Daniel Polak1,2,3, Stephen Cauley2,4,5, Berkin Bilgic2,4,5, Daniel Nicolas Splitthoff3, Peter Bachert1, Lawrence L. Wald2,4,5, and Kawin Setsompop2,4,5
1Department of Physics and Astronomy, Heidelberg University, Heidelberg, Germany, 2Department of Radiology, A. A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, 3Siemens Healthcare GmbH, Erlangen, Germany, 4Department of Radiology, Harvard Medical School, Boston, MA, United States, 5Harvard-MIT Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States
Synopsis
Navigation-free retrospective motion-correction
typically requires estimating hundreds of coupled temporal motion parameters by
solving a large non-linear inverse problem. This can be extremely demanding computationally,
which has impeded implementation/adoption in clinical settings. We propose a
technique that utilizes a single rapid
scout scan (Tadd=3sec) to drastically reduce the computation cost of
this motion-estimation and create a pathway for clinical acceptance. We
optimized this scout along with the sequence acquisition reordering in a 3D Turbo-Spin-Echo
acquisition. Our approach was evaluated in-vivo with up to R=6-fold
acceleration and robust motion-mitigation was achieved using a scout with
differing contrast to the imaging sequence.
Introduction
Navigator-free retrospective motion correction
techniques [1]–[5] often perform image reconstruction
by solving
$$\min_\vec{\theta}\min_x\|\sum_iM_iFCT_{\theta_i}R_{\theta_i}x-k_i\|_2\qquad\quad[1]$$ where $$$k_i$$$ denotes the multi-channel k-space data of shot $$$i$$$, $$$x$$$ the image, $$$T_{\theta_i}R_{\theta_i}$$$ the shot-dependent translation and rotation motions, $$$C$$$ the coil sensitivity, $$$F$$$ the Fourier transformation and $$$M_i$$$ the undersampling mask. All shots are coupled
in this inverse problem and the non-convex estimation of several hundreds of temporal motion
parameters is computationally demanding.
Substantial speed-up was demonstrated in NAMER [6], which can solve for the motion
parameters in each shot separately using a good initial image estimate $$$\hat{x}$$$,
obtained via e.g. Machine Learning.
$$\min_{\vec{\theta}_i}\|M_iFCT_{\theta_i}R_{\theta_i}\hat{x}-k_i\|_2\qquad\qquad\quad[2]$$ However, for this method to work robustly, large
amounts of training data (motion-corrupted & motion-free) are needed to
enable generalizability
to arbitrary motion patterns.
Moreover, repeated updates of the motion parameters $$$\vec{\theta}$$$ are typically required to refine the Machine Learning
generated image $$$\hat{x}$$$ and arrive at the correct motion estimate.
In this work, we propose an alternative
strategy, termed SAME, which does not rely on pre-trained Machine Learning networks.
Inspired by traditional navigator-based motion correction techniques [7]–[9], SAME estimates the motion
parameters quickly using a single low-resolution
scout scan $$$\hat{x}$$$. We optimized this scan along with
the sequence acquisition reordering of our target 3D Turbo-Spin-Echo
acquisition, to achieve accurate motion estimation while minimizing added scan
time (Tadd=3sec). We then
demonstrate motion-mitigation capability of this approach in accelerated 3D-TSE
scans with up to R=6-fold acceleration and achieve motion-robust rapid imaging via
Wave-encoding.
Methods
Sequence ordering: We optimized the encoding reordering
of our prototype T2w SPACE sequence to improve the robustness of its
motion estimation. To test different reorderings, motion-free
data with linear and two types of radial reorderings (with VDS-Poisson) were
acquired at 1 mm isotropic resolution and R=4 acceleration using a 3T scanner (MAGNETOM
Vida, Siemens Healthcare, Erlangen, Germany) and 64-channel head coil. After
instructed subject motion, a separate motion-free scout scan $$$\hat{x}$$$ was acquired (same protocol) and
used to estimate inter-scan motion by
minimizing Eq. 2 (‘fminunc’ in MATLAB). Motion estimation with respect to
the scout scan was performed for each shot (TR) within the T2w SPACE
acquisition which should lead to identical motion-estimates across all TRs.
Scout scan: We optimized the scout acquisition to
minimize added scan time. Specifically, T2w scouts with&without
dummy shots were acquired at 1x4x4 mm3 resolution and R=9-fold acceleration, achieved
with Wave-CAIPI [10]. We then estimated inter-scan motion
parameters between the scout and
imaging sequence and analyzed
how image contrast differences between these scans (due to differences in acquisition
reordering and the absence of dummy-shots) influence motion estimation accuracy.
Inter-shot motion experiments:
Based on the
findings from above, we utilized radial reordering for all motion corrupted in-vivo
acquisitions and added a rapid R=9 Wave-scout prior to each imaging sequence (Tadd=3sec).
For T2w SPACE, acquisitions were performed at R=4 and R=6-fold acceleration
(VDS-Poisson, with&without Wave). For motion-corrupted FLAIR-SPACE (R=6, Wave)
we also assessed the motion estimation provided by either a T2w or
FLAIR-weighted scout scan. Results
Linear reordering does not provide enough k-space
overlap between the low-resolution scout scan (yellow) and the shot data (white), resulting
in poor motion parameter estimates (Fig. 1a). Better results were obtained for
radial reordering (one spoke), however best agreement with the ground truth was
observed for two radial spokes. Removing dummy shots in the acquisition of our
low-resolution scout caused substantial contrast differences (Fig. 1b).
This, however, did not seem to affect the motion estimation.
