Artificial intelligence-assisted compressed sensing CINE enhances the workflow of cardiac magnetic resonance in challenging patients

Abstract

BACKGROUND

A key cardiac magnetic resonance (CMR) challenge is breath-holding duration, difficult for cardiac patients.

AIM

To evaluate whether artificial intelligence-assisted compressed sensing CINE (AI-CS-CINE) reduces image acquisition time of CMR compared to conventional CINE (C-CINE).

METHODS

Cardio-oncology patients (n = 60) and healthy volunteers (n = 29) underwent sequential C-CINE and AI-CS-CINE with a 1.5-T scanner. Acquisition time, visual image quality assessment, and biventricular metrics (end-diastolic volume, end-systolic volume, stroke volume, ejection fraction, left ventricular mass, and wall thickness) were analyzed and compared between C-CINE and AI-CS-CINE with Bland–Altman analysis, and calculation of intraclass coefficient (ICC).

RESULTS

In 89 participants (58.5 ± 16.8 years, 42 males, 47 females), total AI-CS-CINE acquisition and reconstruction time (37 seconds) was 84% faster than C-CINE (238 seconds). C-CINE required repeats in 23% (20/89) of cases (approximately 8 minutes lost), while AI-CS-CINE only needed one repeat (1%; 2 seconds lost). AI-CS-CINE had slightly lower contrast but preserved structural clarity. Bland-Altman plots and ICC (0.73 ≤ r ≤ 0.98) showed strong agreement for left ventricle (LV) and right ventricle (RV) metrics, including those in the cardiac amyloidosis subgroup (n = 31). AI-CS-CINE enabled faster, easier imaging in patients with claustrophobia, dyspnea, arrhythmias, or restlessness. Motion-artifacted C-CINE images were reliably interpreted from AI-CS-CINE.

CONCLUSION

AI-CS-CINE accelerated CMR image acquisition and reconstruction, preserved anatomical detail, and diminished impact of patient-related motion. Quantitative AI-CS-CINE metrics agreed closely with C-CINE in cardio-oncology patients, including the cardiac amyloidosis cohort, as well as healthy volunteers regardless of left and right ventricular size and function. AI-CS-CINE significantly enhanced CMR workflow, particularly in challenging cases. The strong analytical concordance underscores reliability and robustness of AI-CS-CINE as a valuable tool.

Key Words: Cardiac magnetic resonance; CINE imaging; Artificial intelligence; Compressed sensing; Imaging workflow; Acquisition time; Cardiac function; Cardio-oncology; Image quality; Challenging patients

Core Tip: In this prospective study of 89 patients and volunteers, we demonstrate that artificial-intelligence-assisted compressed sensing (AI-CS-CINE) significantly streamlines cardiac magnetic resonance imaging workflows, reducing acquisition time by 84% (37 seconds vs 238 seconds) compared to conventional CINE imaging. Quantitative analysis showed excellent agreement in biventricular volumes and function (intraclass correlation coefficient 0.73-0.98). AI-CS-CINE proved especially valuable in challenging cases, such as for patients with cardiac amyloidosis, enabling faster acquisition and more reliable interpretation. These findings highlight AI-CS-CINE as a robust, time-efficient alternative to conventional methods, with potential to improve clinical efficiency and image quality in diverse cardiac populations.

INTRODUCTION

Cardiac magnetic resonance (CMR) is the gold standard for noninvasive evaluation of cardiovascular disease, providing robust quantitative assessments critical for diagnoses, serial monitoring, and healthcare management[1-4]. CINE magnetic resonance imaging (MRI) is used to acquire fast image acquisitions with very short repetition and echo times, e.g., balanced steady-state free precession (SSFP)[5,6]. Cardiac metrics such as end-diastolic volume (EDV), end-systolic volume (ESV), and ejection fraction (EF) for both ventricles are typically measured from a series of cardiac CINE images[3,7].

Conventional CINE (C-CINE) uses segmented CMR acquisition over multiple cardiac cycles to create a single image, but faces two key challenges. First, image quality relies on controlled breath holding, which many patients, particularly those with heart failure, struggle to maintain, leading to motion artifacts. Second, the prolonged acquisition time needed to combine heartbeats can cause misalignment, reducing image quality and introducing variability in quantitative measurements[5,8].

