Automating Kernel Size Selection in MRI Reconstruction via a Transparent and Interpretable Search Approach

GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA) is a clinical Magnetic Resonance Imaging (MRI) reconstruction method. The kernel size in GRAPPA directly controls the image quality and the optimal kernel size can be manually selected through comparing multiple reconstructed images. However, the optimal kernel size is often impractical to be manually selected in clinical settings. To resolve this issue, we propose an automated kernel size selectionmethod utilizing grid search, whichmaintainsGRAPPA's transparent and interpretable nature in a linear interpolation process. This strategy redefines kernel size selection as an exhaustive search problem and tests all potential kernel sizes within a predefined hyperparameter space. Experimental results, evaluated through both qualitative and quantitative metrics, demonstrate the effectiveness of our method in consistently identifying the optimal kernel size. The proposed approach significantly enhances the efficiency and utility of GRAPPA reconstruction for ensuring high image quality pivotal in accurate clinical diagnoses and treatment plans.