Highly accelerated real-time free-breathing cine CMR for patients with a cardiac implantable electronic device

Abstract

Rationale and Objectives:

To develop a 16-fold accelerated real-time, free-breathing cine CMR pulse sequence with compressed sensing reconstruction and test whether it is capable of producing clinically acceptable summed visual scores(SVS) and accurate left ventricular ejection fraction(LVEF) in patients with a cardiac implantable electronic device(CIED).

Materials and Methods:

A 16-fold accelerated real-time cine CMR pulse sequence was developed using gradient echo readout, Cartesian k-space sampling, and compressed sensing. We scanned 13 CIED patients (mean age=59 years; 9/4 males/females) using clinical standard, breath-hold cine and real-time, free-breathing cine. Two clinical readers performed a visual assessment of image quality in four categories (conspicuity of endocardial wall at end diastole, temporal fidelity of wall motion, any artifact level on the heart, noise) using a 5-point Likert scale (1: worst; 3: clinically acceptable; 5: best). SVS was calculated as the sum of four individual scores, where 12 was defined as clinical acceptable. The Wilcoxon signed-rank test was performed to compare SVS, and the Bland-Altman analysis was conducted to evaluate the agreement of LVEF.

Results:

Median scan time was 3.7-times shorter for real-time (3.5 heartbeats per slice) than clinical standard (13 heartbeats per slice, excluding non-scanning time between successive breath-hold acquisitions). Median SVS was not significantly different between clinical standard (15.0) and real-time (14.5). The mean difference in LVEF was −2% (4.7% of mean), and the limits of agreement was 5.8% (13.5% of mean).

Conclusion:

This study demonstrates that the proposed real-time cine method produces clinically acceptable SVS and relatively accurate LVEF in CIED patients.

Introduction

Although cardiovascular magnetic resonance (CMR) can be performed in cardiac implantable electronic device (CIED) patients with manageable risk (1–5), the diagnostic quality of standard CMR may be low in CIED patients due to substantial device-related image artifacts occurring in 33–48% of cases (6,7). The device-related image artifacts are caused primarily by the pulse generator, less so by intracardiac leads (8). The static magnetic field (B0) across the heart varies on the order of kHz (9), where B0 inhomogeneity severity depends on the size of the CIED and its position relative to the heart. Large B0 variations preclude use of balanced steady state free precession (b-SSFP), which is the standard readout for cine CMR (10). Another sources of image artifacts in CMR are arrhythmias and/or dyspnea that are common in CIED patients. Thus, there is a need to develop robust cardiac cine pulse sequences that address image artifacts associated with CIED, arrhythmia, and/or dyspnea.

One approach to suppress image artifacts associated with CIED is to reduce the echo time (TE) using a combination of short excitation RF pulse and high receiver bandwidth. One proven approach to suppress image artifacts associated with arrhythmia and/or dyspnea is performing highly-accelerated real-time cine CMR during free breathing with clinically acceptable spatial (~2 mm × 2 mm) and temporal (~40 ms) resolutions. Several groups have reported highly-accelerated real-time cine CMR using b-SSFP readouts with Cartesian (11,12) or non-Cartesian k-space trajectories (13–15) or gradient echo readout (GRE) with non-Cartesian k-space trajectories in non-device patients (16,17). These methods, however, may not suitable for CIED patients due to severe off-resonance effects (e.g. banding artifacts for b-SSFP, geometric distortion for non-Cartesian) induced by the device. Accelerated real-time cine CMR offers two key advantages over standard retrospective ECG-gated cine CMR. First, real-time cine is capable of drastically reducing the scan time, particularly since it can be run continuously during free-breathing to sample multiple slices, thereby eliminating non-scanning time (e.g. breathing instruction, recovery from breath-holding) between successive breath-hold acquisitions. Second, real-time cine is capable of reducing RF energy deposition (i.e. less RF excitation per image acquisition), thereby providing a greater margin for MR safety. The purposes of this study were to develop a 16-fold accelerated, real-time, cine CMR pulse sequence using GRE with Cartesian k-space sampling and compressed sensing (CS)(18) and test whether it is capable of producing clinically acceptable image quality, accurate left ventricular functional parameters, and reduced scan time compared with clinical standard cine pulse sequence in CIED patients at 1.5 Tesla.