Compressed sensing (CS) accelerates image acquisition by combining k-space under-sampling with iterative reconstruction, but the required long reconstruction times are a barrier to clinical utility[8,9-12]. In the current study, artificial intelligence (AI) was used to overcome and accelerate CS-CINE reconstruction as well as acquisition, resulting in AI-CS-CINE. AI-CS-CINE integrates CS, parallel imaging, and half Fourier analysis, significantly reducing acquisition and reconstruction time compared to C-CINE[13-17].

In clinical practice, one of the major challenges of CMR is the need for a comprehensive imaging protocol, with most sequences requiring patients to hold their breath to achieve high-quality images. This workflow is particularly difficult for patients with various cardiac conditions that cause inconsistent breathing or, in some cases, an inability to hold their breath. As a result, many patients struggle to complete CMR examinations. Challenging cases make up nearly a quarter of our referrals, often involving individuals with the greatest medical need.

To address these challenges, in the current study, we aimed to explore the incorporation of AI-CS-CINE into the CMR workflow on a 1.5-T scanner to enhance the process by enabling near real-time acquisition, faster reconstruction, and preserve image quality for both diagnosis and longitudinal measurements. We compared AI-CS-CINE with C-CINE for quantifying left ventricle (LV) and right ventricle (RV) structure-function metrics in both referred cardiology patients and healthy volunteers. Additionally, we assessed the feasibility and efficiency of AI-CS-CINE in complex cases, particularly in cardiac amyloidosis, which requires an extensive CMR protocol and longer acquisition times.

MATERIALS AND METHODS

Participant characteristics

The CMR protocol was approved by the institutional ethics committee, and written informed consent was obtained from all participants. Data were prospectively acquired from 100 participants (32 healthy volunteers and 68 patients with various cardiac conditions) between May 2023 and January 2024. Including both groups allowed assessment of variations across a broad range of CMR metrics. Healthy volunteers, with no history of cardiovascular events or risk factors (e.g., hypertension or diabetes), were scanned without gadolinium contrast and had normal cardiac parameters (including normal LV and RV volumes, native myocardial T1 and T2 relaxation times in accordance with Society for Cardiovascular Magnetic Resonance, and institutional reference ranges). We excluded participants with inadequate image quality for analysis due to poor breath holding and/or ECG triggering. Out of 100 acquired cases, three cases were excluded due to poor image quality that rendered them unsuitable for analysis. Eight cases were identified as outliers because the image quality of C-CINE was substantially inferior to that of AI-CS-CINE (Results and Supplementary material).

The study ultimately included 89 participants with an average age of 58 years, comprising 42 males and 47 females (Figure 1). Baseline demographics and characteristics are provided in Table 1. Some patients were referred to CMR to evaluate for more than one potential diagnosis, resulting in overlapping referral categories in Table 1.

Figure 1 The flowchart shows exclusion criteria and the resulting cardio-oncology and healthy participants eligible and included in this study. C-CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted compressed sensing CINE; CMR: Cardiac magnetic resonance.

Table 1 Participant demographics.

	Male	Female	Total
Participants (%)	42 (47)	47 (53)	89
Patients	26 (43)	34 (57)	60
Volunteers	16 (55)	13 (45)	29
Ethnicity (%)	White	Black	Hispanic	Asian	Total
Participants	68 (76)	15 (18)	1 (1)	5 (5)	89
Patients	46 (77)	13 (21)	1 (2)	0 (0)	60
Volunteers	22 (77)	2 (7)	0 (0)	5 (16)	29
Age (year)	Male	Female	Total
Participants	57.3 (19-83)	59.6 (18-83)	58.4 (18-83)
Patients	62.5 (30-88)	63.0 (18-83)	62.8 (18-83)
Volunteers	48.6 (19-77)	50.8 (24-71)	49.6 (19-77)
BSA (m2)	Male	Female	Total
Participants	2.04 (1.70-2.45)	1.84 (1.31-2.87)	1.93 (1.31-2.87)
Heart rate	Male	Female	Total
Participants	66 (46-113)	67 (47-108)	67 (46-113)
Clinical referral indication	Patients
r/o cardiotoxicity	29
r/o amyloidosis	30
r/o myocarditis	3
r/o cardiomyopathy	17
Arrhythmia	9
Chest pain	9

Data of age, body surface area and heart rate are mean (min-max). BSA: Body surface area.

