Quantitative comparison of GRAPPA and RAKI simultaneous multi-slice reconstruction algorithms

The aims of this study were to (1) quantitatively compare various k-space interpolation-based simultaneous multislice (SMS) reconstruction algorithms, including linear GRAPPA and nonlinear, data-driven RAKI methods, and (2) provide an open-source toolbox for GPU-accelerated SMS reconstructions. For a single phantom, fully sampled k-space reference data and SMS-accelerated, R_SMS, data were collected for R_SMS factors of 2 through 5, with in-plane acceleration, R_ip, added retrospectively. Slice-GRAPPA, split-slice-GRAPPA, readout-SENSE-GRAPPA, and their hyperparameter-tuned analogous deep learning-based RAKI reconstructions were performed at different combinations of R_SMS and R_ip factors. Performance of the reconstruction methods were compared quantitatively by testing for significant differences in whole-image SSIM and regions of interest-based measurements of coefficient of variation (CV). The number of epochs, hidden layer size, and the interaction between the two were statistically significant factors for a Type II ANOVA test with SSIM performance (p=0.00005, p=0.00687, and p=0.000281, respectively) for RAKI reconstructions with post hoc tests suggesting that 500 epochs, 3 hidden layers, and 128 neurons per layer strike an ideal balance of speed and performance for this dataset. CV testing through Kruskal-Wallis tests did not yield any significant differences between reconstruction algorithms at any given R_SMS×R_ip. Limited testing indicated that a RAKI method generally resulted in larger SSIM values at any net acceleration factor, (R_net=R_SMS×R_ip), and all RAKI methods had a larger SSIM value than their respective GRAPPA counterparts at R_net>4, with the exception of ROSR at R_SMS=5 and R_net=10, and SPSR at R_SMS=2,R_ip=5. The difference between GRAPPA and RAKI was typically large enough (ΔSSIM of >0.02) to be obvious at R_net≥8. Slice-GRAPPA, slice-RAKI, or readout-SENSE-GRAPPA performed well for net accelerations R_net≤6, and readout SENSE-RAKI or split-slice-RAKI performed well for R_net≥8. Caution should be taken when using these generalizations as observations were dataset-specific and collected on a single vendor, coil array, pulse sequence, and phantom. They may, however, serve as a useful starting point for hyperparameter tuning. More experiments need to be conducted before these results can be translated into in vivo observations.