Learned K-Space Partitioning for Improved Dual-Domain Self-Supervised Image Reconstruction

Motivation: K-space partitioning schemes for self-supervised learning via data undersampling (SSDU) remain unoptimized, with current partitioning typically performed heuristically based on the original sampling distribution. Goal(s): To improve self-supervised image reconstruction by learning optimal k-space partitioning. Approach: We modify SSDU by using a dual domain loss and learning a secondary sampling probability distribution to partition k-space into two sets: a data consistency and loss. Results: Dual domain and learned partitioning produces higher quality images than k-space only self-supervised learning. Impact: Learned k-space partitioning enhances reconstruction quality better utilizing acquired data for reconstructions. Learned k-space partitioning provides a framework for optimizing self-supervised partitioning to diverse k-space trajectories which previously were hand-picked or sub-optimal.