Compressed Sensing Real-Time Cine Reduces CMR Arrhythmia-Related Artifacts

Background and objective: Cardiac magnetic resonance (CMR) is a key tool for cardiac work-up. However, arrhythmia can be responsible for arrhythmia-related artifacts (ARA) and increased scan time using segmented sequences. The aim of this study is to evaluate the effect of cardiac arrhythmia on image quality in a comparison of a compressed sensing real-time (CSrt) cine sequence with the reference prospectively gated segmented balanced steady-state free precession (Cineref) technique regarding ARA. Methods: A total of 71 consecutive adult patients (41 males; mean age = 59.5 ± 20.1 years (95% CI: 54.7–64.2 years)) referred for CMR examination with concomitant irregular heart rate (defined by an RR interval coefficient of variation >10%) during scanning were prospectively enrolled. For each patient, two cine sequences were systematically acquired: first, the reference prospectively triggered multi-breath-hold Cineref sequence including a short-axis stack, one four-chamber slice, and a couple of two-chamber slices; second, an additional single breath-hold CSrt sequence providing the same slices as the reference technique. Two radiologists independently assessed ARA and image quality (overall, acquisition, and edge sharpness) for both techniques. Results: The mean heart rate was 71.8 ± 19.0 (SD) beat per minute (bpm) (95% CI: 67.4–76.3 bpm) and its coefficient of variation was 25.0 ± 9.4 (SD) % (95% CI: 22.8–27.2%). Acquisition was significantly faster with CSrt than with Cineref (Cineref: 556.7 ± 145.4 (SD) s (95% CI: 496.7–616.7 s); CSrt: 23.9 ± 7.9 (SD) s (95% CI: 20.6–27.1 s); p < 0.0001). A total of 599 pairs of cine slices were evaluated (median: 8 (range: 6–14) slices per patient). The mean proportion of ARA-impaired slices per patient was 85.9 ± 22.7 (SD) % using Cineref, but this was figure was zero using CSrt (p < 0.0001). The European CMR registry artifact score was lower with CSrt (median: 1 (range: 0–5)) than with Cineref (median: 3 (range: 0–3); p < 0.0001). Subjective image quality was higher in CSrt than in Cineref (median: 3 (range: 1–3) versus 2 (range: 1–4), respectively; p < 0.0001). In line, edge sharpness was higher on CSrt cine than on Cineref images (0.054 ± 0.016 pixel−1 (95% CI: 0.050–0.057 pixel−1) versus 0.042 ± 0.022 pixel−1 (95% CI: 0.037–0.047 pixel−1), respectively; p = 0.0001). Conclusion: Compressed sensing real-time cine drastically reduces arrhythmia-related artifacts and thus improves cine image quality in patients with arrhythmia.


