BMART-Enabled Field-Map Combination of Projection-Reconstruction Phase-Cycled SSFP Cardiac Cine for Banding and Flow-Artifact Reduction
BBMART-Enabled Field-Map Combination ofPro jection-Reconstruction Phase-Cycled SSFP CardiacCine for Banding and Flow-Artifact Reduction
Anjali Datta, Dwight G Nishimura, and Corey A Baron
Corresponding Author:
Anjali DattaDavid Packard Electrical Engineering350 Jane Stanford Way, Rm. 308Stanford, CA, USA [email protected](817) 269-9548February 24, 2021
Submitted to
Magnetic Resonance in Medicine .Approximate word count: 3200 1 a r X i v : . [ phy s i c s . m e d - ph ] F e b bstractPurpose: To develop a method for banding-free bSSFP cardiac cine with substantiallyreduced flow artifacts.
Methods:
A projection-reconstruction (PR) trajectory is proposed for a frequency-modulated cine sequence, facilitating reconstruction of three phase cycles and a field-map time series from a short, breath-held scan. Data is also acquired during thegradient rewinders to enable generation of field maps using BMART, B mappingusing rewinding trajectories, where the rewind data forms the second TE image forcalculating the field map. A field-map-based combination method is developed whichweights the phase-cycle component images to include only passband signal in the finalcine images, and exclude stopband and near-band flow artifacts. Results:
The weights derived from the BMART-generated field maps mask out band-ing and near-band flow artifacts in and around the heart. Therefore, the field-map-based phase-cycle combination, which is facilitated by the PR acquisition with BMART,results in more homogeneous blood pools and reduced hyperintense regions than root-sum-of-squares.
Conclusion:
With the proposed techniques, using a non-Cartesian trajectory for afrequency-modulated cine sequence enables flow-artifact-reduced banding-free cardiacimaging within a short breath-hold.
Keywords: balanced SSFP (bSSFP); banding; flow artifacts; phase-cycle combina-tion; cardiac cine; B field map; projection-reconstruction; radial; frequency modula-tion ntroduction Balanced SSFP’s high SNR, strong blood-myocardium contrast, and short acquisition timemake it ideal for cardiac imaging [1]. However, the sequence is sensitive to off-resonance –the spectral profile of balanced SSFP (bSSFP) has periodic nulls spaced at the repetitionrate (i.e., TR − ), and images suffer from banding artifacts in regions where the local B fieldstrength corresponds to a null.In phase-cycled, or multiple-acquisition, bSSFP, images are acquired with different RFphase increments (i.e., “phase cycles”). The phase cycling shifts the spectral profile andtherefore the bands in the individual component images, which are then combined into anull-free image [2, 3]. Phase-cycled bSSFP is thus robust to off-resonance, but the scan timeis increased by a factor of the number of phase cycles acquired, which is usually from two tofour. Various methods have been proposed to counteract this scan time increase, enablingphase-cycled bSSFP acquisitions in clinically relevant scan times [4–8].Another challenge to acquiring phase-cycled bSSFP in the heart is that the signal atthe bands is highly sensitive to the through-plane flow rate, which results in near-band flowartifacts [9]. Banding falls within the blood pool in at least one of the phase-cycled com-ponent images. Due to contributions from spins that have flowed out of the slice, flow innear-band regions results in hyperintense artifacts [9], which are mitigated but still presenteven if partial dephasing is used [10]. Therefore, standard phase-cycle combination methodswhich effectively reduce banding, such as maximum-intensity projection, magnitude-mean,and root-sum-of-squares (SOS), fail to eliminate these artifacts. As a result, many of theproposed accelerated phase-cycling methods have not been shown in the heart [5–7], andthose that have, suffer from residual flow artifacts [4, 8, 11]. The poor performance of stan-dard phase-cycle combination methods for cardiac applications stems from their use of bothpassband signal and stopband signal, and thus artifactual near-band flow signal, in the cal-culation of the final combined image. However, if the precession frequency at each location isknown, it could be used to include only passband signal in the combined image, and excludestopband and near-band flow-related signal.For a 2D projection-reconstruction (PR) trajectory, the rewinders necessary to null the3a) k x k y k x k y
