Aliasing Layer: A New Method for Removing Parallel Imaging Aliasing and EPI Ghosting Artifacts in CNN [Presidential Award Proceedings]

Abstract

Purpose : Residual aliasing artifacts are often generated in MR images acquired using parallel imaging (PI) and/or echo-planar imaging (EPI). Existing denoising methods based on convolutional neural networks (CNNs) assume that an image has spatial locality. Since the artifacts do not satisfy the assumption of CNNs, denoising methods based on CNNs cannot remove artifacts efficiently. In this presentation, the author proposes a new method that can significantly reduce residual aliasing artifacts. The proposed method utilizes the locations of aliasing artifacts and/or N-half ghost artifacts, which can be analytically calculated.

Methods : CNNs based on ResNet, with and without aliasing layers (ALs), were used to remove artifacts from the reconstructed MR images. For training and testing the CNNs, 31 fluid-attenuated inversion recovery (FLAIR) images (30 training, 1 testing) were used. PI aliasing artifacts were simulated by adding noise to the k-space. In the case of EPI, even and odd encodes in k-space were shifted to simulate aliasing artifacts.

Results : The use of ALs significantly reduced training and validation errors. The CNN without ALs suppressed the brain structures where signals were weak. In contrast, the correction method using CNNs with the proposed ALs suppressed parallel imaging aliasing and EPI ghosting artifacts, selectively.

Conclusion : The correction method using the proposed AL could effectively remove PI aliasing and EPI ghosting artifacts.