Low-rank and deep learning approach for MR ghost artifact correction

Abstract

The EPI (Echo Planar Imaging) technique is one of the MR imaging sequences that has been widely used for various imaging such as functional magnetic resonance imaging (fMRI) and diffusion weighted imaging (DWI) because of its short scan time. Ghost artifacts always occur because of the sequence characteristics of the EPI technique, and they cause difficulty when interpreting the images. Therefore, there have been many studies on removing these ghost artifacts. In this study, we propose a method to effectively remove ghost artifacts without any additional image acquisition or modification of the MR sequence. First, the images of even and odd lines are separated from each other in the EPI, and then, the low rank property of the Hankel matrix is applied. This proves that the problem of removing ghost artifacts can be solved by transforming it into an interpolation problem in the k-space. Although the low rank-based reconstruction technique effectively eliminates artifacts, it has a disadvantage in that it takes a long time because it requires a large computational complexity. Therefore, a deep learning technique is applied to solve this problem. Recently, the field of deep learning has been applied not only to simple image processing problems but also to various problems such as medical images because its application has been expanded to various fields. In this thesis, based on the theory that the basic convolution neural network is related to the Hankel matrix decomposition, we proposed a deep learning method using the k-space data which are the raw data of the magnetic resonance imaging. By directly restoring the data in the k-space rather than in the image domain, this method is effective in removing the artifacts of magnetic resonance images such as ghost artifacts of echo plane images. Compared with the existing method, the deep learning method learns the basis to reconstruct an image based on training data. Therefore, it is not necessary to restore definite mathematical modeling, and it has better performance than that of the conventional methods. In addition, deep learning methods can produce more accurate results with more data. Additionally, once the learning is completed, it is possible to reconstruct an image very quickly without having to perform many operations on the new data. Furthermore, our results show the possibility of applying the proposed method to not only ghost artifacts but to other various MR artifacts as well.