New arXiv Paper Proposes Magnitude-Informed Deep Learning for Dynamic MRI Reconstruction

A recent preprint available on arXiv (2608.18036) investigates the potential of utilizing k-space measurements based solely on magnitude to enhance the reconstruction of accelerated steady-state dynamic MRI. The researchers highlight that conventional MRI reconstruction typically depends on complex-valued data, but magnitude-only information derived from sparse phase retrieval could provide valuable insights. They report a significant consistency of k-space magnitudes over different time frames and introduce a technique named C+Mag. This approach integrates magnitude-informed physics with deep learning, utilizing an ADMM-based unrolling framework alongside a unique magnitude-aware data fidelity formulation that employs quadratically smoothed optimization.