Motivation: Non-Cartesian GRAPPA shows promise for dynamic imaging reconstruction but demands extensive calibration data and computation. A data-efficient implicit-GRAPPA approach could improve speed and usability without sacrificing image quality. Goal(s): To achieve an efficient GRAPPA kernel with reduced calibration data requirements, matching or exceeding the quality of previous methods for real-time radial trajectory reconstruction. Approach: Our method uses spatially structured source-target mappings to learn data-efficient GRAPPA weights per target point through a multilayer perceptron (MLP) model, learning implicit correlations across calibration trajectories. Results: Our implicit-GRAPPA approach achieves high quality reconstruction with 31.73% lower RMSE than current state-of-art techniques while requiring ~15x less training data. Impact: The proposed implicit-GRAPPA improves the image reconstruction quality and markedly reduces the calibration data requirement of non-Cartesian GRAPPA, and should prove useful for applications requiring robust real-time non-Cartesian reconstructions.
Zhao et al. (Tue,) studied this question.