CMR acquisition

All participants underwent CMR examination on a 1.5-T scanner (uMR570; United Imaging Healthcare, Shanghai, China) with a body coil (12 channel) with ECG triggering and respiratory monitoring (Invivo Corp, Orlando, FL, United States). A short-axis stack of 11 CINE images and 2-, 3-, 4-chamber CINE images in expiratory breath hold was acquired for ventricular volume and function estimates with C-CINE and AI-CS-CINE at matched slice positions to facilitate the intra-participant comparison during the same scanning session. The shortest possible breath hold for C-CINE was 11 seconds to acquire one slice, and 2 seconds for AI-CS-CINE to acquire one slice, which appeared to be rhythmic breathing rather than breath holding. Consequently, AI-CS-CINE with 2 seconds slice acquisition was employed successfully with only rhythmic breathing in a subset of patients. AI was used for both acquisition and reconstruction during AI-CS-CINE. The imaging parameters of C-CINE and AI-CS-CINE are summarized in Table 2.

Table 2 Conventional CINE and Artificial-intelligence-assisted compressed sensing CINE image parameters.

	C-CINE	AI-CS-CINE
ECG mode	Retrospective	Retrospective
TR/TE (millisecond)	3.57/1.75	2.74/1.28
Image matrix	224 × 85	192 × 100
Reconstruction matrix	2.0	1.5
Spatial resolution (mm)	1.89 × 1.61	1.88 × 1.88
Flip angle (°)	80	60
Bandwidth (Hz/pixel)	1500	1200
Temporal resolution (millisecond)	54	41
Reconstructed cardiac phases	25	25
Field of view (mm)	360 × 320	360 × 320
Slice thickness (mm)	8	8
Gap (mm)	0	0
Number of slices	11	11
Longest breath-hold for acquisition (second)	11	11
Number of slices acquired per 11-sec breath hold	1	6
Number of breath-holds	11	2
Shortest breath-hold time to acquire one slice (second)	11	2
Total acquisition time including breath-holds (second)	238	37

C-CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted compressed sensing CINE.

The acquisition time was recorded retrospectively for both C-CINE and AI-CS-CINE. Image reconstruction for both techniques was automatically completed in real-time after the acquisition. Consequently, the primary improvement in the workflow stemmed from the faster acquisition process and real-time reconstruction of AI-CS-CINE. Upon reconstruction, the technologist quickly reviewed image quality to determine if a repeat scan was necessary. In most cases, additional time was required to repeat C-CINE scans for patients with unsatisfactory breath-holds. The extra time spent on repeat scans was also documented. Additionally, participants received a standard-of-care protocol which included T1 and T2 mapping, strain, perfusion, and late gadolinium enhancement for diagnostic determinations.

AI-CS-CINE image reconstruction

The AI model used for image reconstruction is detailed in a previous study[14]. AI-CS-CINE images were reconstructed using a neural network called Res-CRNN, which unrolls over five iterative stages. Each stage refined the output from the previous one and included an independent subnetwork with three bidirectional convolutional gated recurrent unit (GRU) layers, 2D convolutional layers, data consistency layers, and two levels of residual connections. GRU layers captured both forward and backward temporal dependencies, while 2D convolutions offered efficient spatial processing with reduced memory requirements. Residual connections enhanced high-frequency detail recovery, and data consistency layers enforced fidelity to the originally acquired k-space data.

The model simultaneously reconstructed multi-coil images, treating each coil as a separate input channel and combining them using root sum-of-squares. Training was conducted using 1610 retro-cine datasets simulated with the same k-space sampling pattern used in the study. These datasets included a variety of cardiac anatomies and imaging conditions, helping to improve model robustness across different clinical presentations.

Qualitative image assessment between C-CINE vs AI-CS-CINE

Image analysis was qualitatively evaluated by a physician according to four criteria: overall image quality (including contrast and signal-to-noise ratio), ability to assess wall motion and suitability for segmentation, visualization of valve morphology and function, and the presence of imaging artifacts. Each aspect was rated on a 5-point scale, where a score of 5 indicated excellent quality or performance or no artifacts; 4, very good quality or minimal artifacts; 3, adequate quality or some artifacts, though still diagnostic; 2, poor quality or obvious artifacts that impaired interpretation; and 1, non-diagnostic images or severe artifacts.

