Balanced Steady-State Free Precession Cine MR Imaging in the Presence of Cardiac Devices: Value of Interleaved Radial Linear Combination Acquisition With Partial Dephasing

Abstract

Background:

Balanced steady-state free precession (bSSFP) is important in cardiac MRI but suffers from off-resonance artifacts. The interpretation-limiting artifacts in patients with cardiac implants remain an unsolved issue.

Purpose:

To develop an interleaved radial linear combination bSSFP (lcSSFP) method with partial dephasing (PD) for improved cardiac cine imaging when implanted cardiovascular devices are present.

Study Type:

Prospective.

Phantom and Subjects:

Flow phantom adjacent to a pacemaker and 10 healthy volunteers (mean age ± standard deviation: 31.9 ± 2.9 years, 4 females) with a cardioverter-defibrillator (ICD) positioned extracorporeally at the left chest in the prepectoral region.

Field Strength/Sequence:

A 3-T, 1) Cartesian bSSFP, 2) Cartesian gradient echo (GRE), 3) Cartesian lcSSFP, and 4) radial lcSSFP cine sequences.

Assessment:

Flow artifacts mitigation using PD was validated with phantom experiments. Undersampled radial lcSSFP with interleaving across phase-cyclings and cardiac phases (RLC-SSFP), combined with PD, was then employed for achieving improved quality of cine images from left ventricular short-axis view. The image quality in the presence of cardiac devices was qualitatively assessed by three independent raters (1 = worst, 5 = best), regarding five criteria (banding artifacts, streak artifacts, flow artifacts, cavity visibility, and overall image quality).

Statistical Tests:

Wilcoxon rank-sum test for the five criteria between Cartesian bSSFP cine and RLC-SSFP with PD. Fleiss kappa test for inter-reader agreement. A $P$ value < 0.05 was considered statistically significant.

Results:

Based on simulations and phantom experiments, 60 projections per phase cycling and 1/6 PD were chosen. The in vivo experiments demonstrated significantly reduced banding artifacts (4.8 ± 0.4 vs. 2.7 ± 0.7), fewer streak artifacts (3.7 ± 0.6 vs. 2.6 ± 0.7) and flow artifacts (4.4 ± 0.4 vs. 3.7 ± 0.6), therefore improved cavity visibility (4.1 ± 0.4 vs. 2.9 ± 0.9) and overall quality (4.0 ± 0.4 vs. 2.7 ± 0.7).

Data Conclusion:

RLC-SSFP method with PD may improve cine image quality in subjects with cardiac devices.

Cardiovascular implantable electronic devices have previously been regarded a contraindication for MRI scanning. Recently, safety studies have demonstrated an acceptable safety profile and absence of device or lead failure for patients with a pacemaker or implanted cardioverter-defibrillator (ICD) when undergoing MRI at 1.5 T.^1,2 However, while cardiac MRI can be performed safely in these patients, image quality is a concern, given that large ${\text{B}}_0$-field inhomogeneity artifacts can be introduced by the implants, especially adjacent to the heart.³ The balanced steady-state free precession (bSSFP) method is the workhorse of cardiac MRI, and it is used not only for cine imaging but also for T1 and T2 parametric mapping.^4–6 Specifically, bSSFP provides a high signal-to-noise ratio (SNR) and excellent myocardium-blood contrast, which can be achieved with a short repetition time (TR) and thus a high temporal resolution.⁴ However, in patients with ICDs, device-related artifacts can be present in 94% of the bSSFP images and can become relevant to volume calculation in 38%,³ which is due to the well-known banding patterns and flow artifacts in the presence of off-resonance induced by the devices.^4,7,8

In clinical practice, when imaging patients with metallic implants, careful consideration of patient positioning, pulse sequences, and scan parameter choices is required.^9,10 Gradient echo (GRE) imaging has some advantages over bSSFP due to its lower sensitivity to ${\text{B}}_0$-field inhomogeneity in device patients.¹¹ Artifacts are visible in 75% of the GRE cine images and can become relevant in 6%.³ Wideband late gadolinium enhancement (LGE) has shown to be very useful in cardiac MRI to improve image quality by improving the inversion pulse.^12,13 However, no such improvements have been developed and tested for bSSFP-based methods, cine, or T1 or T2 mapping. Typically, GRE cine is used when severe artifacts are present,¹¹ which results in lower temporal resolution, lower contrast-to-noise ratio (CNR), and potentially less accurate contouring of heart chambers to measure diagnostic volumes, since most modern automated segmentation software solutions have been designed for bSSFP cine contrast.^14,15