12 3 (b) Normal Echo: k x k y Rewind Echo: k x k y Figure 1: Illustration of the k-space coverage of the normal and rewind acquisitions. (a) Forthe spoke highlighted in the illustration of the projection-reconstruction trajectory on theleft, the prewinder (red), readout (purple), and rewinder (blue) are shown on the plot onthe right. (b) If neighboring spokes are acquired in opposite directions (i.e., the spokes spanall 360 ◦ , not just 180 ◦ ), the rewinds also sample from all of k-space. Therefore, by acquiringduring the rewinders, a second image with a later TE can be reconstructed from the bSSFPacquisition. Note that, for simplicity, only a small number of spokes are shown here.gradient axes in balanced SSFP cause the k-space trajectory to trace back along the secondhalf of the readout path (Figure 1a). Therefore, by also acquiring data during the rewinders,two images with distinct TEs can be acquired with minimal modifications to the sequencefor non-Cartesian bSSFP acquisitions where the rewinds fully sample k-space (Figure 1b).This is exploited by B mapping using rewinding trajectories (BMART) [12] to estimate amain-field map, which could potentially be used to determine the locations of the passbandsand stopbands during phase-cycle combination, without additional scans.In this work, we demonstrate the feasibility of using a PR trajectory for an acceleratedfrequency-modulated cine sequence in conjunction with BMART to acquire three phase4ycles and a field-map time series within a short breath-hold, enabling use of a field-map-based combination method when reconstructing the final cine images. We hypothesize that,by including only passband signal, and excluding stopband and near-band flow-related signal,such a combination will result in reduced signal contributions from out-of-slice spins and amore homogeneous blood pool than root-sum-of-squares. Methods
Acquisition
A PR cine sequence that acquires three phase cycles for each cardiac phase is proposed tobalance banding reduction and scan time. The scan time is thirteen heartbeats, including aninitial heartbeat of discarded acquisitions for signal stabilization. The frequency modulationscheme presented in [8] is used, obviating the need for any stabilization heartbeats betweenphase cycles. During the first RR interval after the discarded acquisitions, some data forphase cycle 1 is acquired. During RRs 2 and 3, data for phase cycles 2 and 3, respectively,are acquired, before circling back to phase cycle 1 to acquire additional data in the followingheartbeat (see Figure 1a of [8] for an illustration). As a result, the acquisition of the threephase cycles is interleaved.In the PR trajectory, the lines are rotated through all 360 ◦ so that the rewinds cover allof k-space (Figure 1). An undersampled segmented acquisition with twelve view-angles persegment is used, but the sampling pattern is rotated every cardiac phase (Figure 2). Fora given cardiac phase, the sampling pattern is also rotated every RR interval, so differentview angles are acquired for each effective phase cycle. Over the course of the scan, 48k-space diameters are sampled for each effective phase-cycle of each cardiac phase, whichcorresponds to an undersampling ratio of 7.1. Since different sets of spokes are acquiredduring different cardiac phases and effective phase cycles (Table 1), combining the cardiacphases and phase cycles generates fully sampled data for parallel-imaging calibration (similarto earlier Cartesian work [8]) and field-map generation.The sequence was implemented at 1.5 T for use with an eight-channel cardiac coil. It5 x k y Phase Cycle 1 Phase Cycle 2 Phase Cycle 3 Combination C a r d i a c P h a s e C a r d i a c P h a s e C a r d i a c P h a s e C a r d i a c P h a s e Field MapRR 1 2 3 4 5 6 7 8 9 10 11 12Figure 2: View-angle ordering for the proposed frequency-modulated projection-reconstruction cine sequence. Although the sequence acquires diameters of k-space, forclarity, only half of each readout is shown. The sampling pattern is rotated every cardiacphase and between effective phase cycles. Since different view angles are sampled for thethree phase cycles, the phase-cycle combination is less undersampled than the individualphase cycles. Since the view angles sampled during any four consecutive cardiac phases arealso different, the sliding windows of complex-summed combinations used to estimate fieldmaps are even more densely sampled. Note that, since diameters are acquired but only radiiare shown here, both 30 ◦ insets also contain half as many lines as acquired during the normalreadout – the spokes in the opposite quadrants also traverse the highlighted region.6able 1: View angle ordering. With 576 views, the view angle for view v is 360 · v/ v ≤
288 and 360 · ( v − . /
