Sandeep Ganji1,2, Brian Johnson3,4, John Penatzer3, and Johannes M. Peeters5
1MR R&D, Philips Healthcare, Rochestr, MN, United States, 2Mayo Clinic, Rochester, MN, United States, 3Philips Healthcare, Gainesville, FL, United States, 4University of Texas Southwestern Medical Center, Dallas, TX, United States, 5Philips Healthcare, Eindhoven, Netherlands
Synopsis
Despite long scan times, 3D T1-weighted (T1w) MRI is routinely
used for MRI studies to provide high resolution structural and volumetric
information of brain. Volumetric analysis can serve as a biomarker and aid in
clinical diagnosis of certain diseases such as Alzheimer’s, mild cognitive
impairment, and atrophy, however, the standardized 3D T1-weighted scans suffer
from long acquisition times well over 5 minutes. We compared
the results of volumetric brain analysis for 3D T1w images acquired over a
range of compressed SENSE acceleration factors with and without Adaptive-CS-Net
reconstruction against a standard clinical 3D T1w MRI protocol.
Introduction
Despite long scan times, 3D T1-weighted (T1w) MRI is routinely used
for MRI studies to provide high resolution structural and volumetric
information of brain. Volumetric analysis can serve as a biomarker and aid in
clinical diagnosis of certain diseases such as Alzheimer’s, mild cognitive impairment,
and atrophy, however, the standardized 3D T1-weighted scans suffer from long
acquisition times well over 5 minutes. Recent advances in image acceleration
and deep learning-based reconstruction (1 - 5) provide promise for drastically
decreasing acquisition times while maintaining image quality for accurate diagnosis.
Prior studies have already shown the utility of compressed SENSE and deep-learning
based acceleration (1, 5 - 6). Here we compare the results of volumetric brain
analysis for 3D T1w images acquired over a range of compressed SENSE acceleration
factors with and without Adaptive-CS-Net reconstruction (6) against a standard clinical
3D T1w MRI protocol.Methods
Five healthy male subjects (38±6 years) were scanned at 3T
(Philips Elition X, Philips Healthcare, Best, Netherlands) using a 32-channel
head coil. Standard clinical protocol 3D T1w images were acquired using a FOV=256
x 256 x 192 mm3 (160 slices at acquisition resolution of 1.2 x 1.2 x
1.2 mm3) with a SENSE acceleration factor of 1.8 (along the RL
direction). Using the same resolution, additional 3D T1w images were acquired
with increasing compressed SENSE acceleration factors of 1.8 up to 14.4. More details
about the MRI acquisition protocol can be found in Figure 1. Scans using compressed
SENSE were reconstructed using vendor provided compressed reconstruction on the
scanner and with Adaptive-CS-Net artificial intelligence framework (6). Volumetric
segmentation was performed using FastSurfer (7) and FreeSurfer (8) software
tools, which are documented and freely available for download online (https://github.com/Deep-MI/FastSurfer
and http://surfer.nmr.mgh.harvard.edu/, respectively).Results
All acquired images and accelerations were able to be successfully
post-processed using the FastSurfer and FreeSurfer software tools and volumetric
data were generated. Figure 2 shows a sample of the image quality generated as
a function of increasing compressed SENSE acceleration factors, including the
clinical standard image. Visual inspection images showed higher noise as the compressed
SENSE acceleration increased, however with Adaptive-CS-Net reconstruction some
of the noise was reduced. Contrasts for signal to noise and artifacts show
adequate image quality up to an acceleration factor of 14.4 compressed SENSE
acceleration factor. Figure 3 shows the volumetric data from few selected brain
regions, in a Bland-Altman plot, using standard clinical protocol 3D T1w with
SENSE acceleration factor of 1.8 as reference. Overall, the volumetric variation
with increasing compressed SENSE acceleration factors was within 5% the standard
clinical protocol 3D T1w. Moreover, the consistency for large regions volumes,
such as cerebral-white matter and hippocampus was much narrower compared to small
regions, as expected. The heatmap of the averaged data from 5 subjects shown in
Figure 4, demonstrates the variation with respect to standard clinical protocol
3D T1w with SENSE. The deviation from the standard clinical protocol 3D T1w values
increases only at compressed SENSE acceleration factor of 14.4.Discussion
Here we applied compressed SENSE with factors up to 14.4 acquire 3D
T1w images to evaluate the effect on volumetric analysis. Applying an compressed
SENSE acceleration factor of 7.2 produced a 1 minute and 12 seconds scan time
with no evidence of significant differences in volumetric results. Pushing to a
compressed SENSE acceleration to 14.4 yielded generated image contrasts with a
higher noise profile but still clinically diagnostic for a scan time of only 37
seconds. Even at compressed SENSE
acceleration factor of 14.4, the volumetric results were under 5% different
when compared to the volumetric results standard clinical protocol 3D T1w
images. Images acquired with compressed
SENSE acceleration factors with and without Adaptive-CS-Net reconstruction showed
slight variation, which may need to be further investigated in a larger study.Conclusion
Compressed SENSE based acceleration can be used to achieve
clinically viable scan times while maintaining image quality and providing the volumetric
information within 5% of the values of the standard clinical protocol 3D T1w.
Drastically lowering the acquisition time to scans below 2-minutes could lead
to more widespread clinical adoption of incorporating volumetric based measurements.Acknowledgements
No acknowledgement found.References
- Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of
compressed sensing for rapid MR imaging. Magn Reson Med. 2007
Dec;58(6):1182-95. doi: 10.1002/mrm.21391. PMID: 17969013.
- Wang S, Su Z, Ying L, Peng X, Zhu S, Liang F, et al.,
“Accelerating Magnetic Resonance Imaging via Deep Learning,” IEEE Conference on
International Symposium on Biomedical Imaging (ISBI), pp. 514–517, 2016.
- Liang D, Cheng J, Ke Z, Ying L. Deep Magnetic Resonance Image Reconstruction:
Inverse Problems Meet Neural Networks. IEEE Signal Process Mag.
2020;37(1):141-151. doi:10.1109/MSP.2019.2950557
-
T. M. Quan, T. Nguyen-Duc and W.-K. Jeong, "Compressed
sensing MRI reconstruction using a generative adversarial network with a cyclic
loss", IEEE Trans. Med. Imag., vol. 37, no. 6, pp. 1488-1497, Jun. 2018.
- M. Mardani, E. Gong, J. Y. Cheng, S. S. Vasanawala, G. Zaharchuk,
L. Xing, et al., "Deep generative adversarial neural networks for
compressive sensing MRI", IEEE Trans. Med. Imag., vol. 38, no. 1, pp.
167-179, Jan. 2019.
- N. Pezzotti et al., "An Adaptive Intelligence Algorithm for
Undersampled Knee MRI Reconstruction," in IEEE Access, vol. 8, pp.
204825-204838, 2020, doi: 10.1109/ACCESS.2020.3034287.
- Henschel L, Conjeti S, Estrada S, Diers K, Fischl B, Reuter M,
FastSurfer - A fast and accurate deep learning based neuroimaging pipeline,
NeuroImage 219 (2020), 117012. https://doi.org/10.1016/j.neuroimage.2020.117012
- Dale AM, Fischl B, Sereno MI.
Cortical surface-based analysis. I. Segmentation and surface reconstruction.
Neuroimage. 1999 Feb;9(2):179-94. doi: 10.1006/nimg.1998.0395. PMID: 9931268.