提示:文章写完后,目录可以自动生成,如何生成可参考右边的帮助文档
提示:以下是本篇文章正文内容
在图像域中产生的混叠伪影是结构性的和非局部的,仅靠图像域修复不充分,且输入为欠采样的K空间重建图像存在细节扭曲、消失的问题
以往的研究仅限于常规的网络设计,如多层CNN和U-Net,很少有人尝试为欠采样的MRI重建设计定制网络结构。
MRI采集的过程中有多模态图像,但没有引入成像速度快的图像如T1w作为先验知识指导MRI重建过程
Develop a dual domain learning schemein MRI, which allows the network to restore the data in both image and frequency domains in a recurrent fashion.
we propose a Dilated Residual Dense Network (DRD-Net) with a Squeeze-and-Excitation Dilated Dense Residual Block(SDRDB) as the building module.
The DRD-Net is used for both image and frequency domain
restorations.
Our SDRDB is customized for MRI reconstruction task.
we propose to use T1WI as deep prior in both image domain and k-space domain for improving the MRI reconstruction fidelity, given that the structural information in T1WI is highly correlated with that in different MRI protocols
In summary, we propose a Dual Domain Recurrent Network (DuDoRNet) embedding with T1 priors to address these problems by learning two DRD-Nets on dual domains in a recurrent fashion to restore k-space and image domains simultaneously.
F−1 is used for global residual learning in the third part. The second extracted feature F0 is used as SDRDB input.
作用:initial feature extraction (IFE)
数量:two
上面的DRD-Net是整体网络架构,不同于此处的网络类型DRDNet
4个非线性层的密集连接使得具有不同感受野的层的组合多种多样,这比传统的扩张方法能更有效地提取不同尺度的特征
组成:one 1 × 1 and one 3 × 3 convolution layers
Fn为各层SDRDB提取的特征,得到全局特征(GF)
融合全局特征和初始特征,并送入 reconstruction output
3 × 3 convolution layer
1、具有大的感受野的DRD-Net可以感知更多的信号,以达到更好的修复效果;
2、递归学习可以更好地避免直接优化中的过拟合;
3、图像域的恢复可以通过融合K空间恢复的信号而得到加强,反之亦然;
an in-house MRI dataset consisting of 20 patients
3 type of each patient
360 2D images are generated for each protocol
T1, T2, and FLAIR& full k-space sampled
training/validation/test sets with a ratio of 7 : 1 : 2
3种
Cartesian, radial, and spiral trajectories
PSNR
SSIM
MSE
TV-CS
GRAPPA
Four CNN-based algorithms
1、所有的采样模式和重建方法中,我们的DuDoRNet的螺旋模式产生了最好的图像质量
2、我们的重建对输入中的混叠伪影和结构损失是稳健的
3、我们的方法恢复了图像和K空间的信息,并保留了重要的解剖学细节
提示:这里对文章进行总结:
通过具有大感受野的DRD-Nets来循环恢复图像和k-空间领域。T1先验被嵌入到每个循环块中,以深入指导这两个领域的恢复工作。广泛的实验结果表明,虽然以前的单域快速MRI方法在直接减少图像域的混叠伪影方面能力有限,但我们的DuDoRNet可以有效地恢复重构,并且T1先验可以进一步显著提高结构恢复能力