576 degrees otherwise.7cquires a cine loop with 41 ms temporal resolution of an 8-mm-thick axial slice with 1.6mm resolution and 36 cm field of view. The other scan parameters are: 60 ◦ tip angle, 1.3ms TE, and 3.4 ms TR. This TR is 0.18 ms longer than that of a sequence that does notacquire data on the rewinds. Note that, unlike for the 3D acquisitions considered in [12], fora 2D balanced SSFP sequence, the TR is lengthened slightly since the readout rewinders canno longer overlap with the slice-select prewinder (Figure 3). The readout trapezoid uses areceiver bandwidth of ±
125 kHz, and samples are also acquired on the ramps. The rewinderreaches the maximum gradient amplitude, so a bandwidth of ±
250 kHz is used.Three slices of a healthy subject were imaged with informed consent and InstitutionalReview Board (IRB) approval. As in [8], 30 ◦ of partial dephasing was applied in the slice-select direction to mitigate near-band flow artifacts, and the shim was offset prior to imagingto create challenging off-resonance conditions. Gradient delays were calibrated before recon-struction, but no other corrections, e.g., for eddy currents or gradient-amplifier nonlinearity,were applied. Reconstruction
The cine data is reconstructed using ESPIRiT with (cid:96) maps using BMART, both the normal and the rewind imagesare reconstructed with ESPIRiT with Tikhonov regularization using the estimated receiversensitivity maps and an empirically determined regularization parameter. The three effectivephase cycles are complex-summed, and a sliding window four cardiac phases long is used tobalance the temporal resolution of the field-map time series with the SNR of the field maps, soa total of 576 sampled lines are used for the estimation of each map. Since neighboring spokesare acquired in opposite directions, we assume that, for the normal data, the acquisition8YZRFFigure 3: Pulse sequence diagram of projection reconstruction bSSFP acquisition that en-ables BMART. Since data is being acquired on the rewinds, the readout rewinders can nolonger overlap with the slice-select prewinder, so the TR is lengthened slightly.9ime is TE for all of k-space (Figure 1). The average delay between the normal and rewindacquisition at a gridded k-space location thus depends on the acquisition times of rewindsamples that contribute to that grid location. The data is sorted into 25 bins based on thismean effective delay and the B map is generated in MATLAB using the algorithm proposedin [12] (see Figure 2a of that paper for an illustration). The full width at half maximumof the Hanning window low-pass filter used to further increase the SNR of the field maps isone-third of the k-space extent. Field-Map Combination
The precession per TR, φ , of the spins in each voxel is calculated from the B map cor-responding to each cardiac phase and used to weight the phase-cycled component imagesto emphasize passband signal and deemphasize stopband signal. If the most severe flowartifacts are localized near stopbands [9], such a field-map combination method should sub-stantially reduce residual flow artifacts and contributions from out-of-slice spins in the finalreconstructed image.The combined image m c is calculated as m c ( x, y, t ) = (cid:80) Nn =1 (cid:107) m n ( x, y, t ) (cid:107) w (( φ ( x, y, t ) − ψ n ) mod 360 ◦ ) (cid:80) Nn =1 w (( φ ( x, y, t ) − ψ n ) mod 360 ◦ )where x and y are spatial coordinates, t is the cardiac phase, N is the number of phase cycles, m n is the nth component image, ψ n is the average phase increment for the nth effective phasecycle during that cardiac phase, and w is a weighting function. We heuristically used atriangular weighting function w ( φ ) = tri (( φ − ◦ ) / ◦ ) to smoothly merge the passbandsof the three phase cycles while avoiding as much of the stopbands and flow artifacts aspossible (Figure 4). Note that, over the cardiac cycle, in addition to the temporal changes inthe field map, the phase cycling shifts due to the frequency modulation scheme. Therefore,the weighting of the three images at any given location changes with the cardiac phase. Inregions where the mean amplitude over the cardiac cycle of the SOS phase-cycle combinationis less than 5% of the maximum amplitude in the image, magnitude mean is used for phase-cycle combination instead of relying on the noisy field-map estimate to combine the effective10 a) Phase Cycle Spectral Profiles and Weighting Functions (b)
Combined Spectral Profiles
Figure 4: Field-map combination of three phase cycles. (a) The triangular weighting func-tions smoothly merge the passbands of the phase cycles while excluding as much of thestopbands (and associated flow artifacts) as possible. (b) Like other combination methods,field-map combination results in a flatter spectral profile with no nulls. The simulation pa-rameters are T = 1000 ms, T = 150 ms, TR = 3 . .