Linear combination bSSFP (lcSSFP) imaging is an established way to reduce off-resonance artifacts, approximating a uniform spectral response to off-resonance.¹⁶ The immediate problem is that the increased scan time (usually 4-fold) makes it an impractical solution for cardiac cine imaging (more than 40 seconds of breath-hold). However, radial acquisition can be easily implemented with different acceleration factors compared to Cartesian lcSSFP. Using an undersampled radial lcSSFP method provides more reasonable scan times and therefore feasible breath-holds, which has been demonstrated in bSSFP cine for the off-resonance pulmonary veins.¹⁷ Furthermore, flow artifacts in bSSFP due to off-resonance can also impact image quality. The contribution from spins flowing out of the imaged slice can substantially broaden the slice thickness, introduce transient-related artifacts, and near-band hyperintensity that can be over 400% of the signal in absence of flow.⁷ One approach to reduce these artifacts is partial dephasing (PD),¹⁸ which uses a slightly unbalanced bSSFP acquisition that decreases hyper-enhancement at the stopbands from out-of-slice signal contribution, and thus greatly mitigates off-resonance flow artifacts. However, this approach has not been tested in the presence of devices.

To enable bSSFP cine imaging in the presence of an implanted cardiovascular electronic device, we aimed to investigate three different types of artifacts in this study. Specifically, we explored¹ lcSSFP to remove the banding artifacts generated by the device,² an undersampled radial method for feasible breath-hold times with only modest streak artifacts, and³ PD to mitigate flow artifacts. We aimed to explore interleaved radial lcSSFP (RLC-SSFP) with PD combining these three techniques.

Materials and Methods

This prospective study was approved by our institutional review board. All subjects provided written informed consent.

MRI Sequence

Figure 1 illustrates the undersampled pattern of the radial acquisition. Four-pass lcSSFP with four repeated acquisitions or dynamics were acquired, each having a different phase cycling with a radiofrequency (RF) phase increment in each TR of 0°, 90°, 180°, and 270°, respectively (Fig. 1a). The four-pass lcSSFP was preceded with a dummy heartbeat to approach steady state. By combining these four acquisitions, banding artifacts from the standard bSSFP were expected to be reduced.

FIGURE 1:

Illustration of the proposed RLC-SSFP sequence. (a) Interleaved radial acquisitions from each phase cycling can be linearly combined to reduce banding artifacts in cine images. Three heartbeats per dynamic were shown in this figure for simplicity. Five heartbeats were used in the in vivo experiments. (b) Example of one set of segments from the first heartbeat in the first dynamic (blue box in a). The projections are interleaved from phase to phase (see the dotted gray lines if without this interleaving) to take advantage of the temporal correlation, as in UNFOLD. (c) The prephasing moment of slice selection gradients was slightly unbalanced to attenuate the phase coherence of out-of-slice flowing spins.¹⁶

To keep the undersampling artifacts within an acceptable range, two interleaving strategies were implemented. First, the radial views from each acquisition were interleaved by adding a different initial angle offset, alongside with a rearranged sampling pattern where the most complementary sets of radial views were matched to the adjacent phase cycling (eg with Np projections acquired for each dynamic, initial angle $(0\pi )/(4\text{Np})$ and $(2\pi )/(4\text{Np})$ were matched to first and second phase cyclings and $(1\pi )/(4\text{Np})$ and $(3\pi )/(4\text{Np})$ were matched to third and fourth phase cyclings, instead of sequentially increasing the initial angle by $(1\pi )/(4\text{Np})$ in each phase-cycling; Fig. 1a). Since the radial streaks differ in orientation among the four interleaved dynamics, it has been demonstrated that they can be reduced with their combination.¹⁷ Second, projections from the odd and even cardiac phases were also interleaved (i.e. different by $(1\pi )/(2\text{Np})$, Fig. 1b) so that unaliasing in the temporal domain (UNFOLD)¹⁹ can be performed to reject the Nyquist aliasing.

