Reliability of respiratory-gated real-time two-dimensional cine incorporating deep learning reconstruction for the assessment of ventricular function in an adult population

This study aimed to assess the image quality and accuracy of respiratory-gated real-time two-dimensional (2D) cine incorporating deep learning reconstruction (DLR) for the quantification of biventricular volumes and function compared with those of the standard reference, that is, breath-hold 2D balanced steady-state free precession (bSSFP) cine, in an adult population. Twenty-four patients (15 men, mean age 50.7 ± 16.5 years) underwent cardiac magnetic resonance for clinical indications, and 2D DLR and bSSFP cine were acquired on the short-axis view. The image quality scores were based on three main criteria: blood-to-myocardial contrast, endocardial edge delineation, and presence of motion artifacts throughout the cardiac cycle. Biventricular end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV), ejection fraction (EF), and left ventricular mass (LVM) were analyzed. The 2D DLR cine had significantly shorter scan time than bSSFP (41.0 ± 11.3 s vs. 327.6 ± 65.8 s; p < 0.0001). Despite an analysis of endocardial edge definition and motion artifacts showed significant impairment using DLR cine compared with bSSFP (p < 0.01), the two sequences demonstrated no significant difference in terms of biventricular EDV, ESV, SV, and EF (p > 0.05). Moreover, the linear regression yielded good agreement between the two techniques (r ≥ 0.76). However, the LVM was underestimated for DLR cine (109.8 ± 34.6 g) compared with that for bSSFP (116.2 ± 40.2 g; p = 0.0291). Respiratory-gated 2D DLR cine is a reliable technique that could be used in the evaluation of biventricular volumes and function in an adult population.


Introduction
Cardiac magnetic resonance (CMR) is considered the reference standard for the noninvasive assessment of cardiac morphology, biventricular volumes, and systolic function [1]. Most commonly, biventricular volumes and ejection fraction (EF) are calculated by acquiring multiple and contiguous two-dimensional (2D) balanced steady-state free precession (bSSFP) slices during breath-holds through the ventricles on short -axis views [2]. bSSFP provides intrinsic advantages, especially its superior blood-myocardium contrast owing to blood in-flow effects. However, 2D breath-hold bSSFP cine sequences requires the suspension of respiration, and several factors may limit the ability of patients to perform consistent breath-holds, such as anxiety, medical conditions, and age. Moreover, the inconsistent breath-hold position during the multiple breath-hold acquisition may cause misalignment of consecutive slices [3,4].
Free-breathing bSSFP can be partially overcome using multiple signal averaging or respiratory gating, but usually at the expense of degraded image quality or increased.
Recently, a deep learning reconstruction (DLR) was proposed to estimate high-spatial-resolution cardiac MRI bSSFP images from low-resolution short-axis scans with reduced matrix size, although no increased acceleration factors [9]. DLR cine imaging uses a highly accelerated 2D bSSFP sequence with model-based DLR to iteratively reconstruct cardiac cine images directly from sparsely sampled k-space data [10]. It has been evaluated in pediatric patients; however, no data are available concerning the reliability in an adult population [11].
This study aimed to assess the image quality and accuracy of respiratory-gated real-time 2D cine incorporating DLR for the quantification of biventricular volumes and function in comparison with 2D bSSFP cine of an adult population in the clinical setting.

Study population and design
Between May 2021 and November 2021, 24 patients (15 [63%] men, mean age 50.7 ± 16.5 years) with clinical indications at Iwate Medical University Hospital were prospectively recruited. Patients with supraventricular arrhythmias, non-MR compatible devices, or inability to sustain a breath-hold were excluded from the study. Patients with the following clinical indication were enrolled: hypertrophic cardiomyopathy (n = 7), cardiac sarcoidosis (n = 5), suspected or known coronary artery disease (n = 4), dilated cardiomyopathy (n = 4), myocarditis (n = 3), and hypertensive heart disease (n = 1). The study protocol conforms to the ethical guidelines of the 1975 Declaration of Helsinki as reflected in an a priori approval by the institution's human research committee (MH2018-561), and written informed consent for data collection was obtained from each patient before CMR.

