Meeting Banner
Abstract #3815

From Matrix to Tensor: Compressed Sensing Dynamic MRI Using Tensor Based Sparsity

Yeyang Yu1, Jin Jin2, Feng Liu2, Stuart Crozier2, Mingjian Hong3

1The School of Information Technology and Electrical and Engineering, The University of Queensland, Brisbane, QLD, Australia; 2The School of Information Technology and Electrical and Engineering, The University of Queensland, Brisbane, QLD, Australia; 3School of Software Engineering, Chongqing Univeristy, Chongqing, China

In this work, we introduce the conception of tensor sparsity for Compressed Sensing dynamic MRI. Conventionally, the spatial and temporal information were then sparsified independently and sequentially. Therefore the spatial-temporal correlation may not be sufficiently exploited. This work applys the Tucker model based Higher-order Singular Value Decomposition (HOSVD) in the Compressed Sensing dynamic MRI framework. Instead of treating the 3D/4D data as series of 2D images, HOSVD inherits the high-dimensional data format, leading to significantly improved dMRI reconstructions compared with those well-established CS-dMRI methods. The advantages of the tensor sparsity in terms of reconstruction accuracy have been demonstrated in a given cardiac dynamic MRI study.

Abstract PDF Presentation Video

Keywords

ability accelerate accuracy achieved acquisition advantages anal application applied approaches authors basis better briefly cardiac china cine clearly close coefficients combination complex compressed computation computed concentrated concept conception consistent consistently consists construct contrast conventionally core corner correlated correlation correlations datasets decomposed decomposition degree denote denotes dimension dimensional dimensions dynamic electrical employs engineering enlarged equation error evaluation example exists exploited exploiting explore exploring extra fact fashion finding format frame frames framework free full highest highly hong improved independently ineffective inherits inspired instead introduce leading leads lowest maintains majority maps matrices matrix minimize mode model multidimensional normalized novel offering operations original partial precession product proposed quantified quantifies randomly recent reconstruction reconstructions recovery reduces reduction redundancy representation resolution respectively retrospective school sensing separated sequentially series significantly simultaneously singular slice software sourced space sparse sparsity spatial steady subsequent sufficiently summarized system technology temporal tensor terms title tolerance transform transforms treating tucker unfolding vast vector wavelet