Twenty healthy volunteers (20 males) were enrolled. All volunteers provided written informed consent and underwent CS CMRA and conventional CMRA without contrast medium. The age, height, weight, and body mass index of the volunteers ranged from 23 to 43 (mean ± standard deviation [SD], 30.9 ± 5.2) years, 160 to 182 (mean ± SD, 170.8 ± 5.9) cm, 52 to 89 (mean ± SD, 67.2 ± 11.3) kg, and 18.1 to 28.3 (mean ± SD, 22.9 ± 3.0) kg/m2. None of the volunteers had any history or risk of heart disease. The study protocol was approved by the hospital’s Institutional Review Board.

All MRI examinations were conducted using a clinical 3 T MR scanner (Signa Architect 3 T; GE Healthcare, Waukesha, WI, USA). CMRA was performed using electrocardiogram-triggered, and navigator-gated techniques. First, Scout images of the heart in the axial, sagittal, and coronal views were acquired. Subsequently, a long-axis cine sequence with ECG triggering was acquired to determine the CMRA data acquisition time. The voltage-specific acquisition time was set by an investigator according to the phase of minimal right coronary artery (RCA) motion by observing the coronary artery stagnation time. The CMRA scans were acquired using a T2-prepared segmented three-dimensional spoiled GRE. Chemical shift selective was used as fat saturation to improve coronary artery delineation and suppress fat-related aliasing artifacts. Conventional CMRA uses parallel imaging (PI) with an acceleration factor of 2. The CS CMRA employed a combination of PI and CS with an acceleration factor of 3. Currently, PI is widely used in clinical MR, but because of the increased noise at higher acceleration factors, it is standardly limited to a factor of approximately 2 accelerations to maintain reliable visualization [6, 14, 15]. Research indicates that CS acceleration factors typically range from 3 to 9 [16]. Since this is a preliminary study and a high acceleration factor is expected to degrade image quality, an acceleration factor of 3 was first considered. The detailed imaging parameters are listed in Table 1. The imaging order of CS and conventional CMRA scans was randomized. Nitroglycerin sublingual spray (0.3 mg) was administered before the CMRA scan. Respiratory motion was detected by placing the navigator on the right hemidiaphragmatic dome. An edge-detection algorithm was subsequently used to identify the location of the lung-liver inface. The ratio detected by the navigator within the window setting was used as the acceptance rate.

First, PI consistently undersamples the k-space. In this study, we used autocalibrating reconstruction for Cartesian imaging (ARC) as PI [17]. A feature of the ARC is that the central part of the k-space is used for self-calibration, and unfolding is performed in the x-ky-kz domain. The ARC algorithm was supplied by the vendor.

Second, CS is a technique used for randomly undersampling the k-space [5]. The CS algorithm minimizes the L1 norm of the image after a sparse transformation by estimating the undersampled points. CS requires iterative calculations. A feature of CS is that it reduces the imaging time without reducing the SNR, which can be reduced by undersampling. The CS algorithm was supplied by the vendor. To estimate the undersampled k-space, we combine ARC with CS [18].

Third, AIR Recon DL was used as the DLR method [19]. AIR Recon DL performs DLR in the k-space domain rather than in the image domain (Fig. 1). When converting k-space data into an image, a Lorentzian-shaped k-space filter is sometimes used to reduce aliasing artifacts; however, this reduces image sharpness. AIR Recon DL fulfills the outer regions of the k-space without using a k-space filter. This simultaneously reduces aliasing artifacts and improves sharpness and SNR. The research version of the AIR Recon DL algorithm was provided by a vendor. The product version of the AIR Recon DL can be used for PI but not for CS. The research version of the AIR Recon DL can be used for PI and CS. The product and research versions use nearly identical principles. Product version uses Graphic Processor, and post-processing can be done automatically. However, in the research version, post-processing must be done manually on a conventional processor on the host computer. In DLR, ARC, and CS, raw data are treated as complex-number data. AIR Recon DL was performed offline on CS CMRA data after scanning.

AIR Recon DL algorithm. AIR Recon DL estimates and complements high-spatial frequency information in k-space filled with zeros by zero-fill interpolation processing. In addition, AIR Recon DL denoises k-space data. Thus, AIR Recon DL sharpens images, reduces noise and truncation artifacts

Two radiologists with 15 years (reader 1) and 5 years (reader 2) of experience in cardiac imaging independently assessed the qualitative image quality of the coronary arteries. Coronary arteries were assessed separately for the RCA, left anterior descending artery (LAD), and left circumflex artery (LCX). The assessment focused on the vessel sharpness and artifacts. Coronary arteries were assessed mainly in the proximal to middle locations. We used a 4-point subjective scale for qualitative image analysis: 4, excellent (vessel well visualized with sharply defined borders); 3, good (vessel adequately visualized with only mildly blurred borders); 2, fair (coronary vessel visible but with low confidence in diagnosis due to moderately blurred borders); and 1, poor (coronary vessel barely visible or obscured by noise) [6] (Fig. 2). All image quality assessments were performed in the axial orientation. The image quality assessed by reader 1 was used, and the inter-observer agreement with reader 2 was calculated for subsequent analysis.

The 4-point subjective scale for qualitative image analysis

Quantitative image assessment was performed by a radiologist with 5 years of experience. The CSAI, CS, and conventional CMRA images were assessed using the dedicated workstation (SYNAPSE VINCENT; Fujifilm Corp., Ltd., Tokyo, Japan). The regions of interest (ROIs; size: 80–100 mm2) were set without artifacts on the ventricular septum and left ventricular (LV) blood pool on the same slice to define the SNR and CNR (Fig. 3). The ROIs of the three images were set as much as possible at the same site. The SNR and CNR were defined using the following equation: SNR = signal intensity of the myocardium (SImyo)/standard deviation of the myocardium (SDmyo), CNR = (signal intensity of blood [SIblood]/standard deviation of the blood [SDblood])-(SImyo/SDmyo) [20].

The ROIs placement for the calculation of signal-to-noise ratio and contrast-to-noise ratio. The ROIs were set without artifacts on the ventricular septum and left ventricular blood pool on the same slice. ROI, region of interest

Vessel sharpness in the RCA#1, LAD#6, and LCX#11 was evaluated using the following methods: The signal intensity profiles were obtained along a user-defined line perpendicular to the major axes of the vessel. The vessel sharpness was assessed by calculating the 20th and 80th percentile points between the maximum and background signal intensities for each side of the signal intensity profile. The distance between these two points was then determined for each side in millimeters. Vessel sharpness was defined as the reciprocal of the average distance between the two points [21].

Statistical analyses were performed using a statistical software (JMP version 13; SAS Institute, Cary, North Carolina, USA). Continuous variables are presented as mean ± SD or median (first and third quartiles). Paired t tests were used to compare the acceptance rates. The Wilcoxon matched-pairs signed-rank test and Friedman test were used to compare the imaging times, image quality, SNR, CNR, and vessel sharpness. The quadratic-weighted kappa test was used to evaluate the inter-observer agreement of image quality (> 0.81, excellent agreement; 0.61–0.80, good agreement; 0.41–0.60, moderate agreement; 0.21–0.40, fair agreement; and < 0.20, poor agreement). A P value < 0.05, and Bonferroni correction was used to reduce the chance of false-positive results (type I error) when multiple pairwise tests were performed.