This interleaved radial lcSSFP can be performed with different number of phase cyclings. We also tested a three-pass lcSSFP, each dynamic having a RF phase increment of 60°, 180°, and 300°, respectively (instead of 0°, 120°, and 240°). The 0° phase-cycling usually gave the worst images while the 180° phase-cycling gave the best in our study, so we avoided 0° and preserved 180° phase-cycling in the three-pass lcSSFP.

The near-band flow artifacts are mostly due to the transverse magnetization from spins flowing out of the imaged slice.⁷ Those flowing spins do not experience additional RF excitation but are not excluded from the gradient activity, therefore could still contribute to the total SSFP signal, resulting in transient signal and hyperintensity. Figure 1c shows the slightly unbalanced prephaser and rephaser of the bSSFP, achieved by reducing the prephasing gradient moment in the slice selection direction. This technique has been demonstrated to be robust to oscillations during the approach to steady state, eddy currents, as well as through-plane flow artifacts by attenuating the phase coherence of out-of-slice flowing spins.^18,20,21

The sequence used and shown in Fig. 1 was similar to previous study¹⁵ except with the addition of UNFOLD and PD option.

Simulations

All simulations were performed in Matlab (Matlab 2019b; MathWorks, Natick, MA, USA). To investigate the effect of the two interleaving strategies (with interleaved phase-cycling dynamics, idSSFP, and additional interleaved phases, ipSSFP, which we call RLC-SSFP), cine images (a bSSFP cine from one subject without device placement) within a simulated inhomogeneous B0 field were compared. We did not simulate coils, since this was not an essential innovation of our study. The superimposed banding patterns for each phase cycling were based on Bloch equation-derived expression,⁴ assuming a T1/T2 = 1800 msec/200 msec, and TR/flip angle = 3.5 msec/35°. The SNR (using the center region of the left ventricular blood as a region of interest and the left upper corner as background noise, averaged across the cardiac cycle) and root mean square error (RMSE, using the original cine images as reference) were calculated. Different white Gaussian noise levels (from 10 dB to 80 dB on each radial readout, assuming the k-space signal from reference image was noise-free; noise outside this range either corrupted the whole image or did not introduce recognizable noise in image), and different numbers of projections (Np, from 24 to 192 in each phase cycling) were simulated. For four-phase cycling required in lcSSFP, 48–60 projections for each phase cycling allowed acceptable scan times (13–17 seconds), and therefore a focus of the simulations. A 180-degree span (i.e. with Np projections, $\text{angles}=\left[i^{\ast }180/\text{Np}\right]$, for $i=0$ to Np-1) was used to limit eddy current artifacts in the acquired images.

To show the influence of flowing spins, the frequency responses were simulated with different flow velocities (from 0 to 100 cm/sec to represent the typical range of blood flow velocities) and different amounts of prephasing moment reduction (from 0, fully balanced, to 1/2, no prephasing), using the assumption of imperfect slice profile based on the applied Hamming-windowed sinc pulse and out–of-slice contribution at velocity of 0 (no flow) and 100 cm/sec.⁷ We also investigated how the banding behavior in lcSSFP changed with and without PD.

Devices

Implantable cardiac rhythm management devices were provided by Medtronic (Minneapolis, MN, USA). As size of the generator may be expected to influence extent of artifact, devices were chosen to reflect the range in size of current devices used clinically, and thus included one pacemaker (Advisa DR MRI), two implantable cardioverter defibrillators (ICD, Evera MRI XT, DR and Evera XT DR), and two ICDs with capacity for cardiac resynchronization (CRT-D, the largest device, Viva quad XT CRT and Cobalt CRT-D MRI). We included both devices approved as “MRI conditional,” and those not approved for MRI, as both undergo MRI frequently in the current era.^1,2 To eliminate any risk of shock to the subjects or research staff during the experiments, all devices had detection and therapy turned off and were insulated (Ultem Polyetherimide film, 3M, St Paul, MN, which can withstand 3000 V, larger than the 750 V maximum delivered by an ICD). Plastic plugs were inserted into the high-voltage ports with medical adhesive applied to the setscrew. After use, the devices were interrogated using a doughnut magnet, whose differential tone indicated whether the ICD was still turned off, and later with a programmer.

