Automatic Kernel Selection For Optimal GRAPPA Reconstruction

GRAPPA [1] has emerged to be a popular k-space-based parallel imaging reconstruction technique. GRAPPA procedure involves a block-wise reconstruction in which multiple k-space lines from all coils are combined to fit a missing line for each single coil. Recent work has suggested the incorporation of acquired k-space points in both phase encoding (PE) and frequency encoding (FE) directions in the reconstruction [2]. Two main types of error exist for GRAPPA [3]: model error and noise-related error. The model error arises from the use of a limited (or truncated) size of the k-space subset, or kernel, to fit each missing datum. The noise-related erro