Materials and Methods

Patients

This study was conducted in accordance with protocols approved by our institutional review board and was Health Insurance Portability and Accountability Act (HIPAA) compliant. All subjects provided informed consent in writing. We recruited 13 patients (9 males, 4 females; median age = 60, interquartile range [IQR] of age = 10; median left ventricular ejection fraction [LVEF] = 47%, IQR of LVEF = 28 %) with a CIED undergoing clinical CMR examination for evaluation of myocardial scarring or diagnosis of new cardiac symptoms. For each patient, the rhythm status during CMR was estimated by extracting the mean and standard deviation of R-R interval embedded in the Digital Imaging and Communication in Medicine (DICOM) header of retrospective breath-hold ECG-gated cine images. Table 1 summarizes patient characteristics, including risk factors for myocardial ischemia and device types.

Table 1.

Values for age, LVEF, and heart rate represent median and interquartile range. For all other variables, numbers in parenthesis represent percentages. CRT-D: cardiac resynchronization therapy with defibrillator; S-ICD: subcutaneous ICD; LBBB: left bundle branch block; RBBB: right bundle branch block; VT: ventricular tachycardia; VF: ventricular fibrillation; PVC: premature ventricular contraction; NICM: non-ischemic cardiomyopathy; GFR: glomerular filtration rate.

	CIED (N=13)
Ages (years)	60(10)
LVEF (%)	47(28)
Heart rate (bmp)	63(18)
Sex (Male/Female)	9/4 (62%)
ICD	9(69%)
CRT-D	0(0%)
Pacemaker	2(15%)
S-ICD	2(15%)
Diabetes	4(31%)
Hypertension	7(54%)
Obstructive CAD	5(38%)
Atrial Fibrillation	7(54%)
LBBB or RBBB	3(23%)
VT or VF	5(38%)
PVC	0(0%)
Heart Failure	6(46%)

Preliminary Phantom Experiment: Cartesian b-SSFP vs. Cartesian GRE vs. Radial GRE

Clinical standard pulse sequences use Cartesian k-space sampling, because it is simpler to use than non-Cartesian k-space sampling. In highly-undersampled acquisitions, it may beneficial to use non-Cartesian (e.g. radial) k-space sampling over Cartesian k-space sampling, because the former is more robust to motion and produces more incoherent aliasing artifacts. Clinical standard cine MRI uses b-SSFP, because it produces higher contrast-to-noise ratio than GRE. In CIED patients, it may not be feasible to perform b-SSFP imaging due to severe banding artifacts caused by off-resonance.

To determine an optimal readout type and k-space sampling, we performed the following preliminary experiment on an American College of Radiology phantom with a dual-chamber ICD (Evera XT DR, Medtronic, Minneapolis, Minnesota) taped on its side, approximately 10 cm away from the coronal imaging plane. We tested the following three different breath-hold pulse sequences: (a) b-SSFP and Cartesian k-space sampling (i.e. clinical standard for non-device patients), (b) GRE with Cartesian k-space sampling (i.e. clinical standard for CIED patients), and (c) GRE and radial k-space sampling. For the segmented GRE radial acquisition, the scan time was greater than 20 s to ensure Nyquist condition. All three sequences used standard imaging parameters (e.g. spatial resolution = 2 mm × 2 mm × 6 mm, temporal resolution = 40 ms, flip angle = 60° for b-SSFP and 20° for GRE). All three scans were performed during sinus rhythm simulated at 60 beats per minute (bpm). For this phantom experiment, we used standard inline generalized autocalibrating partially parallel acquisitions (GRAPPA)(19) reconstruction for Cartesian k-space acquisitions and inline gridded image reconstruction for radial k-space acquisition. As shown in Figure 1, the GRE sequence with Cartesian k-space sampling produced better images quality than both b-SSFP with Cartesian k-space sampling (severe banding artifacts) and GRE with radial k-space sampling (severe geometric distortion artifact). Based on this preliminary experiment, we determined that GRE with Cartesian k-space sampling is the preferred readout and k-space sampling for in vivo experiment, in which we will compare breath-hold acquisition vs. real-time, free-breathing acquisition. For convenience, clinical standard refers retrospective, ECG-gated, breath-hold GRE with Cartesian k-space sampling, and real-time refers to GRE with 16-fold accelerated, Cartesian k-space sampling.