Phantom Experiments

All imaging experiments were conducted with a 3-T scanner (Magnetom Prisma, Siemens Healthineers, Erlangen, Germany). A phased-array coil with 34 channels was used. The Advisa pacemaker was affixed onto the wall of a cylindrical plastic container to generate the inhomogeneous field, around which a tube (filled with gadolinium-doped water) was wrapped multiple times. Flow was generated by using a peristaltic pump motor (Omegaflex, Stamford, CT, USA) with an average velocity of 80 cm/sec. To show the banding patterns and flow artifacts, images from two slices (distal and proximal to the device) by GRE, standard Cartesian bSSFP, and the proposed RLC-SSFP were compared.

Scan parameters for two-dimensional (2D) Cartesian GRE cine were as follow: TR/echo time $(\text{TE})/\theta =5.3\text{msec}/2.6\text{msec}/{12}^{\circ }$, number of phase encodings = 100, field of view (FOV) = 200 mm, slice thickness (ST) = 3 mm, prospectively electrocardiogram (ECG) triggered using 50 readouts per segment, 192*192 matrix, 1.0 mm*1.0 mm spatial resolution, GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA)²² with 24 integrated reference lines and acceleration factor R = 2. For Cartesian bSSFP cine imaging, the scan parameters were identical except that $\text{TR/TE/}\theta =3.4\text{msec}/1.7\text{msec}/{40}^{\circ }$. For the interleaved radial method, the scan parameters were the same as for Cartesian. The acquisition was repeated with four-phase cyclings (or four averages), Np = 100.

In Vivo Experiments

Ten healthy subjects were included (mean age ± standard deviation: 31.9 ± 2.9 years, 4 females). The proposed method was evaluated using a working ICD, reprogrammed and prepared as outlined above for device specifications, positioned extracorporeally on the chest in the left prepectoral area below the clavicle, mimicking clinical placement. Cartesian bSSFP were compared with the proposed RLC-SSFP with PD. GRE and Cartesian lcSSFP were also tested.

Scan parameters for 2D Cartesian bSSFP cine were as follows: $\text{TR}/\text{TE}/\theta =3.4\text{msec}/1.7\text{msec}/35-{40}^{\circ }$ (constrained by specific absorption rate limits), number of phase encodings = 192, FOV = 380 mm, ST = 8 mm, retrospectively ECG-gated using 12 readouts per segment, 192 × 192 matrix, 2.0 mm × 2.0 mm spatial resolution, 16 heartbeat breath-hold. For the interleaved radial method, the scan was repeated with four-phase cycling, 60 radial projections per k-space, and 21 heartbeat breath-holds, achieving the same spatial and temporal resolution with an acceptable 4 seconds increase of acquisition time. In addition, we also investigated methods to reduce scan time further, either by using fewer radial projections in each phase cycling (Np = 30) or by employing fewer phase cyclings (three-phase cycling, each having RF phase increment of 60°, 180°, and 300°°, as described earlier). GRE cine imaging used the same parameters as Cartesian bSSFP except $\text{TR}/\text{TE}/\theta =5.3\text{msec}/2.6\text{msec}/{12}^{\circ }$. Finally, we also tested the possibility of Cartesian lcSSFP (4 averages, 81.3% FOV phase, 70% phase resolution, 6/8 partial Fourier, 192 × 156 matrix, GRAPPA with 24 integrated reference lines and acceleration factor R = 2, 21 heartbeats).

Reconstruction

The k-space data from Cartesian acquisitions were all reconstructed online. The radial acquisitions were reconstructed offline as shown in Fig. 2. First, prewhitening was conducted with a noise covariance matrix estimated through a prescan with RF excitation off. The data were projected into Cartesian k-space using gridding²³ separately and then combined in a complex sum.

