Meeting Banner
Abstract #2645

Memory-Saving Iterative Reconstruction on Overlapping Blocks of K-Space

Martin Uecker1, Michael Lustig1

1Electrical Engineering and Computer Sciences, University of California, Berkeley, Berkeley, CA, United States

The purpose of this work to develop memory-efficient implementations of iterative reconstruction algorithms, such as SENSE, nonlinear inversion, and ESPIRiT. Although iterative reconstruction provides faster imaging and improved image quality, high computational demand currently limits its application for large 3D data sets. Graphical processing units can reduce computation time, but do not have enough memory for conventional algorithms. The proposed technique exploits the typical structure of iterative reconstruction algorithms to divide the computation into small independent blocks to reduce memory consumption.

Abstract PDF Poster PDF

Keywords

accumulation achieving adjoint advances alley allows although amount apart application applications applying arbitrary arrays audience avoided block blocks board boundary caches channel channels clinically coefficients coil combine combining computation computational computer consistency content convolution count decoupling demand develop dimensions discrete domain edition efficient eigenvalue electrical encoding equivalent estimation even example exploit extended extraction fast full goal graphical head highly implementation implementations improved independent independently invalid inversion iterative joint just language library limits made make makes making maps mask matrix memory much multiplication multiplying needed nonlinear opens operates operations opportunities overlap overlapping padding parallel parallelization part place practical preferable prerequisite processing progress proposed receivers reconstruction reconstructions reduce reduced reduction regularized requires requiring resolution retrospectively reversing runtime sampling save saving savings scalable sciences section sense sensing sensitivities sensitivity series sets space special spirit store stored storing strategies structure submitted table target temp trajectories trans treatment true typical typically units update various virtue zero