75 ms, and tip angle α = 30 ◦ . The individual spectral profiles are normalized such that signal at the center ofany passband is 1. Note that the relative flatness of the combined profiles depends on the T , T , and α . 11hase cycles. This is also done where the SOS amplitude is more than 50% of the maximum,since bright voxels are presumed to consist primarily of fat, where chemical shift may result inphase wrapping in the B map. The proposed method is compared with root-sum-of-squares.Bloch simulations are used to evaluate the flatness of the spectral profiles of flowing spinsafter field-map combination. In particular, we look to see how effective including only thepassband signal is at suppressing hyperintense signal from flow at precession frequenciesnear the stopbands. The spectral profiles for various constant through-plane flow rates ofSSFP with 30 ◦ of partial dephasing are determined using the simulations described in [10].The flow rate is quantified as the percent of blood in the excited slice replaced by freshmagnetization each TR. Spin replacement percentages between 6 .
25% and 29 .
2% are used.The slice thickness is 8 mm. Other simulation parameters are unchanged: TR = 3 . .
75 ms, T = 1000 ms, T = 150 ms (in-slice), T ∗ = 75 ms (out-of-slice), and tip angle α = 60 ◦ . The profiles are normalized so that the signal of stationary, on-resonant spins inbSSFP is 1. These profiles are then shifted to correspond to 180 ◦ , 300 ◦ , and 60 ◦ phase-cycling,and the phase-cycled spectral profiles are combined using field-map combination and SOS.We consider approximately two-fold enhancement of passband signal due to the inflow of freshmagnetization to be desirable [14], whereas we consider hyper-enhancement of near-bandsignal to be artifactual, so the ideal plot after phase-cycle combination would be a constantplane at approximately two (i.e., the desired signal for flowing blood is approximately twicethat of static blood). Results
Across the range of simulated flow velocities (from 14.3 cm/s to 66.7 cm/s), field-map combi-nation effectively excludes the hyperintense flow signal near the stopbands of the simulatedphase-cycled spectral profiles. As a result, the field-map-combined profiles are substantiallyflatter than those combined using SOS (Figure 5).In vivo, the B maps generated by BMART appear slowly-varying over the image, as isexpected, except in regions with fat or phase wrapping (Figure 6). In addition, the fieldmap changes gradually over the cardiac cycle (Supporting Information Video S1). The12 a) b) Figure 5: Simulated spectral profiles for various constant through-plane flow rates between14.3 cm/s (6 .
25% spin replacement) and 66.7 cm/s (29 .
2% spin replacement). The individualphase-cycle spectral profiles in (a) are combined using root-sum-of-squares (SOS) and theproposed field-map-based method to obtain the plots in (b). By including only the pass-band signal in the final profiles and excluding the hyperintense flow signal near the bSSFPstopbands, field-map combination results in substantially flatter combined profiles than SOS.14
MART-Generated B Field Maps (in Hz) − − map mask out the banding and near-band flow artifacts in andaround the heart in the component images (Figure 7). As a result, in all three imaged slices,the field-map combinations have more homogeneous blood pools and substantially reducedhyperintense regions than the SOS combinations (selected cardiac phases are shown in Figure8). This holds true throughout the cardiac cycle, so the field-map-combined cine loopsshow less artifactual intensity variations between cardiac phases than their SOS counterparts(Supporting Information Videos S2, S3, and S4). Discussion
We provide a proof-of-concept that a PR trajectory can be used for a frequency-modulatedcine sequence, enabling reconstruction of three phase cycles and a field-map time series fromdata acquired within a short breath-hold. Acquiring data on the rewinds to enable generationof field maps using BMART requires only a 5% lengthening of the TR but facilitates field-map combination of the phase-cycle images. The resulting combined images show morehomogeneous blood pool and reduced signal contributions from out-of-slice spins.Although a Cartesian trajectory could have been used, acquiring two echoes during eachTR may result in a longer scan time than PR with BMART. On the other hand, acquiring aB map during a separate scan, as done in [15], introduces the possibility of misregistrationbetween the cine images and field map. In addition, any changes in the field map over thecardiac cycle are not captured. PR’s tolerance of undersampling, motion, and flow may also15igure 7: Weighting of component images. For all three phase-cycled component images,the weights determined from the leftmost field map in Figure 6 mask out the hyperintenseflow artifacts (red arrows) near the bands (cyan arrows) in the blood pool. The field-mapcombination image is the sum of the weighted component images. These component imagescorrespond to the leftmost cardiac phase of the middle slice shown in Figure 8.16 uperior Slice S O S F M C Middle Slice S O S F M C nferior Slice S O S F M C Figure 8: Root-sum-of-squares and field-map combinations of cardiac phases evenly dis-tributed through the cardiac cycle. The field-map combinations (lower row of images foreach slice, labelled FMC) have substantially reduced blood pool heterogeneity compared toroot-sum-of-squares combined images (upper row for each slice, labelled SOS).be desirable in this application.The correspondence of the weights with the banding in and near the heart suggests thatthe field maps generated using BMART are accurate in the region of interest. However, thefield maps suffered from phase-wrap errors in the subcutaneous fat, where the additionaloff-resonance from the detuned shim is also largest. While the TE difference is 0.92 ms,note that the dynamic range of the field maps is not well-defined since the delay betweenthe normal and rewind acquisition times decreases as | k | increases; however, the range isgreater than or equal to ±