FIGURE 2:

(a) The reconstruction flow chart. Preprocessing was performed in Matlab and iterative nonlinear inversion reconstruction was performed with the BART toolbox on GPU. The final outputs are images and coil sensitivities. (b) Coil selection was performed to reject some coils. The B0 field of the device can warp the gradient fields, and thus serve as a source point of streaks (See coil #2 and #3), which can be removed with coil rejection before the parallel imaging reconstruction. Intensity of coil #1 was multiplied by 5 for display purpose (purely streaks thus should be rejected, but intensity much smaller than the desired signal so not obvious to see in the coil-combined images).

Prior to iterative parallel imaging reconstruction, coil rejection was used.²⁴ Specifically, this was used to remove coils that contained artifacts due to data inconsistency caused by the off-resonant region and its warping of the gradient field. Automatic outlier pruning and coil removal (OPCR) was based on the streak ratio (SR), defined as $S{R}_j=\frac{{\Vert I_j-L_j\Vert }_2}{{\Vert L_j\Vert }_2}$, assuming that a low-resolution image $L_j$ from k-space center is streak-free. If the coil image $I_j$ is too different from $L_j$, it was rejected immediately. An empirical reject SR threshold of 0.15 was used in our study. Then coils were clustered into two groups by k-means based on the calculated SR. The coil group with higher SR was also rejected before the coil combination in the following iterative reconstruction, similar to a previous study.²⁴ The utility of this is evident in Fig. 2b.

Next, the graphics processing unit (GPU)-accelerated calibrationless nonlinear inversion joint reconstruction (NLINV)^25,26 was implemented to reduce aliasing due to undersampling via parallel imaging. Specifically, NLINV simultaneously generates the images and coil sensitivities based on the measured k-space data. Balancing time and image quality, we used 12 iteratively regularized Gauss-Newton steps with L2 norm penalty on images, based on the BART toolbox²⁷ (version 0.7.00, https://mrirecon.github.io/bart/index.html). NUFFT²⁸ without parallel imaging was used for comparison. Finally UNFOLD processing was performed by applying a low-pass Fermi filter in the temporal Fourier space for each image voxel. UNFOLD was applied after combining all the phase cycling and coils and decreasing the handled data size for computational efficiency.

Image Quality Analysis

To compare the Cartesian bSSFP and RLC-SSFP, image quality evaluation was completed by three independent readers (J.X., J.L., D.C.P. with 1, 8, and 26 years of cardiac MRI experience), who graded the following five criteria using a 5-point scale (from 1 = worst to 5 = best) on images in all the subjects: banding artifacts, streak artifacts (between RLC-SSFP and noninterleaved radial lcSSFP, because no streak should be present in Cartesian bSSFP), flow artifacts, cavity visibility, and overall image quality. For the three artifact scores grading consisted of the following scores: 5 = artifact not present, 4 = mild and not impacting quality, 3 = moderately impairing quality, 2 = severely impairing quality, 1 = artifact resulting in nondiagnostic images. See the example cases in supplemental materials Video S1.

Statistical Analysis

The image quality scores were analyzed using a nonparametric test (two-sided Wilcoxon rank-sum test in Matlab) to compare the standard bSSFP, the noninterleaved radial lcSSFP, and the proposed RLC-SSFP with PD. The unpaired test was performed in order to use all of the data because some studies did not acquire all the sequences. The analysis was based on the average score from the three readers. Inter-reader agreement was assessed using the Fleiss kappa test toolbox (http://www.mathworks.com/matlabcentral/fileexchange/15426). Agreement was considered slight for kappa values of 0.00–0.20, fair for values of 0.21–0.40, moderate for values of 0.41–0.60, substantial for values of 0.61–0.80, and almost perfect for values of 0.81–1.00. A $p$ value < 0.05 was considered statistically significant.

