基于生成对抗网络和域一致性的MRI运动伪影校正方法

Motion artifact correction of MRI based on a generative adversarial network and domain consistency

在临床诊断中,磁共振成像(magnetic resonance imaging,MRI)运动伪影是一个常见的问题,运动伪影的存在会影响医生的诊断,虽然重新采集MRI可以避免这一问题,但这会提高医院和患者的经济成本和时间成本,因此,运动伪影的校正具备实用研究价值.现有的研究主要关注于空域的运动伪影校正或者K空间的运动伪影校正,缺乏对K空间和空域之间数据一致性的保持.为了解决这一问题,本文基于生成对抗网络提出了保持K空间和空域之间数据一致性的MRI运动伪影校正模型.该模型通过频域生成器对K空间数据进行初步校正,然后通过空域生成器对空域中的数据进行精细校正,在优化阶段则采用域间数据一致性损失来保持K空间和空域之间的数据一致性.在公开脑部MRI数据集ADNI,ABIDE,OASIS和Brain上的实验结果表明,本文提出的模型相较于其他方法分别在PSNR,SSIM以及RMSE上最高提升了3.4%,3.07%和15.57%.

In clinical diagnosis,magnetic resonance imaging(MRI)motion artifact is a common problem that will affect the doctor’s diagnosis.Although the reacquisition of MRI can avoid this problem,it will bring extra economic and time cost to hospitals and patients.Therefore,the correction of motion artifacts has a practical research value.Existing studies mainly focus on correcting motion artifacts from the spatial domain or K-space,and ignore the data consistency between them.In order to solve this problem,a motion artifact correction model based on a generative adversarial network is proposed to maintain the data consistency between the K-space and spatial domain.In this model,K-space data are initially corrected by a frequency domain generator,and then spatial domain data are fine corrected by a spatial domain generator.In optimization phase,the data consistency loss between the K-space and spatial domain is used to maintain data consistency.In four public MRI datasets,ADNI,ABIDE,OASIS and Brain,the experimental results show that the performance of the proposed model has increased by 3.4%,3.07%and 15.57%on PSNR,SSIM and RMSE,respectively.