543 Hz. Since field-map combination only uses φ and not B itself,selecting the time difference between the normal and rewind acquisitions at the center ofk-space such that the TR is an integer multiple of it may warrant investigation; this wouldcause any phase-wrapped B values to still correspond to the correct precession per TR.It may alternatively be possible to reconstruct subspace images and linearly combine themto enable water-fat separation [7], using the BMART-generated B -field maps to correctwater-fat swaps, but this is beyond the scope of this work. In addition, measurement of thegradient waveforms and more accurate characterization of the gradient delays may improve18he accuracy of the field maps.Note that, instead of acquiring data during the gradient rewinders, we could alternativelyhave acquired data during the gradient prewinders. Acquiring data on both the prewinds andrewinds and using all three echoes to estimate the field maps may provide better robustness;however, this would double the TR increase.When determining the field maps, instead of using the entire spoke for the normal data,using either half (either the data up to and including TE, or the data at and after TE) stillresults in samples from all of k-space while avoiding averaging together neighboring datapoints that are acquired at substantially different times. The earlier half is an outside-intrajectory – i.e., the same direction as the rewind data – that results in less variation in thedelay between the normal and rewind acquisitions as a function of | k | . However, this causesthe bin with the longest delay to no longer be at the center of k-space, which may necessitatemodifications to the BMART implementation. The latter half matches the trajectory of therewind, with a center-out normal acquisition followed by an outside-in rewind acquisition.However, we found that using the entire spoke resulted in field maps that appear less noisy.The proposed BMART-enabled phase-cycle combination should be compatible with anyacquisition that results in separate phase-cycle images and that can be modified to alsogenerate a B map – i.e, the frequency modulation scheme used here and partial dephasingare not necessary, and field-map combination could be trivially extended to other numbersof phase cycles (e.g., see [15]). However, partial dephasing may cause artifactual flow signalto be more localized to near the bands, which may enable field-map combination to moreeffectively exclude it from the final combined images. Other spoke ordering schemes withsimilar properties to the one used here, such as golden-angle view ordering [16], could alsobe used. Since the frequency modulation scheme results in interleaved acquisition of thephase cycles and golden-angle view ordering covers k-space relatively evenly for any arbitrarynumber of views, it may be possible to reconstruct data from a smaller number of heartbeats,for example, if motion during the breath-hold is a concern. Eddy current artifacts are apotential downside.Although, in our experience, the flow artifact mitigation from BMART-enabled field-mapcombination was robust, with a marked improvement over SOS in most cardiac phases of all19f the slices we acquired, some hyperintensities remained in some cardiac phases (in the rightventricle of the leftmost cardiac phase of the inferior slice in Figure 8, for example). Thesemay have stemmed from inaccuracies in the field maps or from some flow artifacts beingless localized to regions near the stopbands – e.g., if the flow is oblique, out-of-slice spinsmay contribute signal to voxels away from the one they originated in. A more sophisticatedreconstruction for the B maps that can support higher spatial and temporal resolutionwith adequate SNR or a more optimized weighting function in the field-map combinationmay address some residual artifacts, but these were not investigated here. The basic parallelimaging reconstruction with finite-difference regularization used to reconstruct the individualphase-cycle images resulted in good quality. However, it may be possible to better supportthe high acceleration by also exploiting the redundancy between the effective phase cycles,so this merits investigation.Note that, unlike in the original paper, BMART was not used to correct off-resonance-related blurring in this work. Because of the short readout time in the sequence used here,we did not expect noticeable blurring. Trajectories with longer readouts, such as spirals, mayenable more time-efficient acquisition, e.g., to decrease the acceleration factor or increase thenumber of acquired effective phase cycles. Conclusion
In this work, we proposed using a PR trajectory for a phase-cycled cine acquisition tofacilitate estimation of a field-map time series without additional scans with BMART. Afield-map-based phase-cycle combination method was developed that utilizes the BMART-generated field maps to weight the component images to include only passband signal, andexclude stopband and near-band flow-related signal, from the final cine images. Since theweights derived from the field maps mask out banding and near-band flow artifacts, field-mapcombination results in more homogeneous blood pool signal and reduced hyperintense regionscompared to root-sum-of-squares. Therefore, using the presented methods, a non-Cartesianfrequency-modulated sequence can achieve flow-artifact-reduced banding-free cardiac cineimaging within a short breath-hold. 20 cknowledgements
Thank you to GE Healthcare, NIH R01 HL127039, NSF GRFP, and the Hertz Foundationfor their support.
