CUDA Papers

http://impact.crhc.illinois.edu/ftp/conference/xiaolong-2010.pdf

Abstract

Regularization is a common technique used toimprove image quality in inverse problems such as MR imagereconstruction. In this work, we extend our previous GraphicsProcessing Unit (GPU) implementation of MR imagereconstruction with compensation for susceptibility-induced fieldinhomogeneity effects by incorporating an additional quadraticregularization term. Regularization techniques commonly imposethe prior information that MR images are relatively smooth bypenalizing large changes in intensity between neighboring voxels.However, the associated computations often increase data accessand the overall computational load, which can lead to slowerimage reconstruction. This motivates us to adopt a GPU-enabledimplementation of spatial regularization using sparse matrices.This implementation enables the computations for the entirereconstruction procedure to be done on the GPU, which avoidsthe memory bandwidth bottlenecks associated with frequentcommunications between the GPU and CPU. Both the CPU andGPU code of this implementation will be available for release atthe time of the conference.

Authors

Yue Zhuo, Xiao-Long Wu, Justin Haldar, Wen-mei Hwu, Zhi-Pei Liang, Bradley P. Sutton, University of Illinois at Urbana-Champaign