MRI Motion Correction Through Disentangled CycleGAN Based on Multi-Mask K-Space Subsampling

This work proposes a new retrospective motion correction method, termed DCGAN-MS, which employs disentangled CycleGAN based onmulti-mask k-space subsampling (DCGAN-MS) to address the image domain translation challenge. The multi-mask k-space subsampling operator is utilized to decrease the complexit...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on medical imaging 2025, p.1-1
Main Authors: Chen, Gang, Xie, Han, Rao, Xinglong, Liu, Xinjie, Otikovs, Martins, Frydman, Lucio, Sun, Peng, Zhang, Zhi, Pan, Feng, Yang, Lian, Zhou, Xin, Liu, Maili, Bao, Qingjia, Liu, Chaoyang
Format: Article
Language:English
Subjects:
Atomic measurements
CycleGAN
Gadoxetic acid-enhanced human liver MRI
Generators
Image reconstruction
Liver
Magnetic resonance imaging
motion artifact
Motion artifacts
multi-mask k-space subsampling
Rodents
Technological innovation
Transient analysis
Translation
unpaired learning
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This work proposes a new retrospective motion correction method, termed DCGAN-MS, which employs disentangled CycleGAN based onmulti-mask k-space subsampling (DCGAN-MS) to address the image domain translation challenge. The multi-mask k-space subsampling operator is utilized to decrease the complexity of motion artifacts by randomly discarding motion-affected k-space lines. The network then disentangles the subsampled, motion-corrupted images into content and artifact features using specialized encoders, and generates motion-corrected images by decoding the content features. By utilizing multi-mask k-space subsampling, motion artifact features become more sparse compared to the original image domain, enhancing the efficiency of the DCGAN-MS network. This method effectively corrects motion artifacts in clinical gadoxetic acid-enhanced human liver MRI, human brain MRI from fastMRI, and preclinical rodent brain MRI. Quantitative improvements are demonstrated with SSIM values increasing from 0.75 to 0.86 for human liver MRI with simulated motion artifacts, and from 0.72 to 0.82 for rodent brain MRI with simulated motion artifacts. Correspondingly, PSNR values increased from 26.09 to 31.09 and from 25.10 to 31.77. The method's performance was further validated on clinical and preclinical motion-corrupted MRI using the Kernel Inception Distance (KID) and Fréchet Inception Distance (FID) metrics. Additionally, ablation experiments were conducted to confirm the effectiveness of the multi-mask k-space subsampling approach.
ISSN:0278-0062
1558-254X
DOI:10.1109/TMI.2024.3523949