References [1] Kramer CM, Barkhausen J, Flamm SD, Kim RJ, Nagel E. Standardized cardiovascularmagnetic resonance (CMR) protocols 2013 update. J Cardiovascular Magn Reson 2013;15:91.[2] Zur Y, Wood ML, Neuringer LJ. Motion-insensitive, steady-state free precession imag-ing. Magn Reson Med 1990; 16:444–459.[3] Bangerter NK, Hargreaves BA, Vasanawala SS, Pauly JM, Gold GE, Nishimura DG.Analysis of multiple-acquisition SSFP. Magn Reson Med 2004; 51:1038–1047.[4] Wang Y, Shao X, Martin T, Moeller S, Yacoub E, Wang DJ. Phase-cycled simultaneousmultislice balanced SSFP imaging with CAIPIRINHA for efficient banding reduction.Magn Reson Med 2016; 76:1764–1774.[5] Slawig A, Wech T, Ratz V, Tran-Gia J, Neubauer H, Bley T, K¨ostler H. Multifrequencyreconstruction for frequency-modulated bSSFP. Magn Reson Med 2017; 78:2226–2235.[6] Bilgic B, Witzel T, Bhat H, Wald LL, Setsompop K. Joint reconstruction of phase-cycled balanced SSFP with constrained parallel imaging. In Proceedings of the 25thAnnual Meeting of ISMRM, Honolulu, May 2017. p. 0441.[7] Roeloffs V, Rosenzweig S, Holme HCM, Uecker M, Frahm J. Frequency-modulatedSSFP with radial sampling and subspace reconstruction: A time-efficient alternative tophase-cycled bSSFP. Magn Reson Med 2019; 81:1566–1579.[8] Datta A, Nishimura DG, Baron CA. Banding-free balanced SSFP cardiac cine usingfrequency modulation and phase cycle redundancy. Magn Reson Med 2019; 82:1604–1616. 219] Markl M, Pelc NJ. On flow effects in balanced steady-state free precession imaging:Pictorial description, parameter dependence, and clinical implications. J Magn Reson2004; 20:697–705.[10] Datta A, Cheng JY, Hargreaves BA, Baron CA, Nishimura DG. Mitigation of near-band balanced steady-state free precession through-plane flow artifacts using partialdephasing. Magn Reson Med 2017; 79:2944–2953.[11] Fischer A, Hoff MN, Ghedin P, Brau AC. Banding-artifact free bSSFP cine imagingusing a Geometric Solution approach. In Proceedings of the 24th Annual Meeting ofISMRM, Singapore, May 2016. p. 1827.[12] Baron CA, Nishimura DG. B mapping using rewinding trajectories (BMART). MagnReson Med 2017; 78:664–669.[13] Ong F, Lustig M. Sigpy: A Python package for high performance iterative recon-struction. In Proceedings of the 27th Annual Meeting of ISMRM, Montreal, Quebec,Canada, May 2019. p. 4819.[14] Lai P, Cheng JY, Vasanawala SS, Brau AC. Robust free-breathing whole-heart cine MRIusing multi-slab 3D acquisition with isotropic resolution and offline reformattability. InProceedings of the 23rd Annual Meeting of ISMRM, Toronto, Ontario, Canada, June2015. p. 4480.[15] Datta A, Nishimura D. Field map combination method for multiple-acquisition bSSFP.In Proceedings of the 25th Annual Meeting of ISMRM, Honolulu, May 2017. p. 0454.[16] Winkelmann S, Schaeffter T, Koehler T, Eggers H, Doessel O. An optimal radial profileorder based on the golden ratio for time-resolved MRI. IEEE Trans Med Imaging 2007;26:68–76. 223 ist of Supporting Information Videosist of Supporting Information Videos