A quantitative survey of GRAPPA reconstruction in parallel MRI: impact on noise reduction and aliasing

In recent years, several attempts have been made to improve the quality of GRAPPA reconstruction by concomitant reduction of noise amplification and aliasing. While it is true that accuracy of the underlying GRAPPA model depends on the goodness of fit, variations in input noise and channel encoding can serve to introduce differences in reconstruction. In this article, we quantitatively compare the performance of different GRAPPA types, in relation to input signal-to-noise ratio, channel number, and acceleration factor. The GRAPPA types considered in this article are obtained using: (1) calibration matrix modification, (2) regularization of calibration equation, (3) updating of coil coefficients, (4) intermediate domain GRAPPA, and (5) iterative method. All the above GRAPPA types are tested on retrospectively sampled in vivo datasets acquired using head, spine and knee-array coils. For each GRAPPA type and dataset analyzed, reconstruction is performed using number of Autocalibration Signal lines set at 10% of the number of phase-encodes used in full acquisition and kernel size optimized for minimum root-mean square-error (RMSE). Each GRAPPA type is rated based on their individual performance, in relation to a mean performance measure obtained by averaging across different methods. Noise amplification in each GRAPPA type is visually examined, and compared with the respective g-factor map. RMSEs and g-factors of in vivo data are compared against those of simulated coil images with added complex Gaussian noise having a pre-determined covariance structure. A summary of performance ratings and execution times are provided for comparison with standard GRAPPA.