1952
High spatiotemporal resolution cones 4D flow using memory-efficient iterative reconstruction1Electrical Engineering, Stanford University, Stanford, CA, United States, 2Electrical Engineering, University of California, Berkeley, Berkeley, CA, United States, 3Radiology, Stanford University, Stanford, CA, United States
Synopsis
4D flow MRI enables comprehensive cardiovascular assessment, but is limited by long acquisition times and motion corruption. Non-Cartesian sampling strategies, such as radial and cones, exhibit excellent aliasing properties that allow reduced scan times and improve motion robustness. However, trade-offs between spatial and temporal resolution are necessary due to computational burden of iterative 4D non-Cartesian reconstruction. This has restricted cones 4D flow to low temporal resolution venous applications. Here we present a memory-efficient iterative reconstruction utilizing batch processing to enable arbitrary spatiotemporal resolution. We demonstrate feasibility of sub-millimeter, 30 cardiac phase cones 4D flow for coronary artery and valvular assessment.
Purpose
Contrast-enhanced 4D flow MRI1 enables comprehensive assessment of cardiovascular anatomy and function, but long acquisition times and motion corruption continue to limit clinical use especially for pediatrics. Efficient non-Cartesian sampling strategies have been successfully integrated into 4D flow acquisitions to simultaneously reduce scan time and sensitivity to motion artifacts2,3. Cones 4D flow4 is one such technique where velocity encoded data is acquired along spiraling trajectories that lie on concentric cone surfaces. The 3D cones trajectory5 exhibits excellent aliasing properties that makes it inherently well suited for time-resolved reconstructions with parallel imaging and compressed sensing (PICS). However, iterative 4D non-Cartesian reconstructions require vast amounts of computational resources often resulting in necessary trade-offs between reconstructed spatial and temporal resolutions. Previously, this has limited cones 4D flow to low temporal resolution applications such as abdominal venous flow. Here, we present a memory-efficient PICS implementation utilizing batch processing to enable cones 4D flow reconstructions at arbitrary spatial and temporal resolutions on off-the-shelf GPU hardware. In two subjects, we show that sub-millimeter resolution, 30 cardiac phase cones 4D flow reconstructions are feasible and yield exceptional image quality.Methods
Pulse Sequence: Cones 4D flow is based on an RF-spoiled gradient recalled echo (SPGR) sequence with a simple 4-point interleaved velocity encoding scheme (Fig. 1). The cone trajectory ordering is randomized by golden-ratio permutations of the sequential ordering to enhance motion robustness6. Further, temporal resolution can be flexibly chosen since randomized cones ordering ensures approximately uniform k-space coverage regardless of how readouts are retrospectively sorted. Readout durations are kept short to reduce off-resonance, eddy current effects, and maintain an incoherent point spread function.
Reconstruction: Images (x) are reconstructed from non-Cartesian k-space data (y) by iteratively solving the following L1-constrained least-squares problem using a first-order primal dual algorithm7: $$\underset{x}{\text{minimize}} \, || A x - y ||_2^2 + \lambda || D x ||_1.$$ The forward model (A) is comprised of sensitivity maps estimated using ESPIRiT8, the non-uniform FFT (NUFFT) operator, and density compensation weights. The L1 regularization term enforces sparsity in temporal finite differences (D) domain9. Forward and transpose NUFFT operations are implemented on GPU and sequentially process all coils and phases batch-by-batch. In this work, we used a batch size of 12. To accelerate reconstruction time and reduce memory requirements, a reduced oversampling ratio10 of 1.25 is used. The reconstruction is fully implemented in Python using libraries from signal processing package, SigPy11. Cartesian data is reconstructed using l1-ESPIRiT with spatial wavelet and temporal finite differences constraints. To further suppress respiratory motion artifacts in Cartesian images, respiratory motion estimates are computed from butterfly navigators and used to penalize against inconsistent respiratory motion states during reconstruction12.
Experiments: With IRB approval and informed consent, two pediatric subjects referred for contrast-enhanced chest MRI exams were scanned using Cartesian and cones 4D flow sequences on a 3T scanner (GE MR750, Waukesha, WI) with a 12-channel screen-printed pediatric coil13. Cones 4D flow scan parameters include flip angle: 15°, readout duration: 1.6-1.8 ms, spatial resolution: 0.8x0.8x1.6 mm3, matrix size: 320x320x100, 30 cardiac phases, 1931 cone readouts/phase, venc: 250 cm/s, and scan duration: 15 minutes. The acceleration rate was chosen to match the typical Cartesian 4D flow scan time of 15 minutes with similar scan parameters. All data was acquired with subjects freely breathing.