Figure 1.

ACR phantom images acquired with three different pulse sequences: b-SSFP readout with Cartesian k-space sampling (left), GRE with Cartesian k-space sampling (middle), and GRE with radial k-space sampling (right). As shown, severe banding and distortion artifacts are visible in b-SSFP and radial GRE images, respectively. b-SSFP: balanced steady state free precession; GRE: gradient echo. Yellow arrows point to severe banding artifacts for Cartesian b-SSFP and geometric distortion artifact for radial GRE.

CMR Hardware

All patients were scanned on a whole-body 1.5T MR scanner (MAGNETOM Avanto, Siemens Healthcare, Erlangen, Germany), equipped with a gradient system capable of achieving a maximum gradient strength of 45 mT/m and maximum slew rate of 200 T/m/s. Body coil was used for radio-frequency excitation. Both body matrix and spine coil arrays (15–18 elements) were used for signal reception.

Pulse Sequence

Our clinical default retrospective ECG-gated cine CMR with GRE readout and Cartesian k-space sampling was performed using the following imaging parameters: field of view (FOV) = 340 to 380 (frequency-encoding) × 221 to 273 mm² (depending on patient size), acquisition matrix = 192 (frequency-encoding) × 125 to 154 (depending on patient size), spatial resolution = 1.8 to 2.0 × 1.8 to 2.0 mm², slice thickness = 6 mm, slice gap = 4 mm, number of slices = 12 to 16 depending on patient’s heart size, TR/TE = 10.7/5.2 ms, 7 k-space lines per cardiac phase, temporal resolution = 74.9 ms, flip angle = 15°, receiver bandwidth = 260 Hz/pixel, GRAPPA acceleration factor = 1.8, and scan time = 12–15 heartbeats per slice. During in-line image reconstruction, each cine set was interpolated through time to produce 25 frames per cardiac cycle (e.g. 40 ms interpolated temporal resolution for heart rate of 60 bpm). Please note, this study had no control over these imaging parameters established by our clinical practice.

We modified a GRE pulse sequence to employ a 16-fold accelerated k-space sampling pattern and the following relevant imaging parameters: FOV = 400 × 400 mm², acquisition matrix = 192 × 192, nominal spatial resolution = 2.1 × 2.1 mm², slice thickness = 8 mm, slice gap = 2 mm, number of slices = 10 to 14 depending on patient’s heart size, TR/TE = 3.8/1.8 ms, 12 k-space lines per cardiac phase, temporal resolution = 45.1 ms, flip angle = 20°, receiver bandwidth = 745 Hz/pixel, acceleration factor = 16, and scan time = 3–4 heartbeats per slice (including 1 heartbeat dummy scan to approach steady state of magnetization).

As shown in Figure 2, we used a 16-fold accelerated “Lattice-like” Cartesian k-space sampling pattern with variable density along the phase-encoding direction (i.e. highest density at the central parts of k-space), which was then varied through time. This sampling pattern ensures that all k-space lines are sampled at least once through time, which is important for producing artifact-free, self-calibrated coil sensitivity profiles derived from the time average image, while producing sufficiently incoherent aliasing artifacts, as previously described (20). Note, this sampling pattern looks similar to a Poisson disc with variable density (21).

Figure 2.

A schematic flowchart of the image reconstruction pipeline. As shown (leftmost), a “lattice-like” Cartesian k-space sampling pattern with variable density along phase-encoding (ky) direction is used to achieve 16-fold acceleration. After compressing the number of coil elements to eight, the adaptive phased array method was performed on the time average images to self-calculate the coil sensitivity maps. Zero-filled cine images, multi-coil, raw k-space data, k-space sampling masks, and coil sensitivity maps were used as inputs to the iterative CS algorithm using temporal TV and temporal PCA as two orthogonal sparsifying transforms (30 iterations, conjugate gradient with backtracking line search).F is the undersampled FFT operator, S is the estimated coil sensitivities in x-y space, x is the image series to be reconstructed in x-y-t space, y is the acquired multi-coil k-space data, T₁ is TTV operator and T₂ is temporal PCA operator, and λ is the normalized regularization weight that controls the tradeoff between data consistency and sparsity terms.

