efficiency. Compared to currently available methods, the proposed algorithm shows the lowest RMS error rate and highest speed.

Abstract PDF

Keywords

accuracy accurate acquisition activation additional almost although ample automatic available award baker beforehand bioengineering brain caching calculation career coherence combined comp complete compromising computation computationally computed conducted constant control correction cost critical crucial currently datasets design designed developed development director discovery distinct dramatically efficiency efficient electrical enable enables engineering error errors especially example excitation fang fast fastest features fellowship fixed forward foundation frame frames front function functional gauss gives grant great hard hardware highest highly human importance inhibition innovator integrated intense interpolation inverse iteration iterations iterative iteratively known largest least leaving lowest majority many marks matrix motion natl nature necessary neurological neurology newton numerous optimization optimized original parallel phantom precision process processes processing propose proposed quantitative real recent reconstruction reduce reduces registration resolution robustness room sciences scientific segmentation series sets slower smith sort source speed square step strategy studies sufficient system technology template temporal texture throughput towards traditional transform volume wired