Introduction
Cardiac magnetic resonance (CMR) is a major imaging modality for the assessment of left and right ventricular volumes and mass [1][2][3]. Moreover, it provides effective morphologic and kinetic assessment, including of the right ventricle which is not easily evaluated with ultrasounds due to its retrosternal location [4]. Multi-breath-hold segmented balanced steady state free precession (bSSFP) sequences are considered superior to gradient-echo imaging since they provide better endocardium delineation and reproducibility in a shorter scan time [5]. Retrospective electrocardiogram (ECG) gating requires the heart rate (HR) to be a regular periodic phenomenon as pieces of data are continuously acquired on multiple cardiac cycles, time-labelled and merged for the reconstruction of a whole cine slice, which is a weighted representation of successive heartbeats. It allows adapting the length of the acquisition window to the duration of the heartbeat during the continuous acquisition. This enables capturing of the complete cardiac cycle in segmented acquisitions. Typically, k-space interpolation or filtering is applied to retrospectively gate the acquired data to a reference heartbeat [6,7]. In the case of arrhythmia, artifacts occur since reconstruction is performed using incoming data from different frames of the cardiac cycle. Arrhythmia rejection algorithms can be applied with retrospective gating but may end in exceedingly long breath-holds. These arrhythmia-related artifacts (ARA) may be limited using prospectively triggered sequences by setting the acquisition window shorter than the briefest measured RR interval (time laps between two consecutive R peaks) [8]. However, this requires decreasing the number of k-space lines acquired per cardiac frame in order to preserve the widely accepted temporal resolution of 20 phases per cardiac cycle and misses to display the diastolic phases [9,10]. As a result of these adjustments, longer breath-holds and scan time are observed while the last phases of the cardiac cycle are not sampled.
Decreasing the amount of measured data is a simple way to reduce acquisition time. In recent years, compressed sensing was established as a powerful method to drastically reduce scan time [11][12][13][14]. This is achieved by highly undersampling k-space with a random sampling pattern. After Fourier transform, these acquired data result in noise-like, incoherent artifacts. These artifacts are compensated for in the final image with a non-linear iterative reconstruction exploiting the fact that medical images have a sparse representation. In combination with parallel imaging, acceleration rates can be achieved with CS that enable real-time cardiac cine imaging based on a balanced bSSFP readout with spatiotemporal resolution in a similar range to the reference (Cine ref ) acquisitions [15].
Various CMR studies have evaluated real-time CS cine sequences in 1.5 and 3 Tesla magnetic resonance scanners showing promising results for the assessment of left and right ventricles, including in patients with atrial fibrillation [16][17][18][19][20][21]. However, image quality was not specifically assessed in patients with irregular HR. Based on the assumption that real-time CS cine (CS rt ) could reduce ARA, our study aimed at evaluating its image quality as compared to the reference multi-breath-hold segmented bSSFP cine (Cine ref ) in patients suffering from cardiac arrhythmia.

Study Population
From January 2019 to December 2019, 71 adult patients referred to our cardiovascular radiology department for CMR with concomitant arrhythmia during scanning were enrolled. Irregular HR was defined when the coefficient of variation of RR intervals (CV RR ) was greater than 10% while scanning. The CV RR was calculated as the ratio of the standard deviation to the mean of RR intervals' durations which were obtained from digital imaging and communications in medicine (DICOM) fields. Patients under 18 years old, grown-up congenital heart disease, stress CMR, patients undergoing ECG retrogated CMR and patients with sinus rhythm were excluded. A graphic illustration of the study design is provided in Figure S1 (Supplementary Materials). The protocol was approved by our institutional ethics committee and patients gave informed consent. The study was approved by the French National Agency for the Safety of Medicines and Health Devices (ANSM; ID-RCB: 2017-A00852-51).

Imaging Protocol
CMR studies were performed on a 1.5 T scanner (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany). Every patient underwent two series of cine images: first, the reference prospectively triggered and segmented multi-breath-hold Cine ref sequence; second, the prototype single-breath-hold real-time single-shot CS rt cine sequence. Both acquisitions included one left ventricular (LV) and one right ventricular (RV) two-chamber slice, one four-chamber slice and a LV short-axis stack covering both ventricles with an 8 mm slice thickness and a 2 mm gap. Regarding the prospectively gated Cine ref sequence, 20 phases of the cardiac cycle were acquired and the number of views per frame was set to reach this sampling rate. In single-shot CS rt cine imaging, the data acquisition was performed in a single heartbeat. The acquisition was triggered by the R peak on the ECG. With adaptive triggering, the acquisition was stopped with the next R peak, which allowed capturing the complete cardiac cycle. However, for multi-slice acquisition it could lead to a variation in the number of cardiac phases acquired between different slices as the temporal resolution was fixed. Temporal interpolation was applied to generate an additional dataset with a fixed number of cardiac phases (n = 20). This dataset was used to quantify cardiac function using a dedicated post-processing software that required a fixed number of cardiac phases. To evaluate the CS rt sequence in clinical conditions, 40 iterations were used to perform image recovery to maintain an acceptable reconstruction time. An additional phase contrast imaging (PCI) flow sequence was acquired on the aortic root. Segmented Cine ref and CS rt cine sequences parameters are available in Table 1. Data are expressed as mean ± standard deviation in the absence of any indication. a Constant value. b Interpolation was performed to provide a constant frame rate of 20 cardiac phases per cycle for post-processing. c The number of views per frame was set according to the shorter RR interval in order to acquire 20 cardiac phases. Prospective triggering allows data sampling during a fixed acquisition window after each R peak while adaptative triggering allow data sampling until the next R peak occurs. Abbreviations: Cine ref , reference segmented cine; CS rt , real-time compressed sensing cine; ECG, electrocardiogram; n, data represented as numbers; NA, not applicable.