Image Reconstruction

The CS image reconstruction was performed off-line on a graphics processing unit (GPU) workstation (Tesla Volta V100 16GB memory, NVIDIA, Santa Clara, California, USA; Xeon E5–2620 v4 256 GB memory, Intel, Santa Clara, California, USA) equipped with Matlab (R2017B, The MathWorks, Natick, Massachusetts) running on Windows 10 (Microsoft, Redmond, Washington). Figure 2 also illustrates a schematic flowchart of our image reconstruction pipeline. We applied principal component analysis (PCA)-based coil compression (22) to derive 8 virtual coils and used the gpuArray functionality in Matlab, in order to speed up processing. During the pre-processing step, coil sensitivity maps were self-calibrated by performing the adaptive array-combination method (23) on the time average image. In the dealiasing step, multi-coil CS reconstruction was performed by enforcing sparsity along the time dimension, where temporal total variation (TTV) and temporal PCA (TPCA) were used as two orthogonal sparsifying transforms and nonlinear conjugate gradient with back-tracking line search was used as the optimization algorithm with 30 iterations. The normalized regularization weight - same for TTV and TPCA - was determined empirically to achieve a good balance between artifact suppression and temporal blurring based on visual inspection on training data. We established 0.01 (or 1% of maximum value in image domain) as an optimal regularization weight by sweeping over a range from 0.005 to 0.2 (0.005 steps) and identifying the highest regularization weight that minimizes temporal blurring of voxels in the heart.

Image Analysis

One radiologist (xxx) with 6 years and one cardiologist (yyy) with 17 years of experience reading cardiovascular MR images evaluated the diagnostic confidence of our high-accelerated real-time cine images relative to clinical standard images. In total, 26 cine datasets (13 patients × 2 cine per patient) were randomized, de-identified, and grouped as a set of three short-axis (base, mid-ventricular, apex) planes for one composite visual score per category during dynamic display at a workstation. Prior to visual evaluation, the two readers were given training datasets to calibrating their scores together, where a score of 3 is defined as clinically acceptable. Following training, each reader was blinded to image acquisition type (clinical standard vs. real-time), each other, and clinical history. Each set of 3 short-axis planes was graded on a 5-point Likert scale: conspicuity of endocardial wall at end diastole (1 = nondiagnostic, 2 = poor, 3 = adequate, 4 = good, 5 = excellent), temporal fidelity (lack of blurring or ghosting) of wall motion (1 = nondiagnostic, 2 = poor, 3 = adequate, 4 = good, 5 = excellent), any visual artifact level on the heart (1 = nondiagnostic, 2 = severe, 3 = moderate, 4 = mild, 5 = minimal), and apparent noise level (1 = nondiagnostic, 2 = severe, 3 = moderate, 4 = mild, 5 = minimal). The summed visual scores (SVS) was calculated as the sum of four categorical scores with 12 defined as clinically acceptable.

LV Function Assessment

Thirty cine data sets (2 sets per patient) were analyzed by another reader (zzz) with one year of experience as a medical research fellow. Image analysis was conducted using standard methods on a workstation equipped with the QMass 7.2 software (Medis, Leiden, Netherlands). Functional parameters included LVEF, end-systolic volume (ESV), end-diastolic volume (EDV), and stroke volume (SV). For consistency, the most basal slice was defined as the plane, which has ≥ 50% of the blood pool surrounded by myocardium, and the most apical slice was defined as the plane showing blood pool at end diastole.

Statistical Analysis

The statistical analyses were conducted by one investigator (aaa) using Matlab. To estimate the arrhythmia burden, for each patient, we calculated coefficient of variation (CV) of R-R intervals as the ratio of standard deviation of R-R interval and mean R-R interval of all slices obtained from DICOM header, multiplied by 100%. Visual scores by two readers were averaged for statistical analysis. Because our sample size (N=13) is relatively small, we elected to perform non-parametric tests throughout. We conducted Wilcoxon signed-rank test to compare two groups and weighted Cohen’s kappa (24) to assess inter-rater reliability. The level of association and agreement between LV functional parameters derived from clinical default and real-time cine images were calculated using the linear regression and Bland-Altman analyses, respectively. Reported values represent median and interquartile range (IQR). A p-value less than 0.05 was considered statistically significant.

