Generalized PRUNO Kernel Selection by Using Singular Value Decomposition (SVD)

INTRODUCTION Parallel Reconstruction Using Null Operations (PRUNO) is an iterative k-space based reconstruction method for Cartesian parallel imaging. One particular challenge in PRUNO is to select a set of proper nulling kernels. A group of poor kernels will lead to an ill-posed system and affect the accuracy and convergence of the algorithm. The early reported kernel selection method is similar to GRAPPA, in which kernel templates are manually assigned by assuming unbiased symmetry among all coil sensitivity maps [1, 3]. In this work, we demonstrate an improved kernel selection strategy to create generalized PRUNO kernels from the Singular Value Decomposition (SVD) of calibration data. Furthermore, by introducing composite kernels prior to the conjugate-gradient (CG) reconstruction, the reconstruction time wouldn’t increase much when a large number of kernels are used. These new strategies boost the robustness of PRUNO with faster algorithm convergence and lower noise sensitivity.