Junhao Zhang1,2, Zheyuan Yi1,2, Yujiao Zhao1,2, Linfang Xiao1,2, Jiahao Hu1,2, Christopher Man1,2, Yujiao Zhao3, and Ed X. Wu1,2
1Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong, China, 2Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China, 3The University of HongKong, HongKong, China
Synopsis
Conventional
ESPIRiT reconstruction requires accurate estimation of ESPIRiT maps from autocalibration
samples or signals but acquiring such autocalibration signals takes time and
may not be straightforward in some situations. This study aims to deploy deep
learning to directly estimate ESPIRiT maps from uniformly undersampled
multi-channel 2D MR data that contain no autocalibration signals. Results show that the estimated ESPIRiT maps
could be reliably obtained and they could be used for ESPIRiT and SENSE
reconstruction with high acceleration.
Introduction
Conventional
parallel imaging technique requires the explicit knowledge of coil
sensitivities[3] or corresponding relations in k-space from
either calibration scans or autocalibration signals (ACS) for reconstruction.
These methods suffer from subject’s motion and additional imaging time. ESPIRiT[1], one of parallel imaging techniques in
practice, could reconstruct the images from undersampled MR data through estimated
ESPIRiT maps (EMS). However, the accurate estimation of ESPIRiT maps of
dominant eigen values requires additional autocalibration signals[2], which takes time and may not be straightforward
in some situations. The EMS are coil-dependent and each subject has slightly
different coil-subject geometry in MR receiving coil system, leading to small
variations in EMS. It is necessary to minimize such variations by spatial alignment
and then exploit the data correlations from the same MR receiving coil system. This
study aims to estimate the ESPIRiT maps directly from uniformly undersampled multi-channel
2D MR data through a 2D convolutional neural network with the U-Net
architecture and apply the maps for ESPIRiT image reconstruction.Theory and Method
Proposed framework
EMS
explicitly characterizes coil sensitivity function of MR receiving system using
autocalibration signals and apply to reconstruct undersampled data in image
space. EMS are coil-dependent information and different subjects may have
different coil-subject geometry. This inevitably brings variations in shareable
EMS among different subjects. Therefore, there is necessity to incorporate
coil-subject geometry parameters to carry out spatial alignment to minimize the
variations of EMS among different subjects. The framework of the proposed method is summarized in Figure 1A.
Experiment Preparation
Multi-channel
coil data used in this study comes from Calgary-Campinas Public Database[4], including fully sampled human brain datasets
from 67 healthy subjects collected on a 1.5T clinical scanner (GE Healthcare,
Waukesha, WI). T1-weighted (T1W) acquisition parameter was TE/TR/TI =
6.3/2.6/650 ms or TE/TR/TI = 7.4/3.1/400 ms. The datasets were reduced to 6
channels by coil combination[5]. The
matrix size of each channel is Nx×Ny×Nz = 128×128×100. Two parts of complex data were treated as two channels. All the
data were randomly partitioned into a training/validation/test set. The multi-channel MR data are
spatially aligned by performing a rigid-body rotation and translation to
minimize the coil sensitivity variations. The details of data alignment were
demonstrated in Figure 1C. 24 fully
sampled central k-space lines were used to estimate reference
ESPIRiT maps. The kernel size was set to 6×6. The k-space data were uniformly undersampled
at different acceleration factors (Rs=2, 3 and 4). The model used is
modified from typical U-Net[6] and attention module was introduced to
effectively combine information from different channels[7].Results
As shown in Figure
2A, EMS estimated from uniformly undersampled data (R = 4) via deep learning
were comparable to reference EMS derived from 24 ACS lines and the correlation
analysis also demonstrated that the estimated EMS are in good agreement with the
reference EMS (shown in Figure 2B).
Under the acceleration factor R=2, 3 and 4, the estimated EMS were used for
ESPIRiT image-space reconstruction and the reconstructed images showed slight residual
errors without pronounced artifacts compared with fully-sampled reference. Images
reconstructed using estimated EMS were almost comparable to the images
reconstructed using EMS from 24 ACS lines. The PSNR and MSE of reconstructed
images using EMS estimated by deep learning or EMS from 24 ACS lines got worse with
acceleration factor increasing from 2 to 4 as shown in Figure 3. Additionally, the estimated EMS were used for SENSE image
space reconstruction and similar results as ESPIRiT reconstruction were observed
as shown in Figure 4. In Figure 5, the reconstructed images
with large head rotation in MR coil receiving system were shown.Discussion and Conclusions
Without
autocalibration signals, the EMS could be directly estimated from uniformly
undersampled MR data by deep learning and applied to ESPIRiT/SENSE image
reconstruction. The estimated EMS show high correlations to the reference EMS
from 24 ACS lines. Images reconstructed using EMS based on ESPIRiT and SENSE
reconstruction at high acceleration did not show apparent artifacts.
The
EMS are coil-dependent and each subject has different coil-subject geometry,
resulting in variation in the coil-specific EMS. Thus, we exploit the
coil-subject geometry information to minimize the variations in ESPIRiT maps.
Such prior information is conventionally neglected. Phase variations would also
exist in scans among different subjects. Such phase changes were well preserved
to some extent in our deep learning model by using a hybrid loss, one was related
to minimizing the variations in coil-specific ESPIRiT sensitivity maps and
another was related to preserve the phase variations.
ESPIRiT maps of dominant eigen values are
estimated and used for MR image reconstruction. ESPIRiT maps corresponding to
smaller eigen values can also be estimated and employed for MR image
reconstruction in future study, which may further improve the image
reconstruction performance, especially for ultra-high-field MRI with rapid
phase variations.Acknowledgements
This work was supported in part by Hong Kong
Research Grant Council (R7003-19F, HKU17112120 and HKU17127121 to E.X.W., and
HKU17103819, HKU17104020 and HKU17127021 to A.T.L.L.), Lam Woo Foundation, and
Guangdong Key Technologies for Treatment of Brain Disorders (2018B030332001) to
E.X.W..References
[1] Uecker, M., et
al., ESPIRiT-An Eigenvalue Approach to
Autocalibrating Parallel MRI: Where SENSE Meets GRAPPA. Magnetic Resonance
in Medicine, 2014. 71(3): p.
990-1001.
[2] Sodickson, D., Parallel magnetic resonance imaging (or,
scanners, cell phones, and the surprising guises of modern tomography).
Medical Physics, 2007. 34(6): p.
2598-2598.
[3] Pruessmann, K.P.,
et al., SENSE: Sensitivity encoding for
fast MRI. Magnetic Resonance in Medicine, 1999. 42(5): p. 952-962.
[4] Warfield, S.K., K.H.
Zou, and W.M. Wells, Simultaneous truth
and performance level estimation (STAPLE): An algorithm for the validation of
image segmentation. Ieee Transactions on Medical Imaging, 2004. 23(7): p. 903-921.
[5]Buehrer, M., et
al., Array compression for MRI with large
coil arrays. Magnetic Resonance in Medicine, 2007. 57(6): p. 1131-1139.
[6] Ronneberger, O.,
P. Fischer, and T. Brox, U-Net:
Convolutional Networks for Biomedical Image Segmentation. Medical Image
Computing and Computer-Assisted Intervention, Pt Iii, 2015. 9351: p. 234-241.
[7] Hu,
J., L. Shen, and G. Sun, Squeeze-and-Excitation
Networks. 2018 Ieee/Cvf Conference on Computer Vision and Pattern
Recognition (Cvpr), 2018: p. 7132-7141.