Biventricular function analysis

Image analysis was performed by a technologist with 25 years of experience in CMR. LV and RV volumes and functions were evaluated with built-in scanner software (Cardiac Analysis). Short-axis stacks of 11 CINE images were loaded into the software and the software automatically delineated the contours of endo-myocardium and epi-myocardium. Papillary muscles were included in the blood pool. The technologist reviewed the correctness of the automatic segmentation and could manually fine-tune the bounds on a slice-by-slice basis, if necessary. The EDV, ESV, stroke volume (SV), EF of LV and RV, and LV mass (LVM) and wall thickness were obtained and normalized to body surface area and indexed metrics were reported. EF was calculated as follows: EF = (EDV - ESV)/EDV × 100.

Statistical analysis

Statistical analysis was performed using R software version 4.3.2 (R Foundation for Statistical Computing). The mean and SD for each CMR measurement was calculated and further compared between male and female with unpaired t-test. The Shapiro–Wilk Test was used to confirm normal distribution of measurement differences for intra-participant comparison and correlation evaluation.

The time of image acquisition for C-CINE and AI-CS-CINE was compared with a paired Student t-test. Qualitative image quality scores of C-CINE and AI-CS-CINE were compared with the Wilcoxon Rank test. Agreement between quantitative measurements from C-CINE and A-CINE acquisitions methods was evaluated with Bland–Altman plots[18,19]. The differences were plotted against the averages of the two measurements. Bias was calculated as the mean difference, and the variation about this mean was calculated as the SD of the differences. The limits of agreement (LoA_{1.96 SD,diff}) were defined as the bias plus and minus 1.96 times the SD of the differences. If the differences were normally distributed, these provided an interval within which 95% of differences between measurements by the two methods were expected to lie[18,19]. The 95% confidence intervals (CIs) were calculated for both bias and LoA. Intraclass coefficient (ICC) estimates and their 95%CI were calculated based on a single measurement (k = 1), absolute agreement, two-way mixed-effects model. Values less than 0.5, between 0.5 and 0.75, between 0.75 and 0.9, and greater than 0.90 were indicative of poor, moderate, good, and excellent reliability, respectively[20,21].

RESULTS

Participant characteristics

The study ultimately included 89 participants including 60 patients (31 of 60 had confirmed cardiac amyloidosis) and 29 healthy volunteers (Table 1). Notable quantitative differences between the C-CINE and AI-CS-CINE scans related to image quality were observed in eight cases (designated as outliers). Detailed results for comparisons are presented, both for the dataset excluding outliers (n = 89) and the complete dataset including outliers (n = 97; Supplementary material).

Efficiency and reliability of AI-CS-CINE compared to C-CINE in enhancing CMR workflow

The average total acquisition-reconstruction time for the short-axis stack views was 238 seconds with C-CINE and 37 seconds with AI-CS-CINE sequence, resulting in 84% time saving with AI-CS-CINE. The acquisition time included only the actual scans, excluding pauses between breath-holds. During CMR, C-CINE sequences often needed to be repeated due to unsatisfactory image quality. The time loss associated with evaluating and repeating C-CINE sequences during the scan was compared with AI-CS-CINE. Across all 89 participants, time loss for C-CINE ranged from 0 to 8 minutes, and C-CINE had to be repeated in 20 cases [23%; 12 male (approximately 8 minutes) vs 8 female (approximately 5 minutes)], whereas AI-CS-CINE required only one repeat scan (female) out of 89 cases (1%) with only 2 second loss. This highlighted the efficiency and reliability of AI-CS-CINE in the workflow of CMR.

