Andrew Dupuis1, Yong Chen2, Mark A Griswold1,2, and Rasim Boyacioglu2
1Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, 2Radiology, School of Medicine, Case Western Reserve University, Cleveland, OH, United States
Synopsis
Motivation:
Address the challenge of integrating MRF data into existing MRI
analysis via synthetic images without introducing spatial artifacts or
hallucinations possible with CNNs.
Goal(s):
Generate static lookup tables (LUTs) mapping from T1/T2 value space
directly to grayscale visualizations matching clinical contrasts.
Approach:
A simple pixel-wise regression network was trained on a public dataset
of MRF data and weighted images. Static LUTs were generated from
dictionaries of T1/T2 combinations, then applied to MRF-derived maps for
visualization and processing via FSL.
Results:
Successful generation of synthetic contrast LUTs ensures
reproducibility and allows instantaneous visualization or registration
of MRF maps in a more conventional grayscale format.
Impact:
Integration of MRF into traditional analysis pipelines suffers because
quantitative maps have inherently different contrasts from weighted
images. LUTs for instant, deterministic generation of weighted contrasts
from T1/T2 maps allow for direct use of tools like FSL with MRF data.
Introduction
Magnetic
Resonance Fingerprinting (MRF) enables precise quantitative mapping of
T1 and T2 relaxation times in an efficient single scan [1]. However, the
integration of MRF-derived data into traditional MRI analysis pipelines
remains a challenge due to the inherently different contrast
characteristics of quantitative maps versus weighted images.
Contrast
synthesis often consists of direct Bloch simulation of a desired
contrast state or utilizes convolutional [2,3] or patch-based [4] neural
networks for image-to-image translation tasks. While CNNs can produce
high-fidelity results [5], they can also be susceptible to hallucination
artifacts [6], where unintended features or spatial errors are
introduced. Instead, we propose a pixel-wise regression network that,
after training, behaves akin to a colormap through use of a 2D lookup
table (LUT) for MRF T1 and T2 maps.Methods
A
public multimodal brain imaging dataset [7] consisting of MRF maps,
T1-weighted (T1w) Magnetization-Prepared Rapid Gradient Echo (MPRAGE)
images, and T2-weighted (T2w) Turbo Spin Echo (TSE) images for 10
healthy volunteers was used to train our regression network. All code
for our lookup table system, regression network design, training and
inference is publicly available [cite] under a research license. We
utilize TensorFlow and Keras for model training. A fully connected
network model was defined, consisting of a normalization layer, a dense
hidden layer with 64 nodes, and an output layer. Training data was
masked to remove voxels outside of the skull and weighted images were
oriented to match MRF maps based on NIFTI headers. For each MRF dataset,
an initial Bloch-simulation-based synthetic image matching the target
qualitative dataset was generated and a coarse linear registration was
performed using ITK. Target images were normalized by their maximum
value to improve training stability.
Voxel data from 8 subjects
was linearized, then split into training (80%) and testing (20%) sets
before training with a Mean Absolute Error (MAE) loss function ensuring
the predicted synthetic images closely matched the normalized intensity
values of the weighted images. Model structure and weights were
archived, but the primary training output is a static two-dimensional
LUT with input indices of T1 and T2 times in 1ms steps across a range of
values [T1:100-5000ms, T2: 1-500ms] generated by running all
combinations within the dictionary through the inference network. The
resulting LUT removes any dependency on Tensorflow from our
reconstruction environment, ensuring that contrast synthesis performance
is consistent and repeatable regardless of future changes to the
inference package used.
