This cross-sectional, single-center study with a prospectively acquired study cohort was performed in concordance with the Declaration of Helsinki and the International Conference on Harmonization of Good Clinical Practice. Study design, information processing, and implementation were approved by the institutional review board. All participants (healthy volunteers and patients) gave written consent before inclusion in the study.

Study participants

Prospective random enrollment of patients with clinical indications for contrast-enhanced CMR occurred between November 2022 and February 2023. Additionally, healthy volunteers over the age of 18 years old without any known cardiac diseases were prospectively recruited. Inclusion criteria were: age over 18 years old, able to give consent. Exclusion criteria were: pregnancy, implementation of cardiac pacemakers, or other contraindications for examinations on 3.0-T MRI scanners.

CMR protocol and normal resolution reconstructions

All sequences were acquired using a 3.0-T scanner (Philips Ingenia 3.0-T; Philips Healthcare) using a 16-channel torso coil with a digital interface. Patients underwent a routine CMR protocol comprised of electrocardiogram-triggered bSSFP cine imaging acquired at normal resolution (cineNR) in short-axis views (field of view: 250 × 250 mm2, repetition time: 3.1 ms, echo time: 1.54 ms, flip angle: 45°, in-plane resolution: 1.89 × 1.96 mm2 [reconstructed: 1.04 × 1.04 mm2], slice thickness: 8 mm, temporal resolution: 45 ms, compressed sensitivity encoding (Compressed SENSE) factor: 2.5), 4-chamber views, 2-chamber views, and 3-chamber views. Additionally, patients received a standard of care protocol consisting of T2 short-tau inversion recovery sequences, T1 and T2 mapping, and segmented inversion-recovery gradient-echo sequences for late gadolinium enhancement (LGE) using the Look-Locker method [15] after intravenous contrast injection (0.2 mmol/kg of body weight bolus of gadoterate meglumine [Clariscan; GE Healthcare]). For study purposes, healthy volunteers and patients (in addition to the standard of care imaging after contrast injection), underwent an electrocardiogram-triggered low-resolution bSSFP cine was acquisition (cineDL) in short-axis views (field of view: 250 × 250 mm2, repetition time: 2.9 ms, echo time: 1.34 ms, flip angle: 45°, in-plane resolution: 2.98 × 3.00 mm2 [reconstructed: 1.04 × 1.04 mm2], slice thickness: 8 mm, temporal resolution: 45 ms, Compressed SENSE factor: 2.5) and 4-chamber views, which were acquired after the cineNR sequences and reconstructed with an SR DL algorithm.

DL image reconstruction

Images were reconstructed using a vendor-provided prototype (Philips NGSA patch). Only non-industry personnel had full access to all acquired study data. A series of convolutional neural networks (CNNs) were applied to the raw low-resolution k-space acquisitions as previously described [14]. The Aadaptive-CS-Net facilitated sparsity-constrained reconstruction of acquired images with Compressed SENSE-based variable density under-sampling patterns [16,17,18] and was applied during coil combination, removing noise and under-sampling artifacts [19]. Adaptive-CS-Net integrated multiscale sparsification with a CNN-based sparsifying approach with image reconstruction of Compressed SENSE, ensuring data consistency. Domain-specific knowledge such as image background location and coil sensitivity distribution was also incorporated to replace the usual process of wavelet transformation. A multilayer approach outputs each scale transformation consisting of 2D convolutional rectifier layers and a maximum pooling layer for downscaling or a bilinear interpolation for upscaling to down- or upscale transforms via direct and skip connections. Regularization optimization was performed by trained threshold levels for each connection. The Adaptive-CS-Net was pretrained on about 740,000 pairs of images of various contrasts and subsampling levels with applied supervised learning. The final output is a de-aliased, denoised MR image with preserved magnitude and phase. Subsequently, a second CNN, Precise Image Net, was applied to remove ringing artifacts and to replace the traditional zero-filling strategy to increase the matrix size and therewith the sharpness of the images [20, 21]. The combination of these CNNs made up the SR network [22, 23]. The network was trained on over six million pairs of low- and high-resolution data with k-space crops to induce ringing. This was achieved by repeatedly cascading a pair of 2D convolutional and rectifier layers ending with a data consistency check. Supervised learning was performed: for each image, a high-resolution version was downscaled to a lower-resolution image with truncation artifacts. Data consistency checks were implemented to match the resulting k-space with the measured k-space data. This study utilized a moderate level of noise reduction. Reconstructions were performed on scanner hardware equipped with an Nvidia Quadro RTX5000 GPU.