Qualitative image assessment between C-CINE vs AI-CS-CINE

In general, AI-CS-CINE images had a slightly lower contrast but smoother appearance than C-CINE. All anatomical structures and functions were well preserved by AI-CS-CINE (Figures 2 and 3). The use of blood volume dephasing as a qualitative marker of valve regurgitation is flow-rate dependent with high velocity regurgitant jets, regardless of valve, clearly seen as sharp dark phase artifacts in some cases. Mild regurgitant jets (1+ to 2+) were less clearly defined compared with C-CINE. Normal and hyperdynamic wall motion was easily appreciated by both AI-CS-CINE and C-CINE. Segmental regions of hypokinesis suggested on the traditional C-CINE images were more clearly observed and corroborated using the AI-CS-CINE images. Occasional intracardiac thrombus seen in the left atrial appendage were equivalently recognized from AI-CS-CINE and C-CINE images. The origins of coronary arteries from aortic valve cusp and their proximal anatomy were also well seen with AI-CS-CINE and C-CINE when captured in a basal short-axis slice. In general, AI-CS-CINE clarified intraventricular confusion attributable to papillary muscles, trabecular, false tendons, and respiratory motion artifacts by producing clearer ventricular endocardium-blood pool definition.

Figure 2 Magnetic resonance imaging illustrating the exceptional alignment between conventional CINE and artificial-intelligence-assisted compressed sensing CINE in a healthy 20-year-old male volunteer. A: The short-axis stack clearly depicts the apex, middle, and base of the left ventricle in both conventional CINE (C-CINE) (top row) and artificial-intelligence-assisted compressed sensing CINE (AI-CS-CINE) (bottom row); B: Additionally, the 2-, 3-, and 4-chamber views demonstrate excellent agreement between C-CINE and AI-CS-CINE in this volunteer. C-CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted compressed sensing CINE.

Figure 3 Exceptional alignment between conventional CINE and artificial-intelligence-assisted compressed sensing CINE is demonstrated in a 76-year-old female patient diagnosed with cardiac amyloidosis. A: The short-axis stack vividly displays the apex, middle, and base of the left ventricle in both conventional CINE (C-CINE) (top row) and artificial-intelligence-assisted compressed sensing CINE (AI-CS-CINE) (bottom row); B: Furthermore, the 2-, 3-, and 4-chamber views exhibit outstanding agreement between C-CINE and AI-CS-CINE in this patient. The left ventricle appears mildly dilated. Additionally, small-to-moderate pericardial effusion (indicated by arrows) is clearly visualized on both C-CINE (top row) and AI-CS-CINE (bottom row), highlighting the consistency in detecting pathological features across both sequences. C-CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted compressed sensing CINE.

The image quality scores were as follows: (1) General image quality: 5.0 ± 0.0 for both C-CINE and AI-CS-CINE (P > 0.99); (2) Wall motion and segmentation: 5.0 ± 0.0 for both C-CINE and AI-CS-CINE (P > 0.99); (3) Valve morphology and function: 5.0 ± 0.0 for C-CINE and 3.9 ± 0.6 for AI-CS-CINE (P < 0.001); and (4) Artifacts: 5.0 ± 0.0 for both C-CINE and AI-CS-CINE (P > 0.99). No significant differences were observed between C-CINE and AI-CS-CINE in any category except valve morphology and function. Despite the lower score for valve evaluation, AI-CS-CINE provided diagnostically sufficient image quality for assessing valve morphology and function (all scores ranging from 3-5).

Difference in LV and RV metrics between male and female participants with AI-CS-CINE

Expected significant size and function differences were determined between male and female groups for LV and RV EF and body-mass-indexed LVEDV (LVEDVi), LVESVi, LVMi, RVEDVi, and RVESVi, consistent with recent prior reports (all P < 0.05) (Supplementary Table 1). The LVEF of females was marginally increased compared to males (P = 0.08), and female LVEDVi and LVESVi were smaller than males. Relative wall thickness, a metric used with muscle mass to differentiate LV size as normal, concentric remodeling, concentric LV hypertrophy, or eccentric hypertrophy did not differ with sex[22]. RVEF was greater in females, but RVEDVi and RVESVi were smaller than males. No sex difference was noted for LVSVi or RVSVi. Collectively, clinically critical determinants of cardiac structure and function known to vary by sex were consistent with published reports when derived from AI-CS-CINE images[23].

