Abstract
Motion artifacts are a pervasive problem in MRI, leading to misdiagnosis ormischaracterization in population-level imaging studies. Current retrospectiverigid intra-slice motion correction techniques jointly optimize estimates ofthe image and the motion parameters. In this paper, we use a deep network toreduce the joint image-motion parameter search to a search over rigid motionparameters alone. Our network produces a reconstruction as a function of twoinputs: corrupted k-space data and motion parameters. We train the networkusing simulated, motion-corrupted k-space data generated from known motionparameters. At test-time, we estimate unknown motion parameters by minimizing adata consistency loss between the motion parameters, the network-based imagereconstruction given those parameters, and the acquired measurements.Intra-slice motion correction experiments on simulated and realistic 2D fastspin echo brain MRI achieve high reconstruction fidelity while retaining thebenefits of explicit data consistency-based optimization. Our code is publiclyavailable at https://www.github.com/nalinimsingh/neuroMoCo.