Cine Images Quality Assessment
Image quality was evaluated in both groups using four indicators. First, the subjective overall image quality was evaluated using a subjective 4-point Likert scale (1: non diagnostic; 2: poor; 3: good; 4: excellent). Secondly, an objective image quality assessment was carried out based on standardized criteria adapted from the European CMR registry "LV-Function cine SSFP" section (referred to below as "EuroCMR score") [22] (p. 3). Higher scores referred to more frequent artifact occurrence (Table 2).

Modified score (items 1 to 8) 10
Every acquisition using both sequences marked a null score concerning the four last items. Indeed, acquisitions were repeated every time slice orientation was not appropriated (item 9 = 0); all acquisitions (Cine ref and CS rt ) were performed using the same slice thickness and gap which were 8 mm (item 10 = 0) and 2 mm (item 11 = 0), respectively, and both horizontal and vertical long-axis slices were systematically acquired (item 12 = 0). Criteria in italics were not applied, and only bold criteria were used for objective quality assessment in our study, providing a maximum score of 10 points. The more artifacts there were, the higher the score was. Abbreviations: LV, left ventricle; SSFP, steady-state free precession; CMR, cardiac magnetic resonance; Cine ref , reference segmented cine; CS rt , real-time compressed sensing cine; ARA, arrhythmia-related artifacts.
Third, the proportion of short-axis slices affected by ARA in each stack of both sequences was calculated, referred to as ARA rate. ARA were defined as a blurring of all or a part of the LV wall borders [22].
Finally, the edge sharpness (ε) of the boundary between myocardium and blood pool, which is the spatial frequency (in pixel −1 ) reflecting the spatial resolution, was measured on paired Cine ref and CS rt four-chamber slices at end-diastole, accordingly to the literature [23,24]. Additional measurement at end-systole was performed. The edge spread function (ESF), which is the response of the imaging system to a high contrast boundary, was measured on MATLAB (version R2015a, The MathWorks, Natick, MA, USA), by drawing a signal profile line perpendicularly across the edge between the interventricular septum and the LV blood pool (Figure 1a,b) [25]. Then, ε was calculated as the reciprocal of the distance separating the points corresponding to 20% and 80% of the difference between local minimum and maximum signal intensities (Figure 1c,d). Figure 1. Example of edge sharpness assessment at end-diastole for a 56-year-old male patient suffering from atrial fibrillation. The same four-chamber view at end-diastole is acquired with (a) the Cineref sequence and (b) the CSrt sequence. An orthogonal profile line was drawn at mid-cavity across the border between the septal myocardium and the left ventricular blood pool (blue line) on a four-chamber view. It provided intensity profiles (blue curves) along the line for (c) Cineref and (d) CSrt cine. The edge sharpness was the inverse of the distance d (in pixels) between the positions corresponding to 20% and 80% (red stars) of the difference between the maximum and minimum signal intensities (blue crosses). The edge sharpness was expressed in pixel −1 . This measurement was performed at end-diastole and end-systole for both sequences. Note that the peaks (arrows heads) added to the Cineref signal profile curve (c) correspond to the doubling of the interventricular septum border (arrows) on the cine view (a). The same assessment was performed on both sequences at enddiastole and end-systole for the 71 enrolled patients. Abbreviations: Cineref, reference segmented cine; CSrt, real-time compressed sensing cine; εCineref, edge sharpness measured on Cineref sequence; εCSrt, edge sharpness measured on CSrt cine; I, signal intensity; Imin, minimal signal intensity; Imax, maximal signal intensity; d, distance along the profile line.

