Group-Sparsity Based Compressed Sensing Reconstruction for Fast Parallel MRI

Compressed sensing (CS) in parallel magnetic resonance imaging (pMRI) has the potential to reduce the MRI scan time by many folds. Due to the application of CS, conventional linear reconstruction techniques would not work. To reconstruct MR images from under-sampled measurements one needs to solve highly nonlinear optimization problems. Practical implementation of CS in pMRI involves reconstruction quality or accuracy and computational time as its major trade-offs. Since clinical multi-dimensional pMRI requires significant amount of raw data, sequential implementation of complex optimization algorithms would not meet the time constraints set for clinically feasible reconstructions. In this paper, we propose a CS based parallel MRI reconstruction method using wavelet forest sparsity and joint total variation as sparsity inducing regularization constraints. The model is applied to multi-slice multi-channel MRI data. Simulations are carried out in the hybrid CPU-GPU environment for reconstruction of complex valued multi-slice multi-coil MR data within a clinically feasible reconstruction time.