Skip to content

Latest commit

 

History

History
43 lines (29 loc) · 1.6 KB

accelerating-magnetic-resonance-imaging-via-deep-learning.md

File metadata and controls

43 lines (29 loc) · 1.6 KB

Paper:

Accelerating Magnetic Resonance Imaging via Deep Learning

Author:

Shanshan Wang, Zhenghang Su,Leslie Ying,Xi Peng,Shun Zhu Feng Liang, Dagan Feng 5 and Dong Liang

Year:

2016

Notes:

泛读。这篇文章可能早期的在MRI重建中使用深度学习的论文。使用的方法比较简单,大致的做法是先训练一个网络,用这个网络作为一个正则项来重建。训练时损失函数为图像域的2范数。输入是zero-fill重建,输出为去伪影的结果。网络结构为3层卷积网络,使用了ReLU激活函数。

假设训练好的网络为 $C$,作者在论文中提供了三种重建方式:

  1. $C$ 作为正则项 $$ \underset{u}{\operatorname{argmin}}\left{\left|C\left(F^{H} f ; \hat{\Theta}\right)-u\right|{2}^{2}+\lambda\left|f-F{M} u\right|_{2}^{2}\right} $$

  2. 增加其他正则项 $$ \underset{u}{\operatorname{argmin}}\left{\left|C\left(F^{H} f ; \hat{\Theta}\right)-u\right|{2}^{2}+\lambda\left|f-F{M} u\right|_{2}^{2}+\beta \operatorname{Reg}(u)\right} $$

  3. 两阶段重建。先用网络重建 $C\left(F^{H} f ; \hat{\Theta}\right)$,然后用这个作为传统压缩感知算法的初值重建。

Bibtex:

@inproceedings{wang2016accelerating,
  title={Accelerating magnetic resonance imaging via deep learning},
  author={Wang, Shanshan and Su, Zhenghang and Ying, Leslie and Peng, Xi and Zhu, Shun and Liang, Feng and Feng, Dagan and Liang, Dong},
  booktitle={2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI)},
  pages={514--517},
  year={2016},
  organization={IEEE}
}