Conditions of Image Analysis
Images from both sequences were anonymized before transfer to a clinical workstation (Sygno.via VB30A, Siemens Healthcare, Erlangen, Germany). A radiologist with 4 years of experience (PEA) performed the image quality assessment according to the above-cited indicators. Image sets were randomly evaluated in each group. The same observer (PEA) first performed the quality assessment of the reference Cineref images and at least one month later evaluated the CSrt images. For each patient, arrhythmia was quantified by calculating the CVRR. An additional assessment was performed by a radiologist with 8 years of experience (BL) from 30 randomly selected patients to evaluate the interobserver agreement and performed the same assessment regarding subjective quality, Eu- An orthogonal profile line was drawn at mid-cavity across the border between the septal myocardium and the left ventricular blood pool (blue line) on a four-chamber view. It provided intensity profiles (blue curves) along the line for (c) Cine ref and (d) CS rt cine. The edge sharpness was the inverse of the distance d (in pixels) between the positions corresponding to 20% and 80% (red stars) of the difference between the maximum and minimum signal intensities (blue crosses). The edge sharpness was expressed in pixel −1 . This measurement was performed at end-diastole and end-systole for both sequences. Note that the peaks (arrows heads) added to the Cine ref signal profile curve (c) correspond to the doubling of the interventricular septum border (arrows) on the cine view (a). The same assessment was performed on both sequences at end-diastole and end-systole for the 71 enrolled patients. Abbreviations: Cine ref , reference segmented cine; CS rt , real-time compressed sensing cine; ε Cineref , edge sharpness measured on Cine ref sequence; ε CSrt , edge sharpness measured on CS rt cine; I, signal intensity; I min , minimal signal intensity; I max , maximal signal intensity; d, distance along the profile line.

Conditions of Image Analysis
Images from both sequences were anonymized before transfer to a clinical workstation (Sygno.via VB30A, Siemens Healthcare, Erlangen, Germany). A radiologist with 4 years of experience (PEA) performed the image quality assessment according to the above-cited indicators. Image sets were randomly evaluated in each group. The same observer (PEA) first performed the quality assessment of the reference Cine ref images and at least one month later evaluated the CS rt images. For each patient, arrhythmia was quantified by calculating the CV RR . An additional assessment was performed by a radiologist with 8 years of experience (BL) from 30 randomly selected patients to evaluate the interobserver agreement and performed the same assessment regarding subjective quality, EuroCMR score and ARA rates. In the case of mismatch between the two readers, a radiologist with 15 years of experience (FP) performed the quality assessment with the two others to reach consensual scores which were used instead of those set by the first and less experienced observer. Mismatches were defined by discrepancies greater than or equal to 2 points regarding subjective quality score and EuroCMR score, or by a 20% difference in ARA rates. The edge sharpness assessment was automated and was not evaluated for interobserver agreement. Finally, semi-automated segmentation of LV endocardium and epicardium, and manual segmentation RV endocardium were performed on the same workstation with both cine sequences for each patient. LV stroke volume was also measured on PCI sequence.

Statistics Analysis
Categorical data were represented as numbers (percentages), continuous variables as mean ± standard deviation (SD) (95% confidence interval (CI)) in case of normal distribution and median (range: minimum-maximum) in other cases. Sequences were compared using the Wilcoxon signed-rank test regarding the overall subjective quality score and the modified EuroCMR score. Paired Student's t-test was used for ARA rates, edge sharpness comparisons, and ventricular functional parameters comparison. An analysis of variance (ANOVA) was used to compare LV stroke volumes assessed by cine segmentation and PCI flow sequence. Intraclass correlation coefficient and kappa test were applied to assess the interobserver agreement [26]. Values of p < 0.05 were considered statistically significant. Statistical analysis was performed using MedCalc software (version 14.8.1.0, MedCalc Software, Ostend, Belgium).

