SELFIE: SElf-supervised Learning for Fast dynamIc golden-anglE radial MRI reconstruction with auto-extracted representations

Motivation: There is a need for self-supervised reconstruction methods for dynamic MRI, where acquiring a fully-sampled reference is impractical. Current self-supervised reconstructions do not exploit the rich temporal properties of dynamic imaging. Goal(s): To develop a self-supervised reconstruction method for dynamic MRI with high spatial and temporal resolution. Approach: The proposed SELFIE technique leverages the sampling properties of golden-angle radial acquisition to auto-extract anatomical and temporal representations directly from the acquired data and the inherent sparsity of dynamic images to self-supervise network training. Results: SELFIE achieves similar performance to supervised learning for DCE-MRI and free-breathing motion-resolved MRI without the need for compressed sensing references. Impact: SELFIE can achieve comparable performance to supervised deep learning without the limitation of using a compressed sensing reference, which is promising for challenging clinical applications where acquiring a reference is impractical.