Subspace dual-domain-loss for self-supervised deep learning reconstruction of dynamic MRI: Method and Application to CMR Multitasking

Motivation: Supervised deep learning (DL) can reduce reconstruction time for CMR Multitasking, but the lack of ground truth limits the quality of supervised DL to that of iteratively reconstructed labels. Goal(s): Our goal was to develop a self-supervised learning (SSL) reconstruction method, whose performance is not limited by the iterative reconstruction. Approach: We developed a dual-domain subspace SSL reconstruction method for non-Cartesian dynamic MRI, applying it to CMR Multitasking. Results: The proposed method can perform image reconstruction without reference images and shows better interscan consistency than supervised DL. Impact: With the proposed method, image quality of DL reconstruction for CMR Multitasking can potentially surpass iterative reconstruction. We applied subspace constraints to SSL reconstruction, showing an efficient way to relieve the computational burden of dynamic MRI SSL reconstruction.