Biventricular volume and function analysis

Correlation of LV and RV volume and function metrics between C-CINE and AI-CS-CINE: In this study, CMR measurements obtained with AI-CS-CINE were highly comparable to the standard-of-care C-CINE results visually and quantitatively for short-axis stack images (Figures 2, 3, and 4). Bland–Altman analysis and absolute ICC results are summarized in Table 3. The key metrics of EF, EDVi, ESVi, SVi (for both LV and RV), LVMi, and relative wall thickness were highly correlated between C-CINE and AI-CS-CINE acquisitions (Table 3, Figure 4). All LV measurements showed very good to excellent reliability (ICC_abs = 0.89-0.98). The RV ICCs were in the “good” range (ICC_abs = 0.73-0.89). LV metrics exhibited slightly better reliability than the RV determinations, which may relate to the derivation of RV metrics from stacked LV short-axis oriented slices rather than RV optimized slices.

Figure 4 Bland–Altman plots reveal a strong agreement of metrics between conventional CINE and artificial-intelligence-assisted compressed sensing CINE. The solid line represents the mean difference (bias), while dashed lines denote the upper and lower limits of agreement (LoA_1.96SDdiff). Error bars, depicting confidence intervals, are presented for both the average difference and the limits of agreement. Visual inspection of the Bland–Altman plots revealed no obvious relationship between the differences and the magnitude of measurements, nor systematic bias. For each variable, conventional CINE (C-CINE) and artificial-intelligence-assisted compressed sensing CINE (AI-CS-CINE) revealed an acceptable mean difference. Nearly all measurement pairs lie within the limits of agreement, indicating the AI-CS-CINE method is highly comparable to C-CINE. The X-axis represents the mean of C-CINE and AI-CS-CINE values, while the Y-axis shows the difference between C-CINE and AI-CS-CINE. Dashed lines indicate the mean difference and 95% limits of agreement. BSA: Body surface area; LV: Left ventricle; RV: Right ventricle.

Table 3 Alignment of left ventricle and right ventricle volume and function metrics with conventional CINE vs artificial-intelligence-assisted compressed sensing CINE excluding outliers (n = 89).

	mean ± SD		Bland–Altman		Intraclass coefficient
	C-CINE	AI-CS-CINE	Bias ± SD	LoA	r	95%CI
LVEF (%)	56.2 ± 11.0	55.9 ± 11.1	0.28 ± 3.78	-7.13, 7.69	0.94	0.91, 0.96
LVEDVi (mL/m2)	75.9 ± 19.0	73.3 ± 19.0	2.63 ± 5.30	-7.76, 13.01	0.95	0.93, 0.97
LVESVi (mL/m2)	34.4 ± 16.0	33.4 ± 15.7	1.02 ± 3.35	-5.55, 7.59	0.98	0.96, 0.98
LVSVi (mL/m2)	41.5 ± 9.3	39.9 ± 9.6	1.61 ± 4.21	-6.63, 9.85	0.89	0.83, 0.92
LVMi (mL/m2)	58.4 ± 20.3	59.3 ± 20.6	0.91 ± 4.14	-9.01, 7.20	0.98	0.97, 0.99
Relative wall thickness	0.4 ± 0.2	0.4 ± 0.2	0.00 ± 0.05	-0.10, 0.10	0.95	0.93, 0.97
RVEF (%)	50.8 ± 10.0	51.5 ± 10.3	0.66 ± 7.32	-15.01, 13.69	0.73	0.62, 0.82
RVEDVi (mL/m2)	71.0 ± 14.8	69.3 ± 14.7	1.74 ± 6.60	-11.21, 14.68	0.89	0.84, 0.93
RVESVi (mL/m2)	35.1 ± 11.0	33.6 ± 10.4	1.51 ± 5.28	-8.84, 11.87	0.87	0.81, 0.91
RVSVi (mL/m2)	35.9 ± 9.7	35.7 ± 10.5	0.22 ± 7.12	-13.74, 14.18	0.75	0.64, 0.83

C-CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted-CINE; EDV: End-diastolic volume; EF: Ejection fraction; ESV: End-systolic volume; LV: Left ventricle; LVM: LV mass; RV: Right ventricle; SV: Stroke volume; I: Normalized to body surface area.