Objective image analysis

Objective image analysis was performed by two board-certified cardiovascular radiologists (J.A.L. with 12 years of experience in CMR and D.Kr. with 5 years of experience in CMR) using dedicated software (IntelliSpace Portal, version 12.1.4; Philips Medical Systems). Left ventricular ejection fraction (LVEF), left ventricular end-diastolic volume index (LVEDVi), and interventricular septum thickness at diastole (IVSD) were measured in both groups. Apparent signal-to-noise ratios (aSNR) and apparent contrast-to-noise ratios (aCNR) were calculated as previously described [14]. Myocardial global systolic longitudinal, circumferential, and radial strain were calculated by using automatic feature tracking strain analysis software (Medis Suite MR, version 4.0.62.4, Medis Medical Imaging Systems) with manual corrections when necessary. Since cineDL was only applied to 4-chamber and short-axis views, the global longitudinal strain was calculated from 4-chamber views for both cineNR and cineDL.

Subjective image analysis

Subjective image quality analysis was performed by two board-certified cardiovascular radiologists (D.Kr. with 5 years and A.I. with 6 years of experience in CMR). Subjective image quality was rated for cineNR and cineDL short-axis and 4-chamber views on a 5-point Likert scale regarding three image criteria: blood-pool to myocardium contrast, endocardial edge definition, and artifacts, as previously described [3, 24]. Raters were blinded and sequences were presented in random order. A total score was determined by the equal weight average of all three criteria:

1. Non-diagnostic: poor contrast between blood pool and myocardium, endocardial edge poorly defined, and artifacts render the images non-diagnostic.

2. Poor: blood pool barely discernable from the myocardium, washed-out endocardial edge, and blurring of trabeculae and numerous artifacts.

3. Adequate: blood pool discernable from myocardium but features lots of noticeable variation throughout the cardiac cycle, barely distinguishable endocardial edge definition, and some artifacts are present.

4. Good: the blood pool is mostly brighter and discernable from the myocardium, papillary and endocardial trabeculae are discernable but blurred in some images during the cardiac cycle, few artifacts are present but do not hinder image quality.

5. Excellent: the blood pool is hyperintense and clearly discernable from the myocardium in all images, and papillary and endocardial trabeculae are clearly visible with no blurring, and almost no artifacts.

Statistical analysis

Statistical analysis was performed using Prism (version 10.1.0; GraphPad Software) and SPSS (version 29; IBM). Continuous variables for quantitative measurements are reported as means ± standard deviation (SD) and nominal data as percent to absolute frequency. The Shapiro–Wilk test was used to check for normality. Pearson’s correlation was used to compare the correlation between cineNR and cineDL volumetry results, aSNR, and aCNR. Median and interquartile range (IQR) are provided for nonparametric data or when normality cannot be assumed. Volumetric findings and acquisition times were compared using the paired Student’s t-test. The chi-squared test was used for nominal data comparisons. Subjective image scores were compared using the Wilcoxon matched-pairs signed rank test. Inter-rater agreement for subjective image quality, aSNR, aCNR, strain, and volumetry was compared using a two-way mixed effects intraclass correlation coefficient (ICC) model for absolute agreement. ICC was rated as poor (less than 0.5), moderate (0.5–0.75), good (0.75–0.9), and excellent (greater than 0.90) [25]. The level of statistical significance was set to p < 0.05.