Results

Simulation

Off-resonance artifacts were reduced with four-pass lcSSFP (Np = 60) compared with standard bSSFP acquisitions at both end-of-diastole and end-of-systole (Fig. 3a). Figure 3b plots metrics of undersampling artifacts (normalized apparent SNR and RMSE) for lcSSFP, idSSFP, and ipSSFP. Three fully sampled acquisitions showed similar SNR (Fig. 3b, SNR-Noise, Np = 192). Interleaved radials helped to reduce the streaks in undersampled acquisitions, which compete with the input noise, thus provided doubled SNR (Fig. 3b, SNR-Noise, Np = 60). Without any interleaving strategies, the background of highly undersampled lcSSFP was dominated by the streaks, so the apparent SNR was independent on input noise (Fig. 3b, SNR-Noise, Np = 24). To balance image quality and total acquisition time, we chose 60 projections for each dynamic, which provided equivalent SNR to that of idSSFP (/lcSSFP) using 100 (/140) projections. In addition, when high noise presented (10 dB), RMSE of ipSSFP using 60 projections almost reached the same level of fully sampled lcSSFP (Np = 192) in our simulation.

FIGURE 3:

Simulations of interleaved radial lcSSFP. (a) Reconstructed images with 60 projections per phase-cycle at end of diastole and end of systole within a simulated inhomogeneous field. Bands were all reduced. With interleaved dynamics (idSSFP) and additional interleaved phases (ipSSFP), the streaks were lessened. (b) Normalized apparent SNR (nSNR, divided by square root of Np and then normalized by SNR of ipSSFP at 80 dB noise and Np = 192) and RMSE at full sampling behaved the same for lcSSFP, idSSFP, and ipSSFP, but streaks dominated the background noise in highly undersampled lcSSFP. The apparent SNR of ipSSFP using 60 projections was equivalent to the SNR of idSSFP (/lcSSFP) using 100 (/140) projections. When high noise presented (10 dB), RMSE of ipSSFP using 60 projections almost reached the same level of fully sampled lcSSFP (Np = 192).

Figure 4 shows simulations regarding PD. Spins flowing out of the imaged slice kept contributing to the total signal, therefore leading to image hyperintensity (Fig. 4a). By reducing the prephasing moment (Fig. 1c), the frequency response became less dependent on the flow velocity (Fig. 4b). However, a larger PD resulted in expansion of the off-resonance bands or stopbands, even in the absence of flow. Indeed, a less sudden reduction in signal intensity improved the bands elimination with lcSSFP (one example shown in Fig. 4c, using only three-phase cycling) but also decreased the on-resonance signal (Fig. 4b). A good trade-off was found when using a PD factor of 1/6, as indicated by mitigating the flow artifacts and smoothing the band while maintaining a high SNR of bSSFP.

FIGURE 4:

A simulation of the partial dephasing method considered imperfect slice profile and spins flowing out of the slice (a), which do not experience excitation but continue contributing to the total signal and cause hyperintensity dependent on their velocity. The signal vs. frequency response with partial dephasing smooths, reduces the hyperintensity, and its dependence on the flow velocity. In the absence of flow, partial dephasing also expanded the stopband and lowered on-resonance signal (dotted arrow in b). The less sudden drop in the response profile helped a better removal of banding artifacts (solid arrow in c). 1/6 PD was chosen for in vivo experiments.

Phantom Experiments

In the phantom experiments (Fig. 5a), GRE was more robust to the off-resonance and flow artifacts but manifested greater regional dephasing and lower SNR (Figure 5b). In bSSFP, bands were presented in both slices while slice 1, which was closer to the pacemaker had stronger artifacts. In lcSSFP, bands were reduced but residual rippling was still visible at a four times greater spatial frequency, whereas PD further flattened the images and improved the overall quality. With flow on, strong flow artifacts manifested as hyperintensity and transient-related blurring, which were greatly diminished by PD.

FIGURE 5:

Static and flow phantom experiments. (a) Phantom showing the device adjacent to tubing, and two imaged slices (indicated by orange and yellow). (b) The device induced banding was reduced using RLC-SSFP (dashed arrow). GRE showed lower CNR and greater dephasing (yellow circle). When flow was on, new artifacts (solid arrow) were generated, which could be greatly improved using partial dephasing.

In Vivo Experiments

Ten subjects were imaged, and no data were discarded. Of these, seven had complete protocol that included Cartesian bSSFP, noninterleaved radial lcSSFP, and RLC-SSFP with and without PD. We observed that short-axis slices closest to the ICD presented more severe artifacts, impeding visualization of the left ventricular (LV) and right ventricular (RV) walls, which was consistent with the phantom experiment. One noted aspect was how the off-resonant region can serve as a source point of streaks, by opposing the gradient encoding so that the region appears to be a hyperintense point-object. OPCR can remove these streaks, prior to reconstruction (Fig. 2b). Our proposed method of RLC-SSFP removed the banding and flow artifacts in all slices (Fig. 6).

