Haoxiang Li1,2, Jing Chen1, Yuanyuan Liu1, Hairong Zheng1, Dong Liang1, and Yanjie Zhu1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, 2Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, 2Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China
This work proposes
a deep learning based method to accelerate MR parametric mapping both by
reducing the contrast number and undersampling the k-space data.
Fig. 1 (a) The
proposed MRI parameter mapping method consists of two neural networks module
and a pixel-wise curve fitting process. (b) The reconstruction module is based on
deep ADMM-Net for reconstruction of weighted images from under-sampled k-space
data. (c) The generative module is a densely connected neural network to
generate the corresponding weighted images from the reconstructed image pairs
under the supervision of corresponding weighted images.
Fig. 5 The ROI
analysis of test $$$T_{1\rho}$$$ knee data (R=1) and error map of test $$$T_{1\rho}$$$ knee data
(R=1); Proposed: Mapping with 2 acquired fully sampled images and 3 synthetic images (Equivalent R = 2.5); Two contrast images: Mapping with 2 acquired fully sampled images only (Equivalent R = 2.5).
