Motion correction methods for brain MRI using deep learning

Abstract

In magnetic resonance imaging (MRI), motion artifacts can significantly degrade image quality. To reduce motion artifacts in brain MRI, we propose deep learning-based motion correction methods. The first study is the motion correction network for multi-contrast MRI. The proposed method consists of two parts: image alignment and motion correction. Alignment of multi-contrast MR images is performed in an unsupervised manner by a convolutional neural network, yielding transformation parameters to align input images. The motion correction network corrects motion artifacts in the aligned multi-contrast images. The proposed method successfully corrected artifacts of simulated motion and real motion. The second study is the unsupervised motion correction method using deep image prior. This approach takes advantage of the high impedance to noise offered by the neural network parameterization to remove motion artifacts in MR images. The proposed method synthesizes motion-simulated images from the motion-corrected images generated by the convolutional neural network, where an optimization process minimizes the objective function between the synthesized images and the acquired images. Furthermore, k-space segments are clustered based on their estimated motion parameters. The proposed method exhibits significant quantitative and qualitative improvements in correcting rigid and in-plane motion artifacts in MR images acquired using turbo spin-echo sequence.