摘要
文中提出了解决定量磁化率成像中偶极子反卷积的病态逆问题和快速重建高质量无伪影的定量磁化率图像的算法。基于k空间阈值法(TKD)初步重建三维定量磁化率图像(QSM),随后将TKD重建图像输入训练完成的三维卷积神经网络(CNN)模型中获得预测图像。在k空间中将TKD重建图像与CNN预测图像进行融合重建最终QSM图像。结果表明:与金标准相比,算法能够重建视觉上误差较小和无条形伪影的磁化率图像;卷积神经网络可以恢复病态区域的信号,k空间融合方法有效解决了偶极子反卷积的病态性。测试集上的重建结果在标准均方根误差(NRMSE)和高频误差范数(HFEN)重建误差上均低于主流算法。
To solve the ill-posed inverse problem of dipole deconvolution in quantitative magnetic susce-ptibility mapping,rapidly reconstruct high-quality and artifact-free susceptibility maps,an algorithm is proposed in this paper.3 D quantitative susceptibility map is preliminarily reconstructed by threshold-based k-space Division(TKD).The interim image is fed into pretrained three-dimensional Convolutional Neural Network(CNN) model to obtain a predicted image.The final QSM image is reconstructed by merging TKD reconstructed image with CNN predicted image in k-space.The results demonstrate that QSM images reconstructed by the proposed method are more visually satisfying,with less visual error and no stripe artifacts.Convolutional neural network can restore the signal in the ill-conditioned region,and k-space merging strategy can effectively solve the ill-posed dipole deconvolution.The reconstruction error on the test data sets is the lowest among the state-of-the-art algorithms in the sense of normalized root mean square error(NRMSE) and high frequency error norm(HFEN).
作者
刘杰
王一达
童睿
谢海滨
李建奇
杨光
LIU Jie;WANG Yi-da;TONG Rui;XIE Hai-bin;LI Jian-qi;YANG Guang(East China Normal University,School of Physics & Materials Science,Shanghai Key Laboratoryof Magnetic Resonance,Shanghai 200062,China)
出处
《信息技术》
2019年第10期72-76,共5页
Information Technology
基金
国家自然科学基金重点项目(61731009)
关键词
磁共振成像
定量磁化率成像
图像重建
深度学习
卷积神经网络
magnetic resonance imaging
quantitative susceptibility mapping
image recon-struction
deep learning
convolutional neural network