FIGURE 6:

A stack of short-axis slices. Severely corrupted image quality was recovered with the proposed method, showing eliminated banding artifacts (dashed arrow), suppressed hyperintensity and flow-related transient artifacts (solid arrow).

When analyzed in the image-temporal domain, the standard bSSFP images for a representative subject showed corrupted quality from banding, whereas the radial lcSSFP method recovered the image quality (one example in Fig. 7). In addition, in this subject, the high SNR (using a center region of the left ventricular blood as region of interest and the background region unaffected by aliasing as noise) and myocardium-blood contrast of bSSFP was maintained when compared to GRE. Contrast with lcSSFP was especially improved in the end-systolic phases. One Cartesian lcSSFP example (with PD and online reconstruction) is presented for comparison in Fig. 7.

Figure 7:

Comparison of (a) Cartesian bSSFP, (b) RLC-SSFP, (c) GRE, and (d) Cartesian lcSSFP in a volunteer with a device. The profile across the heart is plotted for each frame, showing better temporal fidelity using RLC-SSFP, compared to Cartesian, and better contrast compared to GRE, and improved quality compared to bSSFP. The estimated mean CNR for this subject across the cardiac cycling was 20.1 ± 5.7 vs. 84.8 ± 16.4, and SNR 44.8 ± 1.8 vs. 111.1 ± 16.6, for GRE and lcSSFP, respectively (even though SNR was reduced due to streaking artifacts). The Cartesian lcSSFP shows good quality but was more blurred than radial lcSSFP due to the accelerated protocol.

Figure 8a shows the reduced streaks with interleaved radial views. In Fig. 8b, image quality improved with reduced flow artifacts using PD (examples of the cine movies in supplementary material Video S2). Figure 8c shows that image quality improved in 10 subjects for all five criteria. Analysis of the inter-reader agreement revealed moderate agreement for banding and streak artifacts (kappa of 0.41 and 0.59, respectively, P < 0.05), and fair agreement for the rest (0.32, 0.27, and 0.29, respectively, P < 0.05). Compared with standard Cartesian bSSFP, the proposed RLC-SSFP had fewer banding artifacts (4.8 ± 0.35 vs. 2.7 ± 0.66) and flow artifacts (4.4 ± 0.38 vs. 3.7 ± 0.62), elevated cavity visibility (4.1 ± 0.44 vs. 2.9 ± 0.92), and better overall quality (4.0 ± 0.40 vs. 2.7 ± 0.74). In addition, compared with noninterleaved radial lcSSFP, fewer streak artifacts were observed (3.9 ± 0.38 vs. 2.8 ± 0.76). The lcSSFP sequence did not address the problem of flow artifacts and thus had a similar score when compared to bSSFP for this criterion (3.74 ± 0.62 vs. 3.80 ± 0.41, P = 0.93 for bSSFP and lcSSFP, respectively). Furthermore, it was characterized by less desirable cavity visibility (3.6 ± 0.73 vs. 4.1 ± 0.44, P = 0.13) and overall score (3.3 ± 0.71 vs. 4.0 ± 0.40) than the same method with PD.

FIGURE 8:

Improved cine using lcSSFP. (a, b) One of the better (a, subject 1) and worst (b, subject 2) cines, showing reduced artifacts: banding (dashed arrow), streaks (dotted arrow), and flow (solid arrow) artifacts. (c) Statistical analysis of image quality scores showed that banding was improved with lcSSFP, flow-artifacts were improved with partial dephasing (PD), and in general cavity visibility and overall quality were higher using the methods developed here.

The parallel imaging method using NLINV reconstruction was less generative of streaks at 30 projections (Np = 30, total acquisition time around 11 seconds), compared to NUFFT (Fig. 9b), but small structures with lower intensity might be lost. Three-pass lcSSFP was faster (around 13 seconds) and worked in most of our experiments but was not as robust as four-pass lcSSFP (one example of poor image quality three-pass lcSSFP is shown in supplementary material Video S3). A possible reason for less robustness of three-pass lcSSFP could be that with only three phase-cyclings, contributions from a highly artifacted phase-cycling become more important. Application of PD helped not only with flow artifacts but also with banding artifacts, as shown in the three-pass lcSSFP images of Fig. 9c.