Results

All thirteen patients underwent a successful clinical CMR, including the proposed real-time cine CMR. Median CV of R-R interval was 3.7% (IQR = 6.2%), indicating a modest burden of arrhythmia. Median scan time was 3.7-times (P < 0.001) shorter for real-time cine (3.5 [IQR=0.9] heartbeats per slice, including 1 heartbeat of dummy scan) than breath-hold cine (13.0 [IQR=2.0] heartbeats per slice, excluding non-scanning time between successive breath-hold acquisitions). Median CS image reconstruction time was 49 (IQR = 4) sec per slice with 44 time frames.

Figure 3 shows short-axis images of three representative CIED patients with noticeable signal voids induced by the device and/or intracardiac leads. In two out of three cases, the clinical standard cine produced noticeable temporal blurring and/ghosting artifacts on the endocardial wall, whereas the real-time cine minimized such artifacts. For dynamic display, see Video S1 and S2 in Supplemental Materials for clinical standard cine (one heartbeat) and real-time cine (~2 heartbeats), respectively. As summarized in Table 2, median conspicuity score (3.0 for clinical standard vs. 3.5 for real-time), artifact score (3.0 for clinical standard vs. 3.0 for real-time), and SVS (15.0 for clinical standard vs. 14.5 for real-time) were not significantly different. Median temporal fidelity score was significantly (P = 0.046) higher for clinical standard (4.0) than real-time (3.5), whereas median noise score was significantly (P < 0.001) better for real-time cine (4.5) than clinical standard cine (4.0). For both clinical standard and real-time cine, all individual visual score categories and SVS were above the clinically acceptable cut point 3.0 and 12.0, respectively.

Figure 3.

Representative images of three different CIED patients obtained with: (top row) clinical standard cine and (bottom row) real-time cine. Signal voids induced by CIED and/or intracardiac leads are noticeable in these cases. For dynamic display, see Video S1 and S2 in Supplementary Materials for clinical standard cine and real-time cine, respectively. As shown in videos, wall motion shows ghosting artifacts for patients 2 and 3 in clinical standard cine but not in real-time cine.

Table 2.

Median visual scores for clinical standard cine and real-time cine: conspicuity of endocardial wall at end diastole, temporal fidelity of wall motion, any visible artifact level on the heart, noise level throughout, and SVS. Values represent median (interquartile range). SVS: summed visual scores. *P<0.05 corresponds to significant difference.

Visual Score Category	Clinical standard	Real-time
Conspicuity	3.0 (2.0)	3.5 (1.1)
Temporal fidelity	4.0 (1.1)*	3.5 (1.1)*
Artifact level	3.0 (1.1)	3.0 (0.9)
Noise level	4.0 (0.5)*	4.5 (1.0)*
SVS	15 (2.8)	14.5 (3.6)

Weighted Cohen’s kappa values for the visual scores are summarized in Table 3. For clinical standard cine, there was moderate agreement for conspicuity score (kappa = 0.56), temporal fidelity score (kappa = 0.54), and artifact level score (kappa = 0.42); fair agreement for SVS (kappa = 0.33); none to slight agreement for noise score (kappa = 0.02). For real-time cine, there was fair agreement for temporal fidelity (kappa = 0.33) and noise (kappa = 0.38) scores; moderate agreement for conspicuity (kappa = 0.53), artifact level (kappa = 0.54), and SVS (kappa = 0.55).

Table 3.

Weighted Cohen’s kappa and confidence intervals (parenthesis) for clinical standard cine and real-time cine: conspicuity of endocardial wall at end diastole, temporal fidelity of wall motion, any visible artifact level on the heart, noise level throughout, and SVS.