Subjective Overall Quality Score
The subjective quality score was significantly better (p < 0.0001) for the CSrt sequence with a median score of 3 (range : 1-3). A 0.85 interobserver agreement was reached. The Cineref sequence provided a median score of 2 (range: 1-4; p < 0.0001) with an intraclass coefficient of 0.82 (Table 6). No mismatch was encountered between the readers. The Ci-

Subjective Overall Quality Score
The subjective quality score was significantly better (p < 0.0001) for the CS rt sequence with a median score of 3 (range: 1-3). A 0.85 interobserver agreement was reached. The Cine ref sequence provided a median score of 2 (range: 1-4; p < 0.0001) with an intraclass coefficient of 0.82 (Table 6). No mismatch was encountered between the readers. The Cine ref sequence provided 23 non-diagnostic acquisitions compromising functional and morphological assessments versus only 10 stacks with the CS rt cine (Figure 2; Video S1 (Supplementary Materials)).

Discussion
This prospective monocentric study based on a 71-patient cohort is, to our knowledge, the widest and most comprehensive study to evaluate the CS rt sequence in patients with irregular HR. Previous studies in non-selected patients confirmed that real-time CS cine imaging is a reliable alternative to segmented multi-breath-hold SSFP for the assessment of both ventricles' volumes and function in addition to reducing acquisition time [16][17][18][19][20]. Our study demonstrates a dramatic drop in ARA and a significant improvement of subjective and objective image quality with CS rt in patients suffering from heart rhythm disorders. However, no CS rt set was rated as excellent because of the smooth boundaries rendered by the interpolation process which are mandatory for post-processing. Indeed, since the temporal resolution of the CS sequence is fixed, a variable number of frames will be acquired from one cycle to another in the case of arrhythmia. For segmentation to be achieved, post-processing tools require all cine slices to display the same number of frames per cycle. Consequently, a standardization is performed to display 20 frames per cycle on all slices.
We previously showed on a small sub-group of 25 patients suffering from arrhythmia that CS rt and Cine ref sequences allowed similar image quality [20]. However, the present study does not only suggest equivalent scores but significantly better objective and subjective image quality scores with the CS rt cine. Of note, this sequence still provided non-null EuroCMR scores since the slices were identically located on both sequences; accordingly, most of the wrap-around or metallic artifacts were reproduced on the CS rt acquisition.
Our results are in line with the previous study by Goebel et al. on 20 patients with atrial fibrillation [21]. This study focused on a subjective semi-quantitative 4-point quality score and the evaluation of the variation of the myocardial signal intensity which is the reciprocal of the signal-to-noise ratio (SNR). However, this last parameter, or its reciprocal, is considered as hardly suitable for non-linear iterative reconstructions. Moreover, CS is built to suppress pieces of the image signal, while SNR is suited for fully sampled data [27]. Besides our study being specifically designed to evaluate the image quality using additional quantitative and objective metrics, we also performed a clinically integrated evaluation in a larger population.
The higher edge sharpness of CS rt images reflected the faster signal variation along a distance and a better delineation of the image boundaries. We evaluated the edge sharpness both at end-diastole, when the myocardium is supposed to be relatively still, and at end-systole. This metric, regardless of the non-linearity of the reconstruction process, is more suitable than the SNR or its reciprocal to evaluate image quality [27]. The ESF and its inverse value ε measure the imaging system ability to restitute high contrast transition in images. This parameter, the derivative (the line spread function) and its Fourier transform (the so-called task-based modulation transfer function or task transfer function) are currently considered for image quality assessment in the field of non-linear image reconstructions [24,25,28].
Regarding volumetric evaluation in patients suffering from arrhythmia, CS rt provided significantly lower LV end-diastolic volume (−3.6 ± 7.2 mL) than measured on Cine ref , which was already observed in previous studies [16,20]. As for RV, there was a significant underestimation of all evaluated functional parameters. Nevertheless, these variations compared to the conventional Cine ref should be considered with caution in a population with arrhythmia. Indeed, irregular heartbeats induce variable ventricular preloads and contractions, making the real reference values impossible to determine. Moreover, since CS rt demonstrated a better image quality than Cine ref that was impaired by ARA, one may consider the segmentation to be more reliable on CS rt .
Besides the image quality improvement, the single breath-hold CS rt sequence allowed a dramatic reduction in scan time. Not only was acquisition faster, but the ARA reduction avoided repeating the acquisition of non-diagnostic slices [20]. The workflow improvement being a major issue in the field of CMR, this real-time sequence is very promising and may improve cost-effectiveness [29].

