Fanwen Wang1, Hui Zhang1, Jiawei Han1, Fei Dai1, Yuxiang Dai1, Weibo Chen2, Chengyan Wang3, and He Wang1,3
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, 2Philips Healthcare, Shanghai, China, 3Human Phenome Institute, Fudan University, Shanghai, China
Synopsis
This study proposed a novel MAGnitude Image
to Complex (MAGIC) Network to reconstruct images using deep learning with limited
number of training data. Collecting complex multi-coil data is inconvenient
since it is beyond the routine examination. However, there are many magnitude
images available in hospitals. By applying deformation between the magnitude image
and complex image, MAGIC Net succeeded in synthesizing deformed data for
training and enabled deep learning methods. Results show that with the same
original data, MAGIC-Net outperforms the conventional CG-SENSE in PSNR for all
undersampling trajectories with high resolution b = 0 and b =
1000 s/mm2.
Introduction
Deep learning has been widely used in the
field of MRI reconstruction. Using fully sampled data as labels, many deep
learning networks1,2 have
succeeded in recovering artifact-free images from accelerated MRI acquisitions,
with much better reconstruction performance than model-based parallel imaging
methods. However, the effectiveness of deep learning is limited by the size and
diversity of training data3. Collecting complex multi-coil MRI data is beyond the routine
procedures in hospital and takes a lot of storages. Meanwhile, many magnitude
images are available in the database.
Hence, we proposed a MAGnitude Images to Complex (MAGIC)-Net to synthesize large
number of complex training data from the existing magnitude images to facilitate
deep learning reconstructions.Methods
Data
Acquisition
A total of 768 images from 48
patients were exported from the database of a medical center. Four healthy volunteers were recruited in this study, with 16 slices
scanned each. Subjects were grouped into a source group (1 subject, 16 slices)
and test group (3 subjects, 48 slices). All data were acquired on a 3.0 T MRI
scanner (Ingenia CX, Philips Healthcare, Best, the Netherlands) equipped with a
32-channel head coil. Multi-shot Diffusion Weighted Images (DWI) were acquired, consisting
b-values of 0 and 1000 s/mm2 with interleaved four-shot EPI acquisition.
The following parameters were utilized: TE, 75 ms; TR, 2800 ms; matrix size,
228x228; slice thickness, 4 mm; partial
Fourier factor, 0.702; voxel size, 1.0x1.0 mm2. The acquired
data were reconstructed using MUSE4.
Pipeline of
MAGIC-Net
Firstly, we unwrapped the phase of the source multi-Coil
Sensitivity Maps (CSMs) based on the transportation of intensity equation5. Secondly, we calculated the
geometric deformation between the source and magnitude cases. Thirdly, we applied
the deformation map to the complex data to synthesize data with different
anatomies. The deformation was implemented using the SPM DARTEL6 toolbox.
Experiment
Setup
The deformations were calculated between 48
subjects of magnitude images and 1 subject of source complex data. Hence the number
of synthesized data set is up to 48 subjects (768 slices). To validate the
effectiveness of the deformed samples, we further used sets of 6, 12, 32 and 48
subjects (96, 192, 284 and 768 slices) for comparison.
A state-of-art network MoDL7 was trained to evaluate the performance of MAGIC-Net for deep
learning. MoDL cascaded the data consistency and ResNet together, said
to be relatively insensitive to training data beyond 100 images.
Interleaved undersampling trajectories with
rates of 2, 4 and 6 were applied to simulate the sparse multi-coil
measurements. Results
To ensure the feasibility of the
deformation, we synthesized the source subject to one test subject. Comparison of the source,
target and deformed DWI with b = 0 and b = 1000 s/mm2 were
shown (Fig. 2). The source, target and deformed CSMs with the unwarping procedure
were also shown. It is worth noting that the deformed subject shown in the
figures was NOT included for training. Compared with the directly deformed CSMs, the
unwrapped phases have better smoothness and less distortion, hence
providing better space information (Fig. 3).
Then we also calculated PSNR with
different undersampling trajectories reconstructed by MAGIC Net and model-based
CG-SENSE8,9 (Fig.
4). As the synthesized samples increased from 96 to 768 slices, the MAGIC-Net outperformed
CG-SENSE for high-resolution b =1000 s/mm2 in PSNR (49.17 v.
s. 32.65, 38.66 v. s. 30.42, 33.01 v. s. 28.32) under the undersampling rate of
2,4 and 6 respectively. Furthermore, for all undersampling trajectories with b
= 0 and b =1000 s/mm2, MAGIC-Net shows less error compared to
CG-SENSE and better PSNR without actually collected complex data (Fig.
5).
Conclusion
We proposed a MAGIC-Net to
reconstruct images using deep learning with limited
number of training data. Results show that MAGIC-Net enables the deep learning network by synthesizing diverse data and outperforms the
conventional CG-SENSE with PSNR for different undersampling trajectories. The pre-trained
reconstruction network can be replaced by other deep learning networks. The
model can also be extended for other imaging sequences without acquiring large
number of raw data for training. Acknowledgements
The
work was supported in part by the National Natural Science Foundation of China
(No. 81971583), National Key R&D Program of China (No. 2018YFC1312900),
Shanghai Natural Science Foundation (No. 20ZR1406400), Shanghai Municipal
Science and Technology Major Project (No.2017SHZDZX01, No.2018SHZDZX01) and
ZJLab.References
- Hammernik
K, Klatzer T, Kobler E, et al. Learning a variational
network for reconstruction of accelerated MRI data. Magn Reson Med. 2018;79(6):3055-3071.
- Zhang P, Wang F, Xu
W, Li Y. Multi-channel generative adversarial network for parallel magnetic
resonance image reconstruction in k-space. Paper presented at: International
Conference on Medical Image Computing and Computer-Assisted Intervention2018.
- Ding J, Li X,
Gudivada VN. Augmentation and evaluation of training data for deep learning.
Paper presented at: 2017 IEEE International Conference on Big Data (Big
Data)2017.
- Chen NK, Guidon A,
Chang HC, Song AW. A robust multi-shot scan strategy for high-resolution
diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE). NeuroImage. 2013;72:41-47.
- Zixin Zhao (2020). Robust 2D phase unwrapping algorithm (https://www.mathworks.com/matlabcentral/fileexchange/68493-robust-2d-phase-unwrapping-algorithm), MATLAB Central File Exchange. Retrieved December 14, 2020.
- Ashburner J. A fast
diffeomorphic image registration algorithm. Neuroimage.
2007;38(1):95-113.
-
Aggarwal HK, Mani
MP, Jacob M. MoDL: Model-based deep learning architecture for inverse problems.
IEEE transactions on medical imaging. 2018;38(2):394-405.
- Pruessmann KP,
Weiger M, Scheidegger MB, Boesiger P. SENSE: sensitivity encoding for fast MRI.
Magnetic Resonance in Medicine: An
Official Journal of the International Society for Magnetic Resonance in
Medicine. 1999;42(5):952-962.
- Uecker M, Lai P,
Murphy MJ, et al. ESPIRiT—an eigenvalue approach to autocalibrating parallel
MRI: where SENSE meets GRAPPA. Magn Reson
Med. 2014;71(3):990-1001.