CMR acquisition
All examinations were performed using 1.5 T (SIGNA Artist, GE Healthcare, Waukesha, USA) with a 30-channel phased array receiver coil. We initially recorded localizing scans for determining ventricular long-axis orientation. Short-axis stack cine images were acquired, in random order, in each patient using two different imaging techniques: the conventional 2D bSSFP and 2D DLR cine imaging. The most basal short-axis slice was located immediately on the myocardial side of the atrioventricular junction at the end-diastole prescribed from the previously acquired localizing scans. The 2D bSSFP cine was performed with multiple breath-holds with the following parameters: FOV = 360 × 360 mm, repetition time = 3.7 ms, echo time = 1.0 ms, flip angle = 60°, image matrix = 224 × 160, bandwidth = 892.9 Hz/pixel, slice thickness = 8 mm with a 2-mm gap, voxel size = 1.6 × 2.2 mm, and an acceleration factor of 2D using ASSET parallel imaging. 2D DLR is GE research prototype including a time-varied variable density sampling cine acquisition and a model based DLR algorithm. The 2D DLR cine was performed with free-breathing using the following parameters: FOV = 360 × 360 mm, repetition time = 2.9 ms, echo time = 1.0 ms, flip angle = 60°, image matrix = 160 × 150, bandwidth = 1562.5 Hz/pixel, slice thickness = 8 mm with a 2-mm gap, voxel size = 2.2 × 2.4 mm, and an acceleration factor of 6 using variable density sampling scheme. To reduce the breathing effects on the image quality and volume quantification, the 2D DLR cine sequence was respiratory-gated based on the signal with only approximately 40% data acquisition window near end-expiration accepted for image reconstruction, and all data were acquired within a single heart beat for each slice (real-time acquisition). For 2D bSSFP cine, a views/segment of 12 according the heart rate was used, resulting in a temporal resolution of 47 ms. The 2D DLR cine sequence automatically adapts views/segment based on the patient heart rate at scan time in a temporal resolution of 175 ms. Both sequences were acquired with retrospective ECG vector gating, and the acquired cardiac phases were linearly interpolated to 20 phases in 2D bSSFP cine and 30 phases in 2D DLR cine in a cardiac cycle. The 2D DLR images were reconstructed offline in about 15 min.

Image analysis
Datasets of the cine images were transferred to an offline workstation, processed using commercially available software (Ziostation2, Ziosoft Inc., Tokyo, Japan), and evaluated by two expert readers with 12 and 5 years of experience in CMR performance and analysis, respectively. To estimate interobserver variability, each operator performed a blinded analysis on randomly rearranged sequences of all patients, and one of the two observers repeated the analysis on all patients after at least 1 month.
The image quality scores were based on three main criteria: blood-to-myocardial contrast, endocardial edge delineation, and presence of motion artifacts throughout the cardiac cycle. Details of the image quality criteria are described elsewhere [8]. Each criterion was graded on a scale of 1-5, where 1 was nondiagnostic; 2, suboptimal but still diagnostic for volumetric analysis; 3, adequate; 4, good; and 5, excellent.
The endocardial and epicardial contours were automatically traced at end-systole and end-diastole with manual tracing when required. The most basal section was defined as the section in which at least 50% of the border of the lumen was surrounded by the myocardium. The papillary muscles and trabeculations of the RV were included as part of the blood pool [12]. The RV outflow tract was included as part of the RV blood volume. Subsequently, left ventricular end-diastolic volume (LVEDV), LV end-systolic volume (LVESV), LV stroke volume (LVSV), LVEF, RVEDV, RVESV, RVSV, and RVEF were calculated by automatically tracing the endocardial contours on short-axis stacks in end-diastole and end-systole for both datasets of sequences according to the Society for Cardiovascular Magnetic Resonance guidelines [13]. The LV mass (LVM) was calculated as the difference between the total epicardial volume and the total endocardial volume. The actual time required for imaging acquisition was recorded for both 2D bSSFP and 2D DLR cine.

