0134

Self-navigated, rapid 3D UTE for motion-robust skull imaging

Hyunyeol Lee¹, Xia Zhao¹, Hee Kwon Song¹, Rosaline Zhang², Scott P Barttlett², and Felix W Wehrli¹

¹Radiology, University of Pennsylvania, Philadelphia, PA, United States, ²Plastic Surgery, University of Pennsylvania, Philadelphia, PA, United States

Synopsis

Solid-state MRI via 3D UTE or ZTE methods has potential for bone-selective imaging as a radiation-free alternative to computed tomography, particularly for children with craniofacial abnormalities. However, relatively long scan times make the technique vulnerable to artifacts from involuntary subject movements, thereby impairing image quality. Here, we developed a self-navigated, rapid 3D UTE technique by combining a retrospective motion detection and correction approach with sparsity-constrained image reconstruction. Results from in vivo studies demonstrate the proposed method's feasibility in achieving motion-corrected whole-skull images at 1 mm isotropic resolution in 2.1 minutes scan time.

Introduction

Solid-state MRI via 3D ultrashort echo-time (UTE)¹ or zero TE² methods, capable of detecting signals from protons with very short T₂ relaxation times, has potential for bone-selective imaging^3-5, for instance as a radiation-free alternative to computed tomography for the pre- and post-surgical evaluation of children with craniofacial abnormalities. However, relatively long scan times make the technique vulnerable to artifacts from involuntary subject movements, thereby impairing image quality. Here, we developed a self-navigated, rapid 3D UTE technique by combining a retrospective motion detection/correction approach⁶ with sparsity-constrained image reconstruction. In vivo studies were performed to investigate the feasibility of the proposed method in achieving rapid, motion-resistant whole-skull imaging.

Methods

Motion detection and correction: Figure 1 shows a diagram of the proposed pulse sequence. While retaining the dual-RF/dual-echo configuration⁴ and the view-sharing scheme⁷ for achieving high bone specificity with enhanced imaging efficiency, the method employs a multi-dimensional golden-means (GM) based k-space trajectory⁸ for retrospective motion detection and correction⁶. Specifically, GRE signals acquired as full projections (Fig. 1) are employed to derive the center of mass (COM) using the relationship⁹: $$$\gamma_{COM}=\frac{\int r\Re (r,\theta )dr}{\int \Re (r,\theta )dr} $$$ where $$$\gamma_{COM}$$$ is the projection of COM onto a radial line with the angle θ, and $$$\Re$$$ is the Radon transform of the object. The time-course of COM during data collection is then analyzed for adaptive determination of motion states, within each of which sampling views are distributed near-evenly in 3D k-space thereby allowing reconstruction of low-resolution images representative of a particular motion state. Subsequently, rigid-motion parameters are extracted for individual motion states via FSL¹⁰, leading to correction of acquired k-space datasets. The final, high-resolution motion-corrected images are obtained using the reconstruction method described below. The above procedures are summarized in Fig. 2.

Bone-selective image reconstruction: Given the sparse bone signals in the difference between short and long TE images, bone-specific imaging is further accelerated with fewer radial lines by exploiting such sparsity during image reconstruction^11,12. The following sparse signal recovery problem can then be formulated: $$\min_{I_1,I_2} \frac{1}{2} \sum_{j=1}^{N_c} (\| k_{1,j}-F_{NU}(S_j I_1)\|^2_2 + \| k_{2,j}-F_{NU}(S_j I_2)\|^2_2)+ \lambda \| I_1-I_2e^{-i \varphi }\|_1 [1]$$ where k₁/k₂ are the motion-corrected and view-shared k-space data at TE₁/TE₂, and I₁/I₂ are the corresponding complex images is the non-uniform fast FT (NUFFT), S_j is the receive sensitivity for the j-th coil, N_c and λ are the number of receive coils and regularization parameter, respectively, and φ is the phase accrual during ΔTE. The phase correction with φ in the subtraction is essential, as otherwise residual sparsity may be disrupted. Both S and φ are spatially smooth and thus can be estimated using over-sampled, central k-space data. The solutions (I₁, I₂) are found with an alternating minimization approach that splits Eq. [1] into two sub-problems with respect to I₁ and I₂. The two solutions are iteratively updated until convergence is reached.

In vivo studies: Two subjects were scanned at 3T (Siemens Prisma) using the following parameters: TR/TE₁/TE₂=5.0/0.06/1.84ms, RF₁/RF₂ durations=40/520μs, flip-angle=12° (identical for RF₁ and RF₂), matrix size = 256³, field-of-view = 256³mm³, and readout bandwidth=±125kHz. A 20-channel head/neck coil was used for signal reception. Both subjects were instructed to move the head three to four times during each scan. To test the sequence’s self-navigation effectiveness, data were acquired in the first subject using a relatively large number of views (50,000 for each echo; scan time=8.4min). Following the motion detection/correction steps, images for UTE from RF₁ and GRE from RF₂ were reconstructed using inverse NUFFT. Bone-specific images (I_Bone) were then obtained as $$$I_{Bone}=\frac{I_1-I_2}{I_1+I_2} $$$. In the second subject, data were prospectively undersampled using 12,500 views (scan time=2.1min). Motion-corrected and view-shared k-space datasets were then processed to reconstruct images using Eq.[1].

Results

Figure 3 displays results from each processing step in Fig. 2. The time-course of COM accurately reflects four occurrences of the subject’s head motion (Fig. 3a), leading to five sets of GRE image corresponding to each motion state (Fig. 3b). Correction of k-space datasets using the estimated rigid-motion parameters (Fig. 3c) yields clear depiction of inner and outer table of the cranium in I_Bone after removal of motion-induced image blurring in both UTE and GRE images (Fig. 3d). Figure 4 compares two sets of images from the second subject; one reconstructed directly using inverse NUFFT (Fig. 4a) and one with motion correction followed by sparse reconstruction (Fig. 4b). Image blurring and artifacts due to subject motion and data subsampling are effectively eliminated using the proposed method.

Conclusions

Results suggest the proposed method to be robust to head movement during scanning. Upon further optimization, the method should find applications for bone-selective head imaging as a radiation-free alternative to computed tomography in children indicated for craniofacial surgery.

Acknowledgements

NIH grant R01AR050068

References

1. Robson MD, Gatehouse PD, Bydder M, Bydder GM. Magnetic resonance: An introduction to ultrashort TE (UTE) imaging. J Comput Assist Tomogr 2003;27:825-846.

2. Weiger M, Pruessmann KP, Hennel F. MRI with zero echo time: hard versus sweep pulse excitation. Magn Reson Med 2011;66(2):379-389.

3. Li C, Magland JF, Zhao X, Seifert AC, Wehrli FW. Selective in vivo bone imaging with long-T suppressed PETRA MRI. Magn Reson Med 2016;77(3):989-997.

4. Johnson EM, Vyas U, Ghanouni P, Pauly KB, Pauly JM. Improved cortical bone specificity in UTE MR imaging. Magn Reson Med 2017;77:684-695.

5. Wiesinger F, Sacolick LI, Menini A, Kaushik SS, Ahn S, Veit-Haibach P, Delso G, Shanbhag DD. Zero TE MR bone imaging in the head. Magn Reson Med 2016;75(1):107-114.

6. Anderson III AG, Velikina J, Block W, Wieben O, Samsonov A. Adaptive retrospective correction of motion artifacts in cranial MRI with multicoil three‐dimensional radial acquisitions. Magn Reson Med 2013;69(4):1094-1103.

7. Lee H, Zhao X, Song HK, Zhang R, Bartlett SP, Wehrli FW. Solid-state MRI as a noninvasive alternative to computed tomography for craniofacial imaging. Joint Annual Meeting ISMRM- ESMRMB 2018;332.

8. Chan RW, Ramsay EA, Cunningham CH, Plewes DB. Temporal stability of adaptive 3D radial MRI using multidimensional golden means. Magn Reson Med 2009;61(2):354-363.

9. Larson AC, White RD, Laub G, McVeigh ER, Li D, Simonetti OP. Self‐gated cardiac cine MRI. Magn Reson Med 2004;51(1):93-102.

10. M. Jenkinson and S.M. Smith. A global optimisation method for robust affine registration of brain images. Med Image Anal 2001;5(2):143-156.

11. Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magn Reson Med 2007;58(6):1182-1195.

12. Nam S, Akcakaya M, Basha T, Stehning C, Manning WJ, Tarokh V, Nezafat R. Compressed sensing reconstruction for whole-heart imaging with 3D radial trajectories: a graphics processing unit implementation. Magn Reson Med 2013;69(1):91-102.

Figures

Figure 1. a: Diagram of the dual-RF and dual-echo 3D UTE pulse sequence, in which RF₁ (short ~ 40μs) and RF₂ (long ~ 520μs) are alternately played out while two signals (UTE and GRE) are produced with the gradient polarity reversed. b: comparison of view-orders in distributing projections (number: 4096) in 3D k-space between conventional (left) and the proposed (right) methods. Note that with the 2D golden means based view ordering strategy, any subset of consecutive views is distributed near-evenly in 3D k-space.

Figure 2. Schematic of the proposed, self-navigated motion detection and retrospective motion correction method. Note that the image reconstruction on individual motion states is performed oversampled, central k-space data, while image reconstruction at the last step can be implemented either by directly applying inverse NUFFT or by solving the sparsity-constrained problem (Eq. [1]).

Figure 3. a: Derived COM in the three axes along the time course (resolution=0.5 sec; total number of views = 50,000; scan time = 8.4 min). Arrows represent time points the subject was prompted to move the head. b: GRE images for individual motion states. c: Rigid motion parameters in the five displacement states (relative to state 1), extracted using FSL. d: Images obtained without (top) and with (bottom) motion correction. Note that motion-induced image blurring is effectively removed after correction, allowing clear visualization of inner and outer tables of skull (red arrows).

Figure 4. Comparison of images, acquired using the proposed pulse sequence with 12,500 views (scan time = 2.1 min); top: direct application of inverse NUFFT on uncorrected k-space datasets; bottom: motion-correction followed by sparsity-constrained image reconstruction.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)

0134