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

Area of Science:

• Medical Imaging
• Image Reconstruction
• Magnetic Resonance Imaging (MRI)

Background:

• Simultaneous multislice (SMS) imaging accelerates MRI acquisition by exciting and acquiring data from multiple slices concurrently.
• Traditional reconstruction methods like GRAPPA have limitations, especially at higher acceleration factors.
• Data-driven, nonlinear methods like RAKI offer potential for improved image quality in SMS imaging.

Purpose of the Study:

• To quantitatively compare various k-space interpolation-based SMS reconstruction algorithms.
• To evaluate both linear GRAPPA and nonlinear, data-driven RAKI methods.
• To provide an open-source toolbox for GPU-accelerated SMS reconstructions.

Main Methods:

• Comparison of Slice-GRAPPA, split-slice-GRAPPA, readout-SENSE-GRAPPA, and RAKI reconstructions.
• Quantitative performance evaluation using whole-image SSIM and coefficient of variation (CV).
• Statistical analysis (ANOVA, Kruskal-Wallis) to determine significant differences and optimal hyperparameters for RAKI.

Main Results:

• RAKI methods generally yielded higher SSIM values than GRAPPA counterparts, particularly at net acceleration factors (Rnet) > 4.
• Specific RAKI methods (readout SENSE-RAKI, split-slice-RAKI) performed well for Rnet ≥ 8.
• Optimal RAKI hyperparameters identified as 500 epochs, 3 hidden layers, and 128 neurons per layer for a balance of speed and performance.

Conclusions:

• Nonlinear RAKI methods demonstrate superior performance over linear GRAPPA for SMS reconstruction, especially at high acceleration.
• Specific GRAPPA and RAKI algorithms show good performance within certain acceleration ranges (Rnet ≤ 6 and Rnet ≥ 8).
• Findings provide a foundation for hyperparameter tuning and development of GPU-accelerated SMS reconstruction toolboxes, though dataset-specific limitations exist.