Statistical analysis
All statistical analyses were performed using JMP® 13 (SAS Institute Inc., Cary, NC, USA). Continuous variables are presented as mean ± standard deviation (SD) or median and interquartile range as appropriate, whereas qualitative variables are presented as numbers and percentages. Normality was checked using the Shapiro-Wilk test. Differences between means were evaluated using the paired and unpaired (for independent group comparisons) Student t-tests for normally distributed data and the Mann-Whitney or Wilcoxon signed-rank tests for nonparametric data. The correlation between the parameters calculated from the 2D bSSFP and 2D DLR cine images were analyzed using Pearson's correlation analysis. The Bland-Altman analysis was performed to assess the interchangeability between the 2D bSSFP and 2D DLR cine image parameters, bias, and 95% limit of agreement, which was calculated by multiplying the SD by ± 1.96. The inter-and intra-observer reproducibilities were assessed using intraclass correlation coefficients (ICC) for absolute agreement of single measures with a 95% confidence interval (CI). Two-tailed p < 0.05 was considered significant.

Population
The baseline characteristics of the study population are presented in Table 1. 2D DLR cine sequences were successfully performed and evaluated in all patients.

Inter-technique analysis
The actual time of acquisition was significantly shorter in 2D DLR than that in 2D bSSFP (41.0 ± 11.3 s vs. 327.6 ± 65.8 s, respectively; p < 0.0001) ( Table 2). An analysis of image quality showed significant impairment in 2D DLR compared with 2D bSSFP with inferior results for endocardial edge definition and motion artifact ( Table 2). On the contrary, no

Discussion
To the best of our knowledge, this study is the first to perform clinical validation of respiratory-gated 2D DLR cine for volumetry assessment in a cohort of adult patients. The 2D DLR cine has similar reproducibility to conventional 2D bSSFP sequence, despite some drawbacks in image quality. Notably, the time of acquisition for 2D DLR cine was significantly lower than that for 2D bSSFP. 2D DLR sequences allow an accurate evaluation of biventricular volumes and EF comparable with 2D bSSFP sequences. However, the LVM was underestimated for 2D DLR cine compared with those for 2D bSSFP cine images.

Limitations of breath-hold two-dimensional balanced steady-state free precession
2D bSSFP cine is the most common approach for the evaluation of volumes and systolic function in clinical practice. bSSFP provides intrinsic advantages, especially its superior blood-to-myocardium contrast owing to blood in-flow effects. However, the conventional 2D bSSFP cine sequence requires the suspension of respiration, and several factors may limit the ability of patients to perform consistent breathholds, such as anxiety, medical conditions, and age [3,4].
To overcome the technical limitations of 2D bSSFP cine described above, several free-breathing cine sequences have been adopted to achieve faster CMR without affecting the temporal and spatial resolution of sequences [3,[14][15][16][17]. We reported previously that respiratory-triggered 2D cine was a reliable technique for the evaluation of biventricular volumes and function in patients with repaired tetralogy of Fallot [7]. However, the RVESV was overestimated and the RVSV and RVEF were underestimated using respiratorytriggered 2D cine compared with using 2D bSSFP cine because blood-myocardial boundaries were more blurred in respiratory-triggered 2D cine.
significant differences in blood-to-myocardial contrast were found between 2D DLR and 2D bSSFP cine. In addition, the overall image quality scores of both sequences for all acquisitions were ≥ 3, indicating at least diagnostic quality.
The interobserver variability showed strong correlations for both sequences with ICC ≥ 0.96 for all measurements of the LV (Table 3 A) and RV (Table 3B). The intra-observer differences were also strongly correlated (ICC ≥ 0.98).

