Motion-Resolved MRI Reconstruction with Deep Learning

This paper addresses the challenge of respiratory motion artifacts in MRI, a significant problem in abdominal and pulmonary imaging. The authors propose a two-stage deep learning approach (MoraNet) for motion-resolved image reconstruction using radial MRI. The method estimates respiratory motion from low-resolution images and then reconstructs high-resolution images for each motion state. The use of an interpretable deep unrolled network and the comparison with conventional methods (compressed sensing) highlight the potential for improved image quality and faster reconstruction times, which are crucial for clinical applications. The evaluation on phantom and volunteer data strengthens the validity of the approach.