Hoai Nam Dang1, Vladimir Golkov2,3, Jonathan Endres1, Simon Weinmüller1, Felix Glang4, Thomas Wimmer2, Daniel Cremers2,3, Arnd Dörfler1, Andreas Maier5, and Moritz Zaiss1,4,6
1Institute of Neuroradiology, University Hospital Erlangen, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany, 2Technical University of Munich, Munich, Germany, 3Munich Center for Machine Learning, Munich, Germany, 4Magnetic Resonance Center, Max-Planck-Institute for Biological Cybernetics, Tuebingen, Germany, 5Department of Computer Science, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Erlangen, Germany, 6Department Artificial Intelligence in Biomedical Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany
Synopsis
Keywords: AI/ML Image Reconstruction, Machine Learning/Artificial Intelligence, super-resolution, turbo-spin-echo, joint-optimization
Motivation: TSE flip angle trains can have a strong influence on the actual resolution of the acquired image and have consequently a considerable impact on the performance of a super-resolution task.
Goal(s): We demonstrate the advantage of end-to-end optimization of sequence and neural network parameter compared to pure network training approaches.
Approach: This MR-physics-informed training procedure jointly optimizes radiofrequency pulse trains of a PD- and T2-weighted TSE and subsequently applied CNN to predict corresponding PDw and T2w super-resolution TSE images.
Results: The method generalizes from simulation-based optimization to in vivo measurements and acquired super-resolution images show higher accuracy compared to pure network training approaches.
Impact: Acquired super-resolution image may improve evaluation of the data.
Reduction of acquisition time compared to direct high-resolution acquisition
leads to increase in patient comfort and minimization of motion artifacts.
Introduction
Current MRI super-resolution methods use contrasts acquired
from typical clinical protocols as input for the neural network and disregard
the influence of the MR sequence parameters for optimization.
Using so-called
known operator learning1, we propose an approach that utilizes a MR physics
model during the optimization to not only train a neural network for
super-resolution, but also adapt the refocusing RF pulses to directly influence
the point-spread-function (PSF). This approach also allows the use of the uncorrupted theoretical
contrast as ground truth, which is only available during the simulation. By
using two different encoding schemes in our sequences, we gain additional
information from the two different contrasts PD and T2 that are used as input
for the CNN and both will have different PSFs, thus provide valuable
information for the SR task. Both sequences are optimized jointly to allow
generation of optimal contrasts for the SR task of the neural network. Methods
A single-shot 2D TSE sequence is being used as default sequence for our optimization. The acquisition time for the single-shot 2D TSE is 0.76 s at 1.56 mm in-plane resolution. The refocusing RF pulses of the PDw TSE with TE=12 ms and T2w TSE with TE=96 ms were optimized jointly. All simulations and optimizations were performed in a fully differentiable Bloch simulation framework2.The forward simulation outputs the TSE signal which is conventionally reconstructed to magnitude images, and in addition the corresponding contrast as ground truth target. The entire process – MRI sequence, reconstruction, and evaluation – is modelled as one computational chain and is part of the forward and backwards propagation during the optimization, as depicted in Figure 1.
After the optimization process, all sequences were exported using
the Pulseq standard3 and the pypulseq tool4. Pulseq files could then be
interpreted on a real MRI scanner including all necessary safety checks, and
were executed on a 3T scanner using a 20-channel head coil. As high-resolution reference a vendor-provided TSE
sequence was acquired with following parameters: 32-shot segmented, GRAPPA2,
TE=12/96 ms, TR=12 s, FOV=200 mm×200 mm, matrix of 256×256, FA=180°. All
MRI scans were performed
after written informed consent was obtained.Results
The optimization
process can be seen in Figure 2. Starting from the initialized values, the RF
pulses converge to the optimal RF pulse train, while the NN parameters are
optimized simultaneously.
The original LR TSE image with the zero-filled image and the reconstructed SR
image are compared to our optimized RF pulse train design and a conventional
180° RF pulse train TSE sequence for each contrast in Figure 3. It can be
observed, that in all cases the SR image leads to an improvement over the LR
TSE image by showing clearer resolved borders between white and gray matter.
The optimized RF pulse train further improves the nominal resolution, which can
be observed by a clear increase of sharpness of the sulcus between Gyrus
cinguli and Gyrus frontalis superior as indicated by the red arrows. The
optimized sequence and CNN translate well to in vivo measurements, where
similar improvements as seen in the simulated images can be observed (Figure
3d,e). Further, patient data have been acquired, three different pathologies
with multiple sclerosis, glioblastoma and cavernoma. The PDw and T2w SR image of the
patients are shown in Figure 4 with the red arrow indicating area of interest, respectively.
The SR image demonstrates shaper edges of the lesion with clearer boundaries to
the surrounding white matter.Discussion
We demonstrated a new end-to-end learning process for TSE
super-resolution by jointly optimizing refocusing RF pulse trains and neural
network parameters. This approach utilizes a differentiable MR physics
simulation embedded in the forward and backward propagation. The
joint-optimization outperforms a pure neural network training. Although our
approach is solely based on simulated data, the optimized sequence and trained
CNN translate well to in vivo data. Our approach is compatible with any network architecture e.g. 5-7 to further improve the SR task. By using simulation-based training data, we
are able to use the theoretical uncorrupted contrast as ground truth target. Ground truth HR images measured in vivo are difficult to acquire due to the
longer scan time. Motion artifacts become more significant and to acquire the
same contrast as the LR counterpart the bandwidth has to be increased, leading to a decrease of SNR8. Conclusion
We propose an end-to-end optimization of MR sequence
and neural network parameters for TSE super-resolution. This flexible and
general end-to-end approach benefits from a MR physics informed training
procedure, allowing a simple target-based problem formulation, and outperforms
pure neural network training.Acknowledgements
No acknowledgement found.References
- Maier,
A.K. et al.: Learning with known operators reduces maximum error bounds. Nature
Machine Intelligence 2019 1:8. 1, 373–380 (2019).
- Loktyushin,
A., et al. "MRzero‐Automated discovery of MRI sequences using supervised
learning." Magn Reson Med 86.2 (2021): 709-724.
- Layton,
K.J. et al.: Pulseq: A rapid and hardware-independent pulse sequence
prototyping framework. Magn Reson Med. 77, 1544–1552 (2017).
- Ravi, Keerthi, Sairam Geethanath, and John Vaughan.
"PyPulseq: A Python Package for MRI Pulse Sequence Design." Journal of
Open Source Software 4.42 (2019): 1725.
- Li,
G., Lyu, J., Wang, C., Dou, Q., Qin, J.: WavTrans: Synergizing Wavelet
and Cross-Attention Transformer for Multi-contrast MRI
Super-Resolution. Lecture Notes in Computer Science (including subseries
Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics).
13436 LNCS, 463–473 (2022).
- Sui,
Y., Afacan, O., Gholipour, A., Warfield, S.K.: MRI Super-Resolution Through
Generative Degradation Learning. Lecture Notes in Computer Science (including
subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics). 12906 LNCS, 430–440 (2021).
- Lyu,
Q., Shan, H., Wang, G.: MRI Super-Resolution with Ensemble Learning and
Complementary Priors. IEEE Trans Comput Imaging. 6, 615–624 (2019).
- Portnoy,
S., Kale, S.C., Feintuch, A., Tardif, C., Pike, G.B., Henkelman, R.M.: Information
content of SNR/resolution trade-offs in three-dimensional magnetic resonance
imaging. Med Phys. 36, 1442–1451 (2009).