Real-time 3D MR for Simultaneous PET-MR Motion Tracking
1. Real-time 3D MR Respiratory Motion Acquisition for Simultaneous Human PET-MR Imaging
Liheng Guo1, Mark Ahlman3, Tao Feng2, Michael Guttman4, Elliot McVeigh1, David Bluemke3, Benjamin Tsui2, Daniel Herzka1
1Biomedical Engineering, Johns Hopkins University, Baltimore, MD. 2Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University, Baltimore, MD.
3Radiology and Imaging Sciences, National Institutes of Health, Bethesda, MD. 4Cardiology, School of Medicine, Johns Hopkins University, Baltimore, MD.
References
1. Koesters, et al. Motion-compensated EM PET Reconstruction for Simultaneous PET-MR Exams. ISMRM 2014 0789.
2. Kolbitsch, et al. A 3D MR-acquisition scheme for non-rigid bulk motion correction in simultaneous PET-MR. ISMRM 2014 0788.
3. Huang, et al. A Novel Golden-Angle Radial FLASH Motion-Estimation Sequence for Simult. Thoracic PET-MR. ISMRM 2013 2684.
4. Kolbitsch, et al. T1 and T2 Weighted MR Acquisition for Bulk Motion Correction for Simultaneous PET-MR. ISMRM 2013 3745.
5. Grimm, et al. MR-PET Respiration Compensation Using Self-Gated Motion Modeling. ISMRM 2013 0829.
6. Fürst, et al. Respiratory motion compensation in PET/MR with a self-gating MR sequence. SNMMI 2013.
7. Würslin, et al. Respiratory Motion Correction in Oncologic PET Using T1-Weighted MR Imaging. J Nucl Med 54:3 (2013).
Objective
This work aims to develop a fast MR acquisition technique that can
capture 3D respiratory motion in real time during a simultaneous PET-
MR scan, providing an alternative to retrospective motion acquisition.
The resulting motion vector field is to be used in the PET image
reconstruction to correct for the respiratory motion.
Motivation
• Positron emission tomography (PET) is the primary imaging modality
for cancer imaging due to its superior sensitivity to detect tumors.
• PET images is severely degraded by respiratory motion (Fig. 1).
• Respiratory motion tracking in 3D using MRI during PET imaging is now
possible with newly available simultaneous human PET-MR scanners,
and enables correction of respiratory motion effects.
• Current motion tracking methods [1-8] uses retrospective binning
guided by the diaphragm navigator (Fig. 2,3), which assumes the
repeatability of motion and is vulnerable to irregular motion (Fig. 4).
• This work implements a fast real-time 3D acquisition on the MR that
can measure 3D motion at arbitrary time, without the use of a
navigator or assuming the repeatability of motion. The associated MR
image reconstruction exploits both the spatial and temporal sparsity
of respiratory motion.
Figure 1. Example of respiratory motion. Left:
normal torso PET image reconstructed using
attenuation correction map acquired at a
proper respiratory state. Right: respiratory
motion caused misalignment of attenuation
map, resulting in large portions of image
missing (arrows). (Adapted from: Goerres, et
al. Respiration-induced Attenuation Artifact
at PET/CT: Technical Considerations.
Radiology. 2003; 226:906–910.)
w
MRI: continuous image/motion data collection
PET: normal imaging
Figure 2. The diaphragm navigator: during a simultaneous PET-MR scan, conventional
techniques use the MR to execute the diaphragm navigator to track one point on the
lung-liver interface, providing a 1D displacement (red) that drives the retrospective
binning of continuously acquired MR data.
Ÿ Ÿ Ÿ
Figure 3. Retrospective binning of MR data: the 1D displacement of the diaphragm
navigator (red) is used to divide continuously acquired MR data into several bins, which
are then reconstructed separately to independent 3D images. Motion vector fields are
derived from these 3D images and are used to correct PET image reconstruction.
Lung
Liver
Time
Figure 4. Irregular breathing: Retrospective binning using the 1D diaphragm navigator
assumes repeatability of respiratory motion and is vulnerable to irregularity shown here.
1. For the MR pulse sequence, 3D radial projection imaging was chosen for its tolerance to
undersampling. An in-house fast 3D radial sequence was developed for this work.