Bland–Altman analysis: Visual inspection of the Bland–Altman plots revealed no relationship between the differences and the magnitude of measurements, nor any systematic bias. For each variable, C-CINE and AI-CS-CINE means were very close and the mean difference was acceptable. Nearly all measurement pairs lie within the limits of agreement, indicating strong comparability between methods (Figure 4). A comprehensive summary of Bland–Altman results is provided in Table 3. While some variables showed statistically significant bias, this did not affect clinical interpretation. The limits of agreement supported that the observed differences were acceptable for clinical use.

Discrepancy in anatomy and LV and RV function metrics between C-CINE vs AI-CS-CINE

Among the 100 participants, three encountered challenges with both breath holds and ECG triggering during C-CINE imaging, rendering their scans unsuitable for diagnosis or comparison. Specifically, C-CINE scans were repeated multiple times. However, the image quality still failed to reach diagnostic standards. To avoid exhausting the patients and to preserve their ability to complete the remaining sequences necessary for diagnosis, further attempts at acquiring C-CINE were discontinued. In these cases, AI-CS-CINE successfully achieved diagnostic image quality, allowing for the extraction of volumetric and functional metrics needed for clinical evaluation (Videos 1 and 2). Other disparities between the two methods were observed in eight cases, where data points fell beyond the upper and lower limits of agreement for each metric in the Bland–Altman plot. A thorough review of these eight cases revealed that the image quality of C-CINE was significantly inferior to AI-CS-CINE, with notable issues such as blurriness and artifacts caused by inconsistent breath-holds. These shortcomings resulted in less reliable metrics derived from the C-CINE images. Thus these eight instances were categorized as outliers. Nonetheless, our results, inclusive of outliers (n = 97) are presented in Supplementary material, and showcasing a strong alignment between C-CINE and AI-CS-CINE across all metrics (Figure 5; Supplementary Table 2; Supplementary Figure 1). Overall, AI-CS-CINE produced reliable images for qualitative and quantitative analysis in 100/100 participants while C-CINE did so in only 89/100 participants.

Figure 5 Artificial-intelligence-assisted compressed sensing CINE improved image quality and quantitative reliability in challenging cardiac amyloidosis cases. A: In a 75-year-old male patient diagnosed with cardiac amyloidosis and small pericardial effusion, poor image quality of the short-axis stack in conventional CINE (C-CINE) was observed due to bad breath holds and poor ECG triggering. These cardiac conditions hindered the patient’s ability to hold his breath for the required 11 seconds during C-CINE acquisition. Consequently, the endomyocardium and perimyocardium were poorly delineated, and trabeculae were not clearly visible (upper row). Such images were considered unsuitable for quantitative image analysis. However, the patient was able to hold his breath during the 2-second artificial-intelligence-assisted compressed sensing CINE (AI-CS-CINE) acquisition, enabling successful image analysis on AI-CS-CINE images only (lower row). Quantitative analysis of wall thickness on C-CINE (upper row) yielded substantially different results compared to AI-CS-CINE (lower row); B: In an 84-year-old female patient diagnosed with cardiac amyloidosis, poor ECG triggering resulted in missing imaging frames during the diastolic phase. Both C-CINE (top row) and AI-CS-CINE (bottom row) exhibited satisfactory image quality for quantitative analysis. However, quantitative metrics did not align well between C-CINE and AI-CS-CINE (right column). Due to inadequate ECG triggering, missing image frames during the diastolic phase were observed in C-CINE, with the end of the full diastolic phase not being displayed (top row). This discrepancy is evident in the 3-chamber view (middle column, arrow). In contrast, the end of the full diastolic phase was successfully displayed in AI-CS-CINE (bottom row, arrow). In this outlying case, AI-CS-CINE outperformed C-CINE, making the metrics derived from AI-CS-CINE more reliable. CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted compressed sensing CINE.

Some patients and volunteers encountered difficulties maintaining expiratory breath holds during both C-CINE and AI-CS-CINE image acquisition, leading to blurring of images. Given the longer acquisition time of C-CINE compared to AI-CS-CINE, C-CINE image quality was more adversely affected. Additionally, some patients experienced intermittent arrhythmias, resulting in improper ECG triggering image acquisition. Similar to issues with breath holding, poor ECG triggering had a more pronounced impact on image quality with C-CINE compared to AI-CS-CINE due to the longer acquisition time (Figure 5).

