Abstract
Motion artifacts are a primary source of magnetic resonance (MR) imagequality deterioration with strong repercussions on diagnostic performance.Currently, MR motion correction is carried out either prospectively, with thehelp of motion tracking systems, or retrospectively by mainly utilizingcomputationally expensive iterative algorithms. In this paper, we utilize anovel adversarial framework, titled MedGAN, for the joint retrospectivecorrection of rigid and non-rigid motion artifacts in different body regionsand without the need for a reference image. MedGAN utilizes a uniquecombination of non-adversarial losses and a novel generator architecture tocapture the textures and fine-detailed structures of the desired artifacts-freeMR images. Quantitative and qualitative comparisons with other adversarialtechniques have illustrated the proposed model's superior performance.