2. Projection angles were arranged through time using 2-parameter golden angle [9], which can
distribute any number of projections in the 3D angular space in a uniform, pseudo-random, and
nonrepeating fashion (Fig. 5). As a result: a) one can run an open-ended scan without planning for
the number of projections and be assured that all projections will be evenly distributed, and b) one
can use a sliding-window reconstruction with arbitrary window width and be assured that the
projections within each window are always evenly distributed too (Fig. 6). Human volunteer
experiments: gradient-echo 3D projection scans were performed on Siemens Trio (3T) and Avanto
(1.5T) scanners over a FOV of 400×400×400mm at ~2ms per projection (128 points/projection). A
sliding window width of 1000 projections and increments of 500 projections were used, resulting in
temporal resolution of 2 sec and update rate of 1 frame/sec.
3. To obtain 3D image for each sliding-window frame, the projections in each sliding window were
reconstructed using L1 minimization in a sparse domain (3D wavelet: Daubechies-4 level 4) in order
to exploit the spatial sparsity of anatomical images. A large-scale L1-minimization algorithm was
used [10].
4. To obtain the 3D motion vector field that describe the 3D deformation across all temporal frames,
a 4D image registration technique [11,12] was used to analyze all frames at once and to enforce
smoothness of motion across all frames. This exploits the temporal sparsity of respiratory motion.
4) 4D Image Registration
3) Compressed-Sending
Image Reconstruction
2) Two-Param Golden Angle,
Sliding Window Frames
1) 3D Radial Projections,
Continuous Acquisition
Methods
kz
kx
ky
Continuous 3D Projection Acquisition
Figure 5. 1000, 4000, 8000 projections in the 3D k-space, arranged by a 2-
parameter golden angle scheme, which is capable of maintaining uniformity
and pseudo-randomness of projection distribution given arbitrary number of
projections, without repeating any projection twice.
First
Proj.
Last
Proj.
Figure 6. Projections in 3D k-space inside sliding windows of several
widths and positions during an open-ended scan, as distributed by the
2-parameter golden angle, which can maintain distribution uniformity
and pseudo-randomness at arbitrary window widths and positions.
Projections 1 ~ 200 789 ~ 1234 2500 ~ 3500
1000 Projections 4000 Projections 8000 Projections
Results
The proposed 3D acquisition and reconstruction technique can
generate real-time images that clearly delineate the diaphragm
and the cardiac boundary, as seen in Fig. 7. As seen in Fig. 8,
these 3D images provide sufficient details to generate motion
vector field for motion correction in PET image reconstruction.
Due to the use of golden angle, one has the option of running
open-ended scans and free choice of temporal resolution
(trades off smoothly against image quality) of reconstruction.
Figure 7. Several frames of the coronal view of a real-time 3D scan showing respiratory motion, acquired at 128×128×128 pixels over 400×400×400 mm FOV,
reconstructed to 2-sec temporal resolution and update rate of 1 frame/sec. Slice 60 of 128-slice volume is shown in (a~d), and volume projections are shown in (e~h).
a b c d e f g h
Figure 8. 4D registration results of the real-time 3D scan shown in Fig. 7. Several temporal frames shown here (a~d) are aligned to a mean-deformation frame (e),
which minimizes total deformation across all frames. The motion vector field generated here will be used in PET reconstruction to correct for motion effects.
a b c d e
Conclusion
The proposed technique enables the acquisition of 3D
respiratory motion in real time, without the assumption of motion
repeatability, offering an alternative to the diaphragm navigator
and retrospective binning, which takes several minutes of
scanning. Given its flexibility in arbitrary scan duration, the real-
time motion scan can be run continuously during a PET-MR
exam to monitor motion at all time, or can be run at select times
during the exam and leave room for other diagnostic MR scans.
8. Petibon, et al. Cardiac motion compensation and resolution modeling in simul. PET-MR: a cardiac lesion detection study. Phys Med Biol 58 (2013) 2085–2102.
9. Chan, et al. Temporal Stability of Adaptive 3D Radial MRI Using Multidimensional Golden Means. Magnetic Resonance in Medicine 61:354–363 (2009).
10. J. Yang and Y. Zhang. Alternating direction algorithms for L1-problems in compressive sensing, SIAM Journal on Scientific Computing, 33, 1-2, 250-278, 2011.
(Available at: YALL1: Your ALgorithms for L1 Optimization. http://yall1.blogs.rice.edu/)
11. S. Klein, M. Staring, K. Murphy, M.A. Viergever, J.P.W. Pluim, "elastix: a toolbox for intensity based medical image registration," IEEE Transactions on Medical
Imaging, vol. 29, no. 1, pp. 196-205, January 2010. http://elastix.isi.uu.nl/.
12. C.T. Metz, S. Klein, M. Schaap, T. van Walsum and W.J. Niessen. Nonrigid registration of dynamic medical imaging data using nD+t B-splines and a groupwise
optimization approach, Medical Image Analysis, 15 (2011) 238–249.