Visual Score Category	Clinical standard	Real-time
Conspicuity	0.56 (−0.04, 1.16)	0.53 (−0.12, 1.18)
Temporal fidelity	0.54 (−0.23, 1.31)	0.33 (−0.6, 1.26)
Artifact level	0.42 (−0.33, 1.16)	0.54 (−0.1, 1.17)
Noise level	0.02 (−1.23, 1.27)	0.38 (−0.78, 1.55)
SVS	0.33 (−0.72, 1.39)	0.55 (−0.27, 1.37)

According to Wilcoxon signed-rank t-test, median EDV (151.2 [IQR=76.4] mL for real-time vs. 148 [IQR=56.8] mL for clinical standard) and ESV (83.2 [IQR=62] mL for real-time vs. 71.3 [IQR=46] mL for clinical standard) values were significantly (P <0.05 and P<0.01, respectively) different between two pulse sequences, whereas median SV (61.7 [IQR=18.2] mL for real-time vs. 61.2 [IQR=11.5] mL for clinical standard) and LVEF (40.6 [IQR=13.2] % for real-time vs. 43.4 [IQR=13.1] % for clinical standard) values were not significantly different. According to linear regression analyses on LV functional parameters, the association was strong for EDV (R² = 0.93), ESV (R² = 0.96), SV (R² = 0.66), and LVEF (R² = 0.89). Bland-Altman plots showing the agreement between pulse sequences for the four LV functional parameters are shown in Figure 4. For EDV, the mean difference was 10 mL (6.7% of mean), and the 95% limits of agreements (LOA) was 34 mL (22.7% of mean). For ESV, the mean difference was 8.2 mL (9.2% of mean), and the LOA was 26.7 mL (30% of mean). For SV, the mean difference was 0 mL (0% of mean), and the LOA was 13.8 mL (22.6% of mean). For LVEF, the mean difference was −2% (4.7% of mean), and the LOA was 5.8% (13.5% of mean).

Figure 4.

Bland-Altman plots showing agreement of left ventricular functional parameters (EDV, ESV, SV and LVEF) between clinical standard cine and real-time cine. Continuous and dotted lines indicate mean difference (real-time minus breath-hold) and 95% limits of agreements (mean difference ± 1.96 standard deviation), respectively.

Discussion

This study demonstrates the development of a 16-fold accelerated real-time, free-breathing cine pulse sequence with CS reconstruction for achieving clinically acceptable spatial resolution (2.1 mm × 2.1 mm) and temporal resolution (45 ms). The proposed pulse sequence was shown to produce clinically acceptable image quality (SVS = 14.5 for real-time vs. 15 for clinical standard), accurate LVEF (4.7% percent error compared with clinical standard), and reduced scan time (3.5 heartbeats per slice vs. 13 heartbeats per slice for clinical standard) in CIED patients. The reduction in scan time is even higher when factoring non-scanning time for clinical standard breath-hold cine (recovery in breathing + breathing instructions), which is typically longer than the actual breath-hold scan time. Another benefit of real-time cine over segmented breath-hold cine is reduction in RF energy deposition (i.e. less RF excitation per image acquisition), thereby providing a greater margin for MR safety.

ECG-gated, breath-hold cine CMR with b-SSFP readout (10) is the method of choice for assessing myocardial function in non-CIED patients (25). For CIED patients, however, b-SSFP produces considerable banding artifacts due to off-resonance, as shown in Figure 1. This preliminary observation led us to use a GRE pulse sequence for this study. While GRE pulse sequences using non-Cartesian k-space sampling would be preferable in terms of motion properties and incoherent aliasing artifacts, those pulse sequences produce considerable geometric distortion due to off-resonance (see Figure 1). It may be possible to deblur images obtained using non-Cartesian k-space sampling, but that will require static magnetic field mapping that may be challenging to acquire reliably in CIED patients. Another possibility is to use a GRE pulse sequence at 3 Tesla for increasing signal-to-noise ratio compared with 1.5 Tesla. As of today, three vendors (Medtronic, Minneapolis, Minnesota; Biotronik, Berlin, Germany; Abbott, St. Paul, Minnesota) offers FDA-approved ICDs on the market. Because there are not enough data on MR safety in patients with a non-MR-conditional CIED at 3 Tesla, it remains unclear whether CMR can be conducted safely in patients with a non-MR-conditional CIED at 3 Tesla. As safety issues get resolved, it may be worthwhile to explore the utility of real-time cine CMR at 3 Tesla in patients with a CIED.

