Unsupervised learning for motion correction and assessment in brain magnetic resonance imaging using severity-based regularized cycle consistency

Magnetic resonance imaging is an important non-invasive diagnostic tool, yet is vulnerable to motion artifacts due to the long acquisition time. With the recent development of deep learning, numerous methods for improving motion artifacts have been proposed. Due to the difficulty in constructing paired datasets for motion artifact correction, various unsupervised learning-based models have emerged recently. Cycle-Consistent Generative Adversarial Network (CycleGAN), which is widely used as a base model for unsupervised learning, has ill-posed problem and high parameter complexity that poses limitations towards an unstable training process. To overcome these limitations, this paper proposes an unsupervised learning method for motion artifact correction and assessment by replacing one generator of CycleGAN with a motion artifact simulator and regularizing the cycle-consistency loss with a motion severity prior. In adopting this approach, the overall model's training complexity can be significantly reduced due to the decrease in the quantity of learnable parameters. Moreover, the calculation of the cycle-consistency loss becomes more stable, thereby addressing the inherent ill-posed problem of CycleGAN. As a result, the proposed method demonstrates superior performance quantitatively and qualitatively compared to existing motion artifact correction methods and succeeds in reconstructing clearer images qualitatively compared to supervised learning methods.