Synthetic T1w and T2w images were
generated for the 2 remaining subjects to validate the performance of
the generated LUTs. Additionally, in-vivo MRF datasets were acquired for
two new subjects and reconstructed with dictionary-space quadratic
interpolation in the pattern matching step. Finally, FSL was used to
compare brain extraction performance using T1/T2 maps alone against
LUT-derived synthetic images.Results
Figure
2 demonstrates 6 different 2D T1/T2 LUT’s, including linear T1 and T2
intensity profiles as well as Bloch-derived and regression-derived
MPRAGE and TSE profiles. Figure 3 demonstrates the results of applying
our regression-generated LUT profiles to a subject’s T1/T2 map pair in
comparison to clinical MPRAGE and TSE contrasts. A dataset that had been
reconstructed with a very coarse (5% progressive step-size) MRF
dictionary was used to demonstrate the discretization artifacts that can
occur if excessive discretization is present in the input quantitative
maps. Figure 4 shows the improved fidelity in synthetic images as a
result of using quadratic interpolation in T1/T2 space during pattern
matching to reduce discretization artifacts. Figure 5 demonstrates the
performance of FSL’s BET brain extraction tool when given T1 maps alone,
T2 maps alone, or synthetic MPRAGE created by applying the appropriate
LUT to the same input data prior to running BET.Discussion
To
address the limitations of traditional CNN-based approaches, we
implemented a pixel-wise regression network that takes MRF-derived T1
and T2 maps as inputs and undergoes supervised training against
co-registered MPRAGE or TSE images acquired during the same scanning
session. The training process creates a unique mapping from quantitative
map value tensors to specific intensities in a target qualitative
contrast.
After training and 2D lLUT generation, synthetic
contrasts can be generated instantly. Overall, contrast similarity
versus the target qualitative modalities is adequate for use in existing
software tools for co-registration, skull stripping, or similar as
shown in Figure 5. Importantly, since contrast performance of an
exported lookup table is deterministic, the reproducibility of output
synthetic contrasts is identical to the reproducibility of the MRF maps
used in the generation process.Acknowledgements
No acknowledgement found.References
[1]
Ma, D., Gulani, V., Seiberlich, N., Liu, K., Sunshine, J. L., Duerk, J.
L., & Griswold, M. A. (2013). Magnetic resonance fingerprinting.
Nature, 495(7440), 187-192.
[2] Virtue, P., Tamir, J. I., Doneva, M.,
Yu, S. X., & Lustig, M. (2018). Learning contrast synthesis from MR
fingerprinting. In Proc. 26th Annu. Meeting (ISMRM). icsi. berkeley.
edu (Vol. 676).
[3] Nykänen, O., Nevalainen, M., Casula, V., Isosalo,
A., Inkinen, S. I., Nikki, M., ... & Nieminen, M. T. (2023).
Deep‐Learning‐Based Contrast Synthesis From MRF Parameter Maps in the
Knee Joint. Journal of Magnetic Resonance Imaging, 58(2), 559-568.
[4]
Zhang, X., Flassbeck, S., & Assländer, J. MRI contrast synthesis
from low-rank coefficient images. Proceedings of the ISMRM Annual
Meeting, 2022
[5] Chaudhari, A. S., Sandino, C. M., Cole, E. K.,
Larson, D. B., Gold, G. E., Vasanawala, S. S., ... & Langlotz, C. P.
(2021). Prospective deployment of deep learning in MRI: a framework for
important considerations, challenges, and recommendations for best
practices. Journal of Magnetic Resonance Imaging, 54(2), 357-371.
[6]
Muckley, M. J., Riemenschneider, B., Radmanesh, A., Kim, S., Jeong, G.,
Ko, J., ... & Knoll, F. (2020). State-of-the-art machine learning
MRI reconstruction in 2020: Results of the second fastMRI challenge.
arXiv preprint arXiv:2012.06318, 2(6), 7.
[7] Dupuis, A., Chen, Y.,
Hansen, M., Chow, K., Sun, J. E. P., Badve, C., Ma, D., Griswold, M. A.,
& Boyacioglu, R. (2023). Intrasession, Intersession, and
Interscanner Qualitative and Quantitative MRI Datasets of Healthy Brains
at 3.0T (0.1.0) [Data set]. Zenodo.
https://doi.org/10.5281/zenodo.8183344