Figure 2 shows the inter-shot motion results for
T2w SPACE at R=4-fold acceleration. The SENSE reconstruction
demonstrates substantial motion artifacts and loss of contrast/resolution, which
was largely mitigated using SAME. The fully-separable motion-estimation took TØ=91sec per shot using unoptimized MATLAB-code which is readily parallelizable.
Figure 3 demonstrates the benefit of integrating
Wave-encoding into the imaging sequence. At R=6, Wave-encoding better suppressed
noise amplification (zoom-in) and achieved good image quality with substantial
reduction of motion artifacts. The motion-estimation on Wave data took TØ=2:53min per shot (MATLAB), again readily parallelizable.
In Fig. 4, we tested our approach for FLAIR-SPACE
(R=6, Wave). As evident by the SAME reconstructions and motion trajectory
plots, both T2w and FLAIR scout acquisitions enabled accurate motion
estimation for this FLAIR-SPACE dataset.Discussion
In this work, we proposed a novel motion-correction
technique for 3D acquisitions where the usage of a rapid scout dramatically
reduced the computational footprint of the motion parameter estimation. We
evaluated this technique in-vivo on T2w SPACE data and demonstrated robust
motion estimation and correction without requiring any updated reconstruction
of the imaging volume during motion-estimation [5],[6]. This enables parallelization
across all shots pointing to good potential for online reconstruction with
clinically acceptable reconstruction time (e.g., T=91sec*Nshots/Nthreads
in MATLAB). Moreover, we showed the benefits of Wave-encoding in
highly-accelerated scout and imaging acquisitions. We also tested our approach for
FLAIR-SPACE where robustness to contrast variations between the scout and
imaging sequence was observed. This may eliminate the need for scout re-acquisition
across the clinical exam, where several imaging contrasts are acquired. Future
work will examine potential improvements from using more incoherent sampling
orders [11] and adoption into other 3D Wave
sequences (e.g., MPRAGE [12], GRE [10]) which could enable a rapid
motion-robust 3D brain exam [13].Acknowledgements
This
work was supported in part by NIH research grants: R01EB020613, R01EB019437,
R01 MH116173, P41EB015896, U01EB025162 and the shared instrumentation grants:
S10RR023401, S10RR019307, S10RR019254, S10RR023043.References
[1] M. W. Haskell, S. F. Cauley, and L. L.
Wald, “TArgeted Motion Estimation and Reduction (TAMER): Data consistency based
motion mitigation for mri using a reduced model joint optimization,” IEEE
Trans. Med. Imaging, vol. 37, no. 5, pp. 1253–1265, 2018.
[2] A.
Loktyushin, H. Nickisch, R. Pohmann, and B. Schölkopf, “Blind multirigid
retrospective motion correction of MR images,” Magn. Reson. Med., 2015.
[3] L.
Cordero-Grande, R. P. A. G. Teixeira, E. J. Hughes, J. Hutter, A. N. Price, and
J. V. Hajnal, “Sensitivity Encoding for Aligned Multishot Magnetic Resonance
Reconstruction,” IEEE Trans. Comput. Imaging, 2016.
[4] D.
Atkinson, D. L. G. Hill, P. N. R. Stoyle, P. E. Summers, and S. F. Keevil,
“Automatic correction of motion artifacts in magnetic resonance images using an
entropy focus criterion,” IEEE Trans. Med. Imaging, 1997.
[5] L.
Cordero-Grande, E. J. Hughes, J. Hutter, A. N. Price, and J. V. Hajnal,
“Three-dimensional motion corrected sensitivity encoding reconstruction for
multi-shot multi-slice MRI: Application to neonatal brain imaging,” Magn.
Reson. Med., 2018.
[6] M. W.
Haskell et al., “Network Accelerated Motion Estimation and Reduction
(NAMER): Convolutional neural network guided retrospective motion correction
using a separable motion model,” Magn. Reson. Med., vol. 82, no. 4, pp.
1452–1461, 2019.
[7] D.
Gallichan, J. P. Marques, and R. Gruetter, “Retrospective correction of
involuntary microscopic head movement using highly accelerated fat image
navigators (3D FatNavs) at 7T,” Magn. Reson. Med., 2016.
[8] R.
Bammer, M. Aksoy, and C. Liu, “Augmented generalized SENSE reconstruction to
correct for rigid body motion,” Magn. Reson. Med., 2007.
[9] M. D.
Tisdall, A. T. Hess, M. Reuter, E. M. Meintjes, B. Fischl, and A. J. W. Van Der
Kouwe, “Volumetric navigators for prospective motion correction and selective
reacquisition in neuroanatomical MRI,” Magnetic Resonance in Medicine.
2012.
[10] B.
Bilgic et al., “Wave-CAIPI for highly accelerated 3D imaging,” Magn.
Reson. Med., vol. 73, no. 6, pp. 2152–2162, 2015.
[11] L.
Cordero-Grande, G. Ferrazzi, R. P. A. G. Teixeira, J. O’Muircheartaigh, A. N.
Price, and J. V. Hajnal, “Motion corrected MRI with DISORDER: Distributed and
Incoherent Sample Orders for Reconstruction Deblurring using Encoding
Redundancy,” Oct. 2019.
[12] D. Polak
et al., “Wave-CAIPI for highly accelerated MP-RAGE imaging,” Magn.
Reson. Med., vol. 79, no. 1, pp. 401–406, 2018.
[13] D. Polak
et al., “Highly-accelerated volumetric brain protocol using optimized
wave-CAIPI encoding,” in Proc. Intl. Soc. Mag. Res. Med., 2018, p. 0937.