There are a number of differences between clinical standard cine and real-time cine that warrant further discussion. One, clinical cine was performed before real-time cine due to higher priority given to clinical scans. On this note, clinical cine was performed immediately after administration of contrast agent, before clinical LGE, whereas real-time cine was performed after LGE, at which the concentration of gadolinium in the blood pool is expected to be considerably lower than immediate after injection. This difference in delayed imaging after contrast agent administration favors clinical cine in terms of image quality. Despite this bias towards clinical cine, the proposed real-time cine produced image quality scores that were clinically acceptable. Second, the receiver bandwidth was considerably lower for clinical cine (260 Hz/pixel) than real-time cine (745 Hz/pixel), favoring clinical cine in terms of signal-to-noise ratio. We elected to use a higher receiver bandwidth for real-time to achieve high temporal resolution. Third, TE was considerably shorter for real-time cine (1.8 ms) than clinical cine (5.2 ms), favoring real-time cine in terms of signal loss and geometric distortion due to device. We elected to use a short TE for real-time cine to minimize signal loss due to device. Fourth, clinical cine images were obtained with 6 mm slice thickness, whereas real-time cine images were obtained with 8 mm slice thickness. This difference in slice thickness may have contributed to differences in visual scores and LV functional parameters. Fifth, clinical cine was acquired with retrospective ECG gating during breath-holding, whereas real-time was acquired with prospective ECG gating during free-breathing. Differences in respiratory state (breath-holding vs. free-breathing) may have contributed to differences in LV functional parameters. Differences between retrospective ECG gating and prospective ECG triggering may have contributed to differences in LV functional parameters. We minimized this difference by scanning real-time cine over multiple heartbeats and choosing the most end-diastolic phase image, as previously described (11,12). Sixth, real-time cine images were obtained with nominal temporal resolution of 45.1 ms, whereas clinical cine images were obtained with 74.9 ms temporal resolution. Because clinical cine images were interpolated inline to 40 ms temporal resolution (25 frames), temporal fidelity scores were significantly higher for clinical cine that real-time cine. This finding needs to be interpreted with caution, since the temporal resolution of clinical cine was artificially inflated through temporal interpolation. Seventh, we used a standard coil array equipped with our 1.5 Tesla scanner. It may be possible to further accelerate clinical standard cine by approximately 50% (from acceleration factor 2 to 3) using a high-end (e.g. 32-channel) cardiac coil array, which may also benefit compressed sensing by boosting signal-to-noise ratio. A future study using a high-end cardiac coil array is warranted to investigate whether the trends reported in this study are valid with a high number of coils.

This study has several limitations that warrant further discussion. First, imaging parameters were different between clinical cine and real-time cine. These discrepancies may have contributed to differences in visual scores and LV functional parameters. Second, delayed imaging time post contrast agent was different between two scans. This was unavoidable since clinical scans have higher priority over research add-on scans. Third, arrhythmia burden of our patient cohort was modest (CV of R-R interval = 4.6%). In patients with even higher burden of arrhythmia, the proposed real-time cine CMR may produce greater improvement in image quality compared with clinical breath-hold cine. Fourth, this study included only 13 patients, because our clinical volume for CMR in CIED patients is relatively low compared to CMR in non-device patients. A future study including more patients with a variety of conditions and device types is warranted to fully evaluate the clinical utility of the proposed real-time cine CMR.

In conclusion, this study demonstrates feasibility of 16-fold accelerated real-time free-breathing cine CMR with CS reconstruction for producing clinically acceptable image quality, relatively accurate LVEF measurement, and reduced scan time in CIED patients. Further study includes rigorous testing in a large patient cohort with a variety of heart diseases and CIED types.

Supplementary Material

1

Supporting Information Video S1: Dynamic display of clinical standard cine CMR image sets corresponding to Figure 3 (upper row).

2

Supporting Information Video S2: Dynamic display of real-time cine CMR image sets corresponding to Figure 3 (lower row).