Meeting Banner
Abstract #3792

Iterative Hard Thresholding and Matrix Shrinkage (IHT+MS) for Low-Rank Recovery of K-T Undersampled MRI Data

Mark Chiew1, Karla L. Miller1, Peter J. Koopmans2, Elizabeth M. Tunnicliffe3, Stephen M. Smith1, Thomas Blumensath4

1FMRIB Centre, University of Oxford, Oxford, United Kingdom; 2Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, Nijmegen, Netherlands; 3AVIC, Nuffield Department of Clinical Medicine, University of Oxford, Oxford, United Kingdom; 4ISVR, University of Southampton, Southampton, Hampshire, United Kingdom

In matrices that are low rank or approximately so, matrix completion strategies can be used to recover data in the presence of undersampling. Here we present a novel algorithm, iterative hard thresholding + matrix shrinkage (IHT+MS), for the recovery of low rank approximations to k-t undersampled MRI data. Performance of the IHT+MS algorithm is compared to other matrix completion techniques in retrospectively undersampled cardiac cine and fMRI data. Results indicate that good reconstruction fidelity is observed in both cardiac and fMRI data, even at high undersampling factors, and that IHT+MS produces the best results in many cases.

Abstract PDF Presentation Video

Keywords

ability affected although amount amounts amplitudes anal apparent appear approximate approximation assessed axis background basis best beyond brain called cardiac cine clinical cognition coherent coil coils comp companion components compressed consider continuation contrast core correlations courses decay degradation degrees depend describe despite dynamic effectively efficient emerged enhanced entries error evaluate even ever excellent expect factorization fast fidelity filtered fixed freedom full fully future gives graceful hard harm hence improves indistinguishable inspection instead interpolated irrelevant iterative kingdom least look magnitude mark match math matrix measured medicine miller novel occur operator optimal overall oxford particularly peak performance portion power preserve principal prior produce produced qualitatively quickly random rank rather real reasonably recent reconstructed reconstruction reconstructions recover recovery regularly remaining represents require required resting sampled sampling schematic selection selects sensing shrinkage shrunk singular sliding slow smith spatial specified spread start step successful suitable survive systole targeted temporal thresholding true useful visual visually window years