Limitations
Although the overall subjective image quality was improved with CS rt cine, 10 stacks were still considered as non-diagnostic. Indeed, iterative reconstructions occasionally failed or were not completely achieved on this prototype sequence. Nevertheless, such failures are now rare since the release of the final version of the sequence. Moreover, ECG-related issues occurred when R peaks were occasionally missed, which made the system consider two consecutive heartbeats as one single cycle. The corresponding slices then display a double heart cycle which could not be used for post-processing and was ranked as non-diagnostic. Special attention should be paid to skin preparation before ECG electrode placement.
Other fast real-time sequences, such as radial acquisition, have previously been reported [30,31]. Our study does not compare CS rt to other types of real-time sequences. To our knowledge, no such evaluation has been published and further study would be required for comparison.
A methodological limitation of our study is the impossibility to perform blinded evaluation of the sequences since CS rt cine displayed smoother boundaries than Cine ref sequence. Consequently, observers could recognize the type of sequence they were evaluating. However, paired CS rt and Cine ref stacks from the same patients were separated, randomized, and assessed during different sessions.
Regarding the sampling of heart cycles for the assessment of ventricular volumes and mass, the fixed temporal resolution leads to variable sampling rates from shorter cycler to longer ones. As a consequence, in the case of HR faster than 60 bpm, the recommended 20 frames per cycle could not be acquired [9,10]. However, in the field of analog-to-digital signal conversion, a 16-time oversampling is reputed sufficient for the signal restitution to be accurate, corresponding to a 75 bpm HR [32]. High CV RR in HR may be encountered, and some slices may display undersampled heart cycles. In our study, 38% of the patients demonstrated a mean HR above 75 bpm for whom the undersampling of the cardiac cycle should be considered cautiously during interpretation, especially for volume segmentations. Nevertheless, it must be balanced by the reduction of ARA provided by CS rt .

Conclusions
In addition to reducing acquisition time, CS rt sequence drastically reduces arrhythmiarelated artifacts and improves image quality in patients with irregular heart rate. This rapid imaging technique allows practitioners, in daily practice, to improve quality, workflow and accessibility of CMR for patients with challenging cardiac conditions. Supplementary Materials: The following are available online at https://www.mdpi.com/article/10 .3390/jcm10153274/s1, Figure S1: Study design. Video S1: Viability assessment in a 77-year-old man with premature ventricular contractions caused by ischemic scars.
Author Contributions: B.L.: study conception and design, data collection, interpretation and analysis, drafting of the manuscript, critical revision for important intellectual content; P.-E.A.: data collection, interpretation and analysis, drafting of the manuscript; C.V.G.: data collection and interpretation, critical revision for important intellectual content; A.C.: study conception, critical revision for important intellectual content; J.H.: data collection and interpretation, critical revision for important intellectual content; A.S.: data collection and interpretation, critical revision for important intellectual content; M.S.: study conception and design, critical revision for important intellectual content; C.F.: study conception and design, critical revision for important intellectual content; S.T.: study conception and design, critical revision for important intellectual content; D.M.: study conception, critical revision for important intellectual content; F.P.: study conception and design, data collection, interpretation and analysis, drafting of the manuscript, critical revision for important intellectual content. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.

Institutional Review Board Statement:
The study was approved by the research ethics committee of Lille University Hospital.