Correlations of the LV and RV functional parameters
The Bland-Altman plots showed strong agreement between the two sequences ( Figs. 1 and 2), and the linear regression yielded good agreement between the two techniques (r ≥ 0.75) ( Table 4).
Compared with 2D bSSFP, 2D DLR demonstrated no significant difference in terms of biventricular EDV, ESV, SV, and EF. However, the LVM for 2D DLR was underestimated compared with that for 2D SSFP (2D DLR = 109.8 ± 34.6 g vs. 2D SSFP = 116.2 ± 40.2 g, respectively; p = 0.0291) ( Table 5). Representative cases of 2D bSSFP and 2D DLR imaging in patient with dilated cardiomyopathy, cardiac sarcoidosis, hypertrophic cardiomyopathy, and chronic myocarditis are shown in Figs. 3, 4, 5 and 6. observed with the DLR cine, likely related to undersampling and image reconstruction process and contributed to slightly lower image quality scores when compared with bSSFP. Additionally, both the low spatial and temporal resolution of the 2D DLR cine should have affected the image quality and volumetric analysis. However, no significant differences were found in the blood-to-myocardial contrast between 2D DLR and 2D bSSFP sequences. Moreover, compared with 2D bSSFP sequences, 2D DLR sequences allow an accurate evaluation of biventricular volumes and EF including RV parameters. Our study results showed good reproducibility of biventricular volumes and EF between 2D bSSFP and 2D DLR sequences. These results were better than that reported in similar studies comparing bSSFP to nonconventional short-axis techniques [7,8,19,20]. On the contrary, the LVM for 2D DLR was underestimated compared with that for 2D SSFP. In this study, the image quality scores of endocardial edge definition and motion artifacts were significantly impaired in 2D DLR sequences. Therefore, the total epicardial volume and the total endocardial volume for the calculation of LVM should be more affected by the technical limitations of 2D DLR cine. This technical limitation should be considered; thus, further innovations are needed to reduce blurring in the 2D DLR cine. However, 2D DLR cine enabled free-breathing scans and therefore made rapid cine imaging accessible to more patients. Thus, the results

Development of respiratory-gated real-time twodimensional cine incorporating deep learning reconstruction
In recent years, DL methods have markedly progress to accomplish various radiology tasks, focusing primarily on lesion detection, diagnosis, or classification, risk prediction, and image segmentation or quantification. In recent years, numerous methods have been developed to enhance and expedite CMR data acquisition, image reconstruction, post-processing and analysis [9,18]. In comparison with DL super-resolution approaches, model-based DLR method directly leverages the raw k-space data and coil sensitivity information to allow highly accelerated cine acquisition. This approach demonstrated better image quality and biventricular volumetry than state-of-the-art CMR with parallel imaging and compressed sensing in healthy adult volunteers [10]. A recent study also reported that free-breathing 2D DLR cine is an accurate and reproducible technique for the evaluation of biventricular volumes and function in children and young adults [11].

Clinical application of respiratory-gated real-time two-dimensional cine incorporating deep learning reconstruction
As shown in the case examples in our study, differences in endocardial edge definition and motion artifacts were

