Accelerated Cardiac Cine Using Locally Low Rank and Total Variation Constraints

INTRODUCTION: Simultaneously achieving high spatial and temporal resolution remains a challenge in 3D cardiac cine imaging. Constrained reconstruction promoting the low rank property of dynamic image matrices has been proposed for accelerated dynamic MRI [1,2]. Recently, locally low rank (LLR) constraint was proposed to exploit spatially-varied local rank-deficiency [3]. Moreover, the combination of global low rank with sparsity constraints has been shown in various forms to improve image recovery rate and reconstruction performance [1,4]. In this study, we combine LLR with temporal total variation constraints (LLR+tTV), and evaluate against current state-of-art methods on the reconstruction of highly undersampled cardiac cine images. METHODS: Datasets: Six fully-sampled cardiac cine datasets distributed by the 2013 ISMRM Recon Challenge committee [5] were used in this study. The data was acquired using a 2D cine breath-held bSSFP sequence with 32-channel cardiac receiver coils. Three of the datasets were midventricular short-axis, and the other three were vertical long-axis. Approximate parameters: image matrix 210×426, spatial resolution 1×1 mm, 30 timeframes per cardiac cycle. The datasets were retrospectively under-sampled in two dimensions using two sampling patterns: a variable-density random and Cartesian golden-angle radial. Acceleration factors (R) ranged from 10 to 50. Image Reconstruction: The cost function of the LLR and temporal TV regularized optimization problem was formulated as: