Hélène Lajous1,2, Tom Hilbert1,3,4, Christopher W. Roy1, Sébastien Tourbier1, Priscille de Dumast1,2, Yasser Alemán-Gómez1, Thomas Yu4, Patric Hagmann1, Mériam Koob1, Vincent Dunet1, Tobias Kober1,3,4, Matthias Stuber1,2, and Meritxell Bach Cuadra1,2,4
1Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, 2CIBM Center for Biomedical Imaging, Lausanne, Switzerland, 3Advanced Clinical Imaging Technology (ACIT), Siemens Healthcare, Lausanne, Switzerland, 4Signal Processing Laboratory 5 (LTS5), Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
Synopsis
Accurate
characterization of in utero human brain maturation is critical. However, the
limited number of exploitable magnetic resonance acquisitions not corrupted by motion in this cohort of
sensitive subjects hinders the validation of advanced image processing
techniques. Numerical simulations can mitigate these limitations by providing a
controlled environment with a known ground truth. We present a flexible
framework that simulates magnetic resonance acquisitions of the fetal brain in
a realistic setup including stochastic motion. From simulated images
comparable to clinical acquisitions, we assess the accuracy and robustness of
super-resolution fetal brain magnetic resonance imaging with respect to noise
and motion.
Introduction
There
is a growing awareness of the importance of processes occurring during early
brain development on health later in life 1,2. Magnetic
Resonance Imaging (MRI) is a powerful tool for depicting brain tissue contrast
and therefore has the potential to investigate equivocal neurological patterns
in utero. Before birth, fast 2D spin echo sequences are typically used to minimize
the effects of unpredictable fetal motion during the acquisition. Access to large-scale
data, usually corrupted by stochastic movements of the fetus, remains relatively scarce, hampering
the development and evaluation of advanced image processing methods. Numerical phantoms
provide a controlled environment where the ground truth is known to meet a variety of challenges related to the
acquisition process for accurate, robust and reproducible research 3,4. In this context,
we have developed a simulation framework for Half-Fourier Acquisition
Single-shot Turbo spin Echo (HASTE). Since motion correction is crucial in
fetal MRI, the potential of this new tool is highlighted by an application
example: the assessment of super-resolution (SR) reconstruction techniques that
combine orthogonal series of 2D thick slices into an isotropic 3D
high-resolution volume of the fetal brain with reduced intensity artefacts and
motion sensitivity 5–9.Methods
Figure
1 provides an overview of the simulation pipeline of HASTE acquisitions
implemented using MATLAB (MathWorks, R2019a). Segmented high-resolution
anatomical images from a normative spatiotemporal MRI atlas 10 are used as
a model of the normal fetal brain. Segmented brain tissues are labeled as gray
matter, white matter or cerebrospinal fluid, and are assigned corresponding T1
and T2 relaxation times at 1.5T 11–15. The extended phase
graph (EPG) simulation 16 allows for
computing the decay of the transverse magnetization over time in every voxel of
the anatomical images from the reference T1 and T2 maps and from simulated
intensity non-uniformities 17, according to the
HASTE sequence pulse design. The Fourier transform of the resulting 4D matrix
is used to sample the actual k-space of the simulated HASTE images as described
in Figure 2. While intra-slice motion is neglected, inter-slice random rigid
movements of the fetus are implemented during k-space sampling according to
motion estimation from clinical data 18. Three levels of
motion are defined: low, moderate and strong, that are characterized by less
than 5%, 10% and 20% of corrupted slices respectively, and an amplitude of translation
in every direction of [-1, 1]mm, [-2, 2]mm and [-2, 2]mm and of rotation of
[-2, 2]°, [-4, 4]° and [-4, 4]° respectively. Complex Gaussian noise is added
to simulate thermal noise during the acquisition. The final simulated HASTE
images are reconstructed by 2D inverse Fourier transform.
HASTE
acquisitions of the fetal brain are simulated in three orthogonal orientations
with a shift of ± 1.6 mm in-plane for acquisitions in the same orientation. The
acquisition parameters comply with the clinical protocol set up at our local
hospital for fetal brain examination.
A
case study on SR is presented using a previously reported reconstruction
pipeline 8,19. The quality of
SR reconstruction is evaluated based on the normalized root mean squared error
(NRMSE), the local structural similarity (SSIM) index 20 and its mean
(MSSIM) over the image compared to a 3D 1.1-mm isotropic HASTE simulation uncorrupted
by noise or movement. Six realizations are simulated for each SNR. The impact
of low, respectively moderate movements of the fetus on the quality of SR
reconstruction is studied using a reference series without motion, respectively
with low motion amplitude.Results
Figure
3 compares simulated motion-corrupted HASTE images of the fetal brain at 26, 30
and 32 weeks of gestational age to clinical acquisitions. Figure 4 shows the
NRMSE and the MSSIM between SR reconstructions from various numbers of
low-resolution series and a 3D isotropic high-resolution reference. The NRMSE
decreases when increasing the number of series used for SR reconstruction.
Noisier images lead to a slight decrease in the MSSIM, which in turn increases
with the number of series. The quality of SR reconstructions from simulated
data with an SNR similar to that observed in clinical acquisitions is the same
as for SR reconstructions from simulated data with twice as much signal. The
addition of motion-corrupted low-resolution series to reconstruct a SR volume of the fetal brain does not increase the MSSIM.
In the case of moderate motion, the MSSIM is smaller than in the case of low
motion. Figure 5 illustrates the improved sharpness in a region-of-interest of SR
reconstruction when increasing the number of motion-corrupted series at an SNR
similar to that of clinical acquisitions.Discussion and Conclusion
We
developed a new framework for fetal brain MRI simulations and explored its
potential for evaluation and optimization of SR reconstruction techniques. This
powerful tool simulates HASTE images comparable to real clinical acquisitions.
Its flexibility in the choice of the sequence parameters but also other settings
such as the gestational age makes it possible to simulate MR images of the
fetal brain throughout the growth of the fetus with various SNR and amplitude
of fetal movements. The controlled environment implemented demonstrates that the
SR reconstruction algorithm used is robust to noise and motion. Such a
numerical phantom provides a valuable framework for reproducibility studies and validation of advanced image processing techniques.Acknowledgements
This
work was supported by the Swiss National Science Foundation through grants
141283 and 182602, the Centre d’Imagerie BioMédicale (CIBM) of the UNIL, UNIGE,
HUG, CHUV and EPFL, and the Leenaards and Jeantet Foundations.References
1. Kwon EJ, Kim YJ. What is fetal
programming?: a lifetime health is under the control of in utero health. Obstet
Gynecol Sci. 2017;60(6):506-519. doi:10/gf5fj6
2. O’Donnell KJ, Meaney MJ. Fetal origins
of mental health: The developmental origins of health and disease hypothesis. Am
J Psychiatry. 2017;174(4):319-328. doi:10/f92t6n
3. Wissmann L, Santelli C, Segars WP,
Kozerke S. MRXCAT: Realistic numerical phantoms for cardiovascular magnetic
resonance. J Cardiovasc Magn Reson. 2014;16:63. doi:10/gf6bpx
4. Roy CW, Marini D, Segars WP, Seed M,
Macgowan CK. Fetal XCMR: a numerical phantom for fetal cardiovascular magnetic
resonance imaging. Journal of Cardiovascular Magnetic Resonance.
2019;21(1):29. doi:10/gf33j3
5. Rousseau F, Kim K, Studholme C, Koob M,
Dietemann JL. On super-resolution for fetal brain MRI. International
Conference on Medical Image Computing and Computer-Assisted Intervention.
2010;13(Pt 2):355-362. doi:10/bns47p
6. Gholipour A, Estroff JA, Warfield SK.
Robust super-resolution volume reconstruction from slice acquisitions:
application to fetal brain MRI. IEEE Transactions on Medical Imaging.
2010;29(10):1739-1758. doi:10/b2xmdp
7. Kuklisova-Murgasova M, Quaghebeur G,
Rutherford MA, Hajnal JV, Schnabel JA. Reconstruction of fetal brain MRI with
intensity matching and complete outlier removal. Medical Image Analysis.
2012;16(8):1550-1564. doi:10/f4hxwz
8. Tourbier S, Bresson X, Hagmann P, Thiran
J-P, Meuli R, Cuadra MB. An efficient total variation algorithm for
super-resolution in fetal brain MRI with adaptive regularization. NeuroImage. 2015;118:584-597.
doi:10/f7p5zx
9. Kainz B, Steinberger M,
Wein W, et al. Fast volume
reconstruction from motion corrupted stacks of 2D slices. IEEE Transactions
on Medical Imaging. 2015;34(9):1901-1913. doi:10/f3svr5
10. Gholipour A, Rollins CK, Velasco-Annis C,
et al. A normative spatiotemporal MRI atlas of the fetal brain for automatic
segmentation and analysis of early brain growth. Sci Rep. 2017;7(1):476.
doi:10/gf39sn
11. Hagmann CF, De Vita E,
Bainbridge A, et al. T2 at MR imaging
is an objective quantitative measure of cerebral white matter signal intensity
abnormality in preterm infants at term-equivalent age. Radiology.
2009;252(1):209-217. doi:10/bqkd9r
12. Blazejewska AI, Seshamani S, McKown SK, et
al. 3D in utero quantification of T2* relaxation times in human fetal brain
tissues for age optimized structural and functional MRI. Magn Reson Med. 2017;78(3):909-916.
doi:10/gf2n9z
13. Vasylechko S,
Malamateniou C, Nunes RG, et al. T2*
relaxometry of fetal brain at 1.5 Tesla using a motion tolerant method. Magnetic
Resonance in Medicine. 2015;73(5):1795-1802. doi:10/gf2pbh
14. Nossin-Manor R, Card D, Morris D, et al.
Quantitative MRI in the very preterm brain: Assessing tissue organization and
myelination using magnetization transfer, diffusion tensor and T1 imaging. Neuroimage.
2013;64:505-516. doi:10/f4jgtg
15. Yarnykh VL, Prihod’ko IY, Savelov AA,
Korostyshevskaya AM. Quantitative assessment of normal fetal brain myelination
using fast macromolecular proton fraction mapping. American Journal of
Neuroradiology. 2018;39(7):1341-1348. doi:10/gdv9nf
16. Weigel M. Extended phase graphs:
Dephasing, RF pulses, and echoes - pure and simple. Journal of Magnetic
Resonance Imaging. 2015;41(2):266-295. doi:10/gghctd
17. BrainWeb: Simulated brain database.
https://brainweb.bic.mni.mcgill.ca/brainweb/
18. Oubel E, Koob M, Studholme C, Dietemann
J-L, Rousseau F. Reconstruction of scattered data in fetal diffusion MRI. Medical
image analysis. 2012;16(1):28-37. doi:10/fhbqtj
19. Tourbier S, Bresson X, Hagmann P, Meuli R,
Bach Cuadra M. Sebastientourbier/Mialsuperresolutiontoolkit: MIAL
Super-Resolution Toolkit v1.0. Zenodo; 2019. doi:10.5281/zenodo.2598448
20. Wang Z, Bovik A, Sheikh
H, Simoncelli E. Image quality assessment: From error visibility to structural
similarity. IEEE Transactions on Image Processing. 2004;13:600-612.
doi:10/c7sr27