Ho-Joon Lee1, Yeonah Kang1, Marc Lebel2, Min Soo Park3, and Joonsung Lee3
1Department of Radiology, Haeundae Paik Hospital, Busan, Republic of Korea, 2MR Collaboration and Development, GE Healthcare, Calagary, AB, Canada, 3GE Healthcare Korea, Seoul, Republic of Korea
Synopsis
Sequences that allow volumetric parametric maps are being developed but when these techniques will be available clinically is uncertain. MDME sequence is a robust method, allowing rapid acquisition of T1, T2 relaxation times, PD from which one can generate synthetic multi-contrast images. However, acquisition at thin slices is challenging. With a deep learned reconstruction method, we demonstrate that interleaved thin slice acquisition of MDME, can produce quasi-volumetric synthetic MRI at an isotropic resolution. Limitations include motion misregistrations between acquisitions, parallel imaging/partial volume/pulsation related artifacts, which we believe can be overcome with technical development.
Introduction
With the accumulation of evidence regarding quantitative MR imaging biomarkers, and increasing use of volumetric acquisitions for brain imaging, sequences that can yield volumetric multi-parametric maps are being introduced1-2 and validated3-4. However, these techniques are currently based on research use and availability is limited.
A 2D multi-dynamic multi-echo (MDME)5 sequence for rapid simultaneous measurement of T1 and T2 relaxation times and proton density (PD), with correction of B1 field inhomogeneity, was previously proposed and was shown to be overall robust even across different scanners6. In addition, brain tissue volumes, and myelin volume fraction can be calculated. However, thin slice acquisition below 2 mm has not been widely adopted, presumably due to low SNR. Also, due to the cross talks, an inter-slice gap is recommended for MDME acquisition, thus continuous acquisition is not achievable in a single sequence.
DLRecon is a new deep learning-based MR reconstruction, which comprises a deep convolutional residual encoder network trained using a database of over 10,000 images to achieve images with high SNR and high spatial resolution, and can enable thin slice MDME acquisitions.
In this preliminary evaluation, we demonstrate a quasi-volumetric synthetic MRI method based on deep learned reconstruction of slice-interleaved MDME acquisition. Methods
[Data Acquisition]
Images were acquired on a 3T MRI scanner
(Signa Architect, GE Healthcare,
Waukesha, WI, USA), equipped with a 48-channel head coil.
The MDME sequence was acquired in the axial
plane, with the following parameters: TR 6261 msec, TE 16.2/81.1 msec, FOV 208 × 177 mm, matrix size 160 x 160, slice thickness 1.3 mm with 1.3 mm
interslice gap, ASSET factor 3, saturation delays calculated automatically). The
acquisition was repeated with imaging slab shifted by 1.3 mm leading to an acquisition
time of 4:36 x 2 (9:12). Each acquisition produced eight real and imaginary
image pairs with different echo time and saturation delay combinations.
[Image Reconstruction]
The
acquired data were retrospectively reconstructed with and without DLRecon at an
empirical denoising level of 70% (DL70). T2, PD, T1FLAIR and PSIR contrasts were
synthesized from the source images using software installed on the console
(MAGIC, GE Healthcare, Waukesha,
WI, USA).
[Post-processing]
With in-house-codes, interleaved images were combined and interslice variation of image intensity due to interleaved acquisition was corrected by a joint entropy-based weighted least squares estimation method7.Results
While the images based on conventional
reconstruction appear coarse, images synthesized from DL70 inputs show reduced
noise (Figure 2.).
When reformatted in the coronal or sagittal
plane, there are ‘even-odd’ effects. These are reduced by the inter-slice intensity
variation reduction method (Figure 3.).
Reformats of the synthetic images can be achieved at an acceptable quality by the application of both DL70 and intensity standardization
(Figure 4.). Discussion and Conclusion
There are still several limitations to this
approach. First, registration between the interleaved stacks needs to be implemented
to account for motion. Second, due to the high ASSET factor, there parallel
imaging-related artifacts may be present at the mesial portion of the brain. Third,
artifacts on FLAIR related to the partial volume are still present even at a thin
slice (Figure 5.). Last but not least, the scan time, which takes approximately
9 minutes, is relatively long.
These challenges may be overcome by further
developments including implementation of motion correction8-9, methods to decrease artifacts10-12, and methods for further acceleration.
In conclusion, we have demonstrated a method utilizing
deep learned reconstruction to acquire quasi-volumetric synthetic MRI of the brain, which have the potential for clinical application. Acknowledgements
We thank Sang Hyuk Park for scanner arrangements. References
- Ma, D. et al. Development of
high-resolution 3D MR fingerprinting for detection and characterization of
epileptic lesions. J. Magn. Reson. Imaging 49, 1333–1346 (2019).
- Cheng, C. C., Preiswerk, F., Hoge, W. S.,
Kuo, T. H. & Madore, B. Multipathway multi-echo (MPME) imaging: all main MR
parameters mapped based on a single 3D scan. Magn. Reson. Med. 81,
1699–1713 (2019).
- Fujita, S. et al.
Three-dimensional high-resolution simultaneous quantitative mapping of the
whole brain with 3D-QALAS: An accuracy and repeatability study. Magn. Reson.
Imaging 63, 235–243 (2019).
-
Fujita, S. et al. 3D quantitative
synthetic MRI‐derived cortical thickness and subcortical brain volumes: Scan–rescan
repeatability and comparison with conventional T 1 ‐weighted images.
J. Magn. Reson. Imaging jmri.26744 (2019) doi:10.1002/jmri.26744.
- Milshteyn, E. et al. Development of
high resolution 3D hyperpolarized carbon-13 MR molecular imaging techniques. Magn.
Reson. Imaging 38, 152–162 (2017).
- Hagiwara, A. et al. Linearity,
Bias, Intrascanner Repeatability, and Interscanner Reproducibility of
Quantitative Multidynamic Multiecho Sequence for Rapid Simultaneous Relaxometry
at 3 T. Invest. Radiol. 54, 39–47 (2019).
- Schmidt, M. A Method for Standardizing MR
Intensities between Slices and Volumes. (2005) doi:10.7939/R3R49GG8D.
- Rohlfing, T., Rademacher, M. H. &
Pfefferbaum, A. Volume reconstruction by inverse interpolation: Application to
interleaved MR motion correction. in Lecture Notes in Computer Science
(including subseries Lecture Notes in Artificial Intelligence and Lecture Notes
in Bioinformatics) vol. 5241 LNCS 798–806 (2008).
- Nam, H., Lee, Y. J., Jeong, B., Park, H.
J. & Yoon, J. Motion correction of magnetic resonance imaging data by using
adaptive moving least squares method. Magn. Reson. Imaging 33,
659–670 (2015).
- Hagiwara, A. et al. Improving the
quality of synthetic FLAIR images with deep learning using a conditional
generative adversarial network for pixel-by-pixel image translation. Am. J.
Neuroradiol. 40, 224–230 (2019).
- Ryu, K. et al. Data‐driven synthetic
MRI FLAIR artifact correction via deep neural network. J. Magn. Reson.
Imaging 50, 1413–1423 (2019).
- Eo, T. et al. KIKI-net: cross-domain
convolutional neural networks for reconstructing undersampled magnetic
resonance images. Magn. Reson. Med. 80, 2188–2201 (2018).