Amyloidosis patients vs non-amyloidosis participants

Patients with significant cardiac amyloidosis often experience more symptoms than other cardiac patients due to the unique pathophysiology of the disease; they require a more comprehensive CMR protocol, making CMR challenging for some in this subgroup. Key metrics, including EF, EDVi, ESVi, SVi (for both LV and RV), LVMi, and relative wall thickness, showed high correlation between C-CINE and AI-CS-CINE acquisitions (Table 4; Figure 6) in the subgroup of cardiac amyloidosis. LV measurements demonstrated excellent reliability (ICC_abs = 0.91-0.98), while RV measurements showed moderate-to-good reliability (ICC_abs = 0.65-0.86) in the subgroup of cardiac amyloidosis.

Figure 6 Conventional CINE and artificial-intelligence-assisted compressed sensing CINE still show good correlation in volume and function metrics in left and right ventricle in the subgroup of cardiac amyloidosis (n = 31). The X-axis represents the mean of conventional CINE (C-CINE) and artificial-intelligence-assisted compressed sensing CINE (AI-CS-CINE) values, while the Y-axis shows the difference between C-CINE and AI-CS-CINE. Dashed lines indicate the mean difference and 95% limits of agreement. BSA: Body surface area; LV: Left ventricle; RV: Right ventricle.

Table 4 Alignment of left ventricle and right ventricle volume and function metrics with Conventional CINE vs Artificial-intelligence-assisted compressed sensing CINE in patients with cardiac amyloidosis (n = 31).

	mean ± SD		Bland–Altman		Intraclass coefficient
	C-CINE	AI-CS-CINE	Bias ± SD	LoA	r	95%CI
LVEF (%)	55.9 ± 13.2	57.0 ± 13.2	-1.05 ± 4.56	-10.00, 7.89	0.94	0.87, 0.97
LVEDVi (mL/m2)	75.1 ± 21.1	72.0 ± 20.2	3.16 ± 7.07	10.70, 17.02	0.93	0.85, 0.97
LVESVi (mL/m2)	34.8 ± 19.0	32.2 ± 17.2	2.55 ± 6.12	-9.45, 14.55	0.93	0.86, 0.97
LVSVi (mL/m2)	40.4 ± 10.2	39.8 ± 11.1	0.61 ± 4.52	-8.26, 9.47	0.91	0.82, 0.95
LVMi (mL/m2)	69.9 ± 29.0	70.4 ± 29.5	-0.42 ± 5.81	11.81, 10.97	0.98	0.96, 0.99
Relative wall thickness	0.5 ± 0.2	0.5 ± 0.2	0.00 ± 0.06	-0.13, 0.12	0.96	0.91, 0.98
RVEF (%)	48.9 ± 14.6	52.1 ± 13.0	-3.14 ± 9.09	20.96, 14.68	0.76	0.56, 0.88
RVEDVi (mL/m2)	70.8 ± 16.2	70.9 ± 15.1	-0.02 ± 8.72	17.12, 17.07	0.84	0.70, 0.92
RVESVi (mL/m2)	37.1 ± 16.5	34.5 ± 13.5	2.64 ± 7.65	12.36, 17.64	0.86	0.72, 0.93
RVSVi (mL/m2)	33.7 ± 10.7	36.4 ± 10.5	-2.66 ± 8.57	19.46, 14.13	0.65	0.40, 0.81

C-CINE: Conventional CINE; AI-CS-CINE: Artificial-intelligence-assisted-CINE; EDV: End-diastolic volume; EF: Ejection fraction; ESV: End-systolic volume; LV: Left ventricle; LVM: LV mass; RV: Right ventricle; SV: Stroke volume; I: Normalized to body surface area.

DISCUSSION

Cardiovascular diseases are the leading global cause of death, with rising trends across all income levels[24]. CMR plays a crucial role in assessing ventricular volume, function, and myocardial health, as well as monitoring disease progression and treatment responses. However, a major barrier to expanding CMR access worldwide is the complexity of CMR workflows and the shortage of experienced technologists. AI-CS-CINE represents a significant step forward in overcoming these challenges. Our results show that AI-CS-CINE enhances CMR workflow by accelerating image acquisition, preserving anatomical details, and reducing motion