FIGURE 9:

Acquisition time can be further reduced. (a) Cartesian bSSFP in the presence of a cardaic device. (b) With fewer projections (Np = 30) in each phase cycling, streaks can be further removed by NLINV with a total variation penalty. (c) Partial dephasing has potential in helping banding removal, as shown here using three-phase cycling.

Discussion

We propose RLC-SSFP combining three techniques to improve bSSFP cine image quality, which is related to off-resonance artifacts in an extremely distorted field caused by a cardiovascular electronic device. The performance of lcSSFP in the presence of devices has not been explored, and there are no studies using it to improve cardiac cine in patients with cardiac implants. Banding artifacts were removed in simulations, phantoms, and during in vivo experiments with our RLC-SSFP method. The interleaved radial acquisition across the four dynamics and cardiac phases generated fewer streaks. Our results suggest the capability to further reduce the stopbands of bSSFP, which could provide better reconstructed images even in the absence of flow. Residual rippling was still visible in a four-pass lcSSFP at a four times greater spatial frequency, and this rippling was reduced by PD.

The GRE cine technique was robust to the device-induced artifacts in our study but provided lower SNR and myocardium-blood contrast. However, standard bSSFP contrast is preferred since many LV contouring methods, especially using machine learning-based automated methods,^14,15 were developed for bSSFP and utilize the higher contrast (thus easily detectable edges). Parametric mapping (T1 and T2 mapping) is usually acquired with bSSFP contrast and is less reliable using GRE versions (especially for T1 mapping).⁶ Therefore, bSSFP quality improvement is still needed in patients with devices, even though GRE cine is usually almost free of artifacts.

A prior study used undersampled radial lcSSFP with interleaved projections for each phase cycling to improve pulmonary vein cine images.¹⁷ However, in this previous study, the off-resonance was due to the lungs and not the much more extreme off-resonance encountered with cardiac devices. Because of the more challenging situation of devices, we implemented an interleaving strategy in the temporal domain, and parallel imaging reconstruction to further improve the image quality, while flow artifacts were not addressed as in the present study by using PD.

Although our current method acquired the four dynamics in a single breath-hold (21 heartbeats, around 17 seconds), further improvements in breath-hold time and artifact suppression are needed. Potential solutions include acquiring fewer projections in each phase cycling combined with more advanced reconstructions, such as NLINV with variational penalties.²⁶ Alternatively, three-pass lcSSFP (16 heartbeats) worked well in many subjects but may require larger PD as banding artifacts are less reduced. RLC-SSFP technique would benefit from any techniques which reduce radial undersampling artifacts. Here, we used parallel imaging, UNFOLD, and coil-rejection based on artifact level, but other acceleration approaches might also be valuable.^24,29

Limitations

A small number of subjects were included, but the statistics analysis showed significantly improved image quality. The positioning of the ICD on the subject chest (and the absence of the implanted leads) instead of a clinical implantation is another limitation of this study. However, it seems unlikely that this considerably impacted the image quality improvement. Future studies will evaluate these MR techniques in patients with clinically implanted full ICD systems, as well as including generators which may be placed in positions other than the typical prepectoral area (such as the subcutaneous device), and those made by other manufacturers. While we did not quantitatively compare heart chamber volumes, the distortion observed in bSSFP images was so extensive that an obtained improvement seemed evident. Finally, the study was conducted using a 3-T MRI system instead of a 1.5-T MRI system, which is more frequently used in clinical cardiac MRI. As 3-T imaging of patients with implanted devices is more challenging due to the larger B0 inhomogeneity and thus off-resonance,³⁰ we would expect much better results at 1.5 T.

Conclusion

Our proposed RLC-SSFP sequence implementing PD may reduce banding, flow, and radial streak artifacts, thus improving cardiac cine image quality in patients with cardiac devices without compromises for spatial or temporal resolution.