Limitations
First, 2D bSSFP and 2D DLR cine were only evaluated in the short-axis plane, and translation to other planes has not been confirmed. Further studies are needed, in which other planes were corrected to make accurate measurements of ventricular volumes. Second, patients with the congenital heart disease (CHD) who have a dilated RV was not included in this study. CMR parameters of RV size, function, and hypertrophy have been identified as predictors of death and heart failure especially in patients with CHD [21]. However, lesser interobserver agreement in general for RV parameters and machine-derived versus radiologist measurements was    LV end-systolic volume, LVSV: LV stroke volume, LVEF: LV ejection fraction, LVM: LV mass, RVEDV: right ventricular (RV) end-diastolic volume, RVESV: RV end-systolic volume, RVSV: RV stroke volume, RVEF: RV ejection fraction Fig. 3 A representative case of 2D bSSFP and 2D DLR imaging in a patient with dilated cardiomyopathy A 43-year-old man underwent cardiac magnetic resonance imaging for dilated cardiomyopathy. End-diastole and end-systole with the base were located through apex slices. The image quality scores were based on three main criteria: blood-to-myocardial contrast, endocardial edge delineation, and presence of motion artifacts throughout the cardiac cycle. Each criterion was graded on a scale of 1-5, where 1 was nondiagnostic; 2, suboptimal but still diagnostic for volumetric analysis; 3, adequate; 4, good; and 5, excellent. A representative case shows similar diagnostic image quality between bSSFP (score 5 for each criterion) and DLR for dilated ventricular volumetry, with only subtle motion blurring and degradation of the papillary muscle edge delineation with 2D DLR cine imaging (score 5 for blood-to-myocardial contrast, score 4 for endocardial edge delineation and motion artifact; white arrows) bSSFP: balanced steady-state free precession, DLR: deep learning reconstruction to evaluate the diagnostic accuracy of 2D DLR cine in each cardiac disease. Fifth, it is difficult to discuss whether volumetric analysis errors are due to scan technique or resolution of the image in this study. This study focused on how to overcome the technical limitations of 2D bSSFP using respiratory-gated real-time 2D cine incorporating DLR. Therefore, spatial and temporal resolution of 2D DLR cine in this study are lower than those of 2D bSSFP to reduce the scan time. Further studies are needed to investigate the image quality and the volumetric analysis between the 2D bSSFP and DLR cine under the same spatial and temporal resolution. Finally, owing to the relatively small sample size, additional studies with larger sample are needed to evaluate the robustness of the 2D DLR cine technique. attributed to intrinsically more complex RV geometry and inherent challenges in contour automation, respectively [7,8,19,22]. Moreover, delineating the endocardium of the RV is more challenging than that of LV owing to a greater proportion of trabeculated myocardium in patients with CHD [23]. Further studies are need for the evaluation of dilated RV in patients with CHD. Third, the patients with inability to sustain a breath-hold were excluded from the study. Further studies are needed to evaluate the image quality between the 2D bSSFP and DLR cine in patients who unable to hold their breath to apply the DLR cine to clinical situation. Fourth, we did not evaluate regional myocardial wall motion and other morphologic abnormalities in both sequences. For the 2D DLR cine to ultimately be adopted, it must demonstrate the similar diagnostic accuracy and reader confidence to 2D bSSFP across the spectrum of conditions in each cardiac disease. Further studies are needed Fig. 4 A representative case of 2D bSSFP and 2D DLR imaging in a patient with cardiac sarcoidosis A 70-year-old women underwent cardiac magnetic resonance imaging for suspected for cardiac sarcoidosis. End-diastole and end-systole with the base were located through apex slices. A representative case shows similar diagnostic image quality between bSSFP (score 5 for each criterion) and DLR for the detection of the thinning in the inter-ventricular septum (white arrowheads). There was subtle motion blurring and degradation of the endocardial edge delineation with 2D DLR cine imaging (score 5 for blood-to-myocardial contrast, score 4 for endocardial edge delineation and presence of motion artifact; white arrows) bSSFP: balanced steady-state free precession, DLR: deep learning reconstruction critically for important intellectual content. All authors approved the submitted version.
Funding Not applicable.

Data availability
The dataset analyzed during the present study is available from the corresponding author on reasonable request.

Declarations
Ethics approval and consent to participate The study was approved by the institution's human research committee (Iwate Medical University). All methods were carried out in accordance with relevant guidelines and regulations, and informed consent was obtained from all individual participants included in the study.

Consent for publication
Written informed consent for publication was obtained from all participants.

Conclusion
In this study, 2D DLR cine enabled motion-robust singleheart-beat imaging during free-breathing, making rapid cine imaging accessible in a cohort of adult patients. We successfully found good reproducibility of biventricular volumes and EF between 2D bSSFP and 2D DLR cine sequences despite some drawbacks in the quality of 2D DLR cine images. Fig. 6 A representative case of 2D bSSFP and 2D DLR imaging in a patient with chronic myocarditis A 61-year-old men underwent cardiac magnetic resonance imaging for suspected for chronic myocarditis. End-diastole and end-systole with the base were located through apex slices. There was slight degradation of endocardial edge delineation and slight motion blurring even in bSSFP (score 5 for blood-to-myocardial contrast, score 4 for endocardial edge delineation and motion artifact), but DLR cine image appeared ill-defined most notable in the left ventricular inferolateral wall due to substantial motion artifact and blurring (score 4 for bloodto-myocardial contrast, score 3 for endocardial edge delineation and motion artifact; white arrows) bSSFP: balanced steady-state free precession, DLR: deep learning reconstruction