Motion artifact reduction for magnetic resonance imaging with deep learning and k-space analysis

Abstract

Motion artifacts deteriorate the quality of magnetic resonance (MR) images. This study proposes a new method to detect phase-encoding (PE) lines corrupted by motion and remove motion artifacts in MR images. 67 cases containing 8710 slices of axial T2-weighted images from the IXI public dataset were split into three datasets, i.e., training (50 cases/6500 slices), validation (5/650), and test (12/1560) sets. First, motion-corrupted k-spaces and images were simulated using a pseudo-random sampling order and random motion tracks. A convolutional neural network (CNN) model was trained to filter the motion-corrupted images. Then, the k-space of the filtered image was compared with the motion-corrupted k-space line-by-line, to detect the PE lines affected by motion. Finally, the unaffected PE lines were used to reconstruct the final image using compressed sensing (CS). For the simulated images with 35%, 40%, 45%, and 50% unaffected PE lines, the mean peak signal-to-noise ratio (PSNRs) of resulting images (mean±standard deviation) were 36.129±3.678, 38.646±3.526, 40.426±3.223, and 41.510±3.167, respectively, and the mean structural similarity (SSIMs) were 0.950±0.046, 0.964±0.035, 0.975±0.025, and 0.979±0.023, respectively. For images with more than 35% PE lines unaffected by motion, images reconstructed with proposed algorithm exhibited better quality than those images reconstructed with CS using 35% under-sampled data (PSNR 37.678±3.261, SSIM 0.964±0.028). It was proved that deep learning and k-space analysis can detect the k-space PE lines affected by motion and CS can be used to reconstruct images from unaffected data, effectively alleviating the motion artifacts.

Swin Transformer and the Unet Architecture to Correct Motion Artifacts in Magnetic Resonance Image Reconstruction

“…The performance of MACS-Net was compared to those of MoCo-Net [ 11 ], Modified-2D-Net (which uses a CNN to correct out-of-field-of-view MAs) [ 12 ], Namer-Net [ 26 ], the MC-Net motion correction network [ 43 ], Stacked-Unet (which uses self-assisted priors) [ 44 ], and Mark-Net (which performs MA reduction and k-space analysis) [ 45 ]. These methods are mainly used in encoder-decoder frameworks, particularly Unet, and executed convolutional operations.…”

Training Robust T1-Weighted Magnetic Resonance Imaging Liver Segmentation Models Using Ensembles of Datasets with Different Contrast Protocols and Liver Disease Etiologies

“…Aside from obtaining more data, however, future work that might involve re ning the data that we have involves image de-noising. Cui et al recently used a 2D convolutional neural network and k-space analysis to reduce and remove motion artifacts from corrupted T2-weighted brain MR images 27 .…”

“…Aside from obtaining more data, however, future work that might involve refining the data that we have involves image de-noising. Cui et al recently used a 2D convolutional neural network and k-space analysis to reduce and remove motion artifacts from corrupted T2-weighted brain MR images 27 . Given both the prevalence of motion artifacts in DLDS and the fact that, as Macdonald et al noted, such artifacts are not uncommon in a clinical setting 6 , an algorithm that can be applied to remove motion artifacts from liver MR images would expedite the training of robust deep learning segmentation models to assist in preventive surgery.…”

MRI motion artifact reduction using a conditional diffusion probabilistic model (MAR‐CDPM)

“…In addition, the supervised Unet model was used to reduce noise level where the denoised kspace data were line-by-line compared with ground truth k-space to reconstruct motion-free images. 20 However, compressed-sensing algorithms are computationally lengthy and are susceptible to reconstructing images with residual artifacts. 21,22 Supervised Unet deep learning models were developed to remove motion artifacts from quantitative MRI 23,24 and T1W anatomical MRI images, 25 where the motion artifacts were simulated in the spatial domain and frequency (k-space) domain, respectively.…”

“…However, extending this method to nonrigid movement is challenging such as reducing motion artifacts in the lung region with deformable changes. In addition, the supervised Unet model was used to reduce noise level where the denoised k ‐space data were line‐by‐line compared with ground truth k ‐space to reconstruct motion‐free images 20 . However, compressed‐sensing algorithms are computationally lengthy and are susceptible to reconstructing images with residual artifacts 21,22 .…”