Results
High temporal resolution cones 4D flow images provide excellent delineation of fine heart anatomy, including coronary arteries, which are not well represented in a time-averaged gridding reconstruction (Figure 2). Both Cartesian and cones compressed sensing reconstructions are able to depict small but fast hemodynamics caused by valvular dysfunction (Fig. 3) Cones 4D flow reduces the appearance of respiratory motion artifacts near the diaphragm, but also depicts stair-casing artifacts caused by temporal regularization (Fig. 3). Full time-resolved gridding reconstructions of all four flow encodes are completed in 2-3 minutes, whereas compressed reconstructions complete in 11-12 hours on a single NVIDIA Titan Xp graphics card.Discussion & Conclusions
High spatiotemporal resolution cones 4D flow reconstructions are feasible despite the large computational burden of high-dimensional non-Cartesian iterative reconstruction algorithms. Cones 4D flow currently provides similar image quality to Cartesian 4D flow, although with slightly worse regularization artifacts. This can likely be mitigated by incorporating other more accurate regularizers14 or by resolving respiratory motion states in addition to cardiac motion states15. The current reconstruction time of 11-12 hours is clinically infeasible; however, we believe the method presented here is highly scalable. Batch processing during NUFFT steps could be easily parallelized across multiple GPUs, and reduce overhead of transferring data between GPU and CPU.Acknowledgements
National Science Foundation Graduate Research Fellowship (DGE-114747), NIH R01EB009690, NIH R01EB026136, GE HealthcareReferences
1. J. Y. Cheng, K. Hanneman, T. Zhang, M. T. Alley, P. Lai, J. I. Tamir, M. Uecker, J. M. Pauly, M. Lustig, and S. S. Vasanawala, “Comprehensive motion-compensated highly accelerated 4D flow MRI with ferumoxytol enhancement for pediatric congenital heart disease,” J. Magn. Reson. Imaging, vol. 43, no. 6, pp. 1355–68, 2016.
2. T. Gu, F. R. Korosec, W. F. Block, S. B. Fain, Q. Turk, D. Lum, Y. Zhou, T. M. Grist, V. Haughton, and C. A. Mistretta, “PC VIPR: A high-speed 3D phase-contrast method for flow quantification and high-resolution angiography,” Am. J. Neuroradiol., vol. 26, no. 4, pp. 743–749, 2005.
3. A. Sigfridsson, S. Petersson, C. J. Carlhäll, and T. Ebbers, “Four-dimensional flow MRI using spiral acquisition,” Magn. Reson. Med., 2012.
4. C. M. Sandino, J. Y. Cheng, M. T. Alley, M. Carl, and S. S. Vasanawala, “Accelerated abdominal 4D flow MRI using 3D golden-angle cones trajectory,” in ISMRM, Paris, France., 2018.
5. P. T. Gurney, B. A. Hargreaves, and D. G. Nishimura, “Design and analysis of a practical 3D cones trajectory,” Magn. Reson. Med., vol. 55, no. 3, pp. 575–582, 2006.
6. E. J. Zucker, J. Y. Cheng, A. Haldipur, M. Carl, and S. S. Vasanawala, “Free-breathing pediatric chest MRI: Performance of self-navigated golden-angle ordered conical ultrashort echo time acquisition,” J. Magn. Reson. Imaging, vol. 47, no. 1, pp. 200–209, 2018.
7. A. Chambolle and T. Pock, “A first-order primal-dual algorithm for convex problems with applications to imaging,” J. Math. Imaging Vis., 2011.
8. M. Uecker, P. Lai, M. J. Murphy, P. Virtue, M. Elad, J. M. Pauly, S. S. Vasanawala, and M. Lustig, “ESPIRiT - An eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA,” Magn. Reson. Med., vol. 71, no. 3, pp. 990–1001, 2014.
9. L. Feng, R. Grimm, K. T. Block, H. Chandarana, S. Kim, J. Xu, L. Axel, D. K. Sodickson, and R. Otazo, “Golden-angle radial sparse parallel MRI: combination of compressed sensing, parallel imaging, and golden-angle radial sampling for fast and flexible dynamic volumetric MRI,” Magn. Reson. Med., vol. 72, no. 3, pp. 707–717, 2014.
10. P. J. Beatty, D. G. Nishimura, and J. M. Pauly, “Rapid gridding reconstruction with a minimal oversampling ratio,” IEEE Trans. Med. Imaging, vol. 24, no. 6, pp. 799–808, 2005.
11. F. Ong, “SigPy,” 2018. [Online]. Available: https://github.com/mikgroup/sigpy.
12. K. M. Johnson, W. F. Block, S. B. Reeder, and A. Samsonov, “Improved least squares MR image reconstruction using estimates of k-Space data consistency,” Magn. Reson. Med., vol. 67, no. 7, pp. 1600–1608, 2012.
13. J. R. Corea, A. M. Flynn, B. Lechêne, G. Scott, G. D. Reed, P. J. Shin, M. Lustig, and A. C. Arias, “Screen-printed flexible MRI receive coils,” Nat. Commun., 2016.
14. C. Santelli, M. Loecher, J. Busch, O. Wieben, T. Schaeffter, and S. Kozerke, “Accelerating 4D flow MRI by exploiting vector field divergence regularization,” Magn. Reson. Med., vol. 75, no. 1, pp. 115–125, 2016.
15. L. Feng, L. Axel, H. Chandarana, K. T. Block, D. K. Sodickson, and R. Otazo, “XD-GRASP: Golden-angle radial MRI with reconstruction of extra motion-state dimensions using compressed sensing,” Magn. Reson. Med., vol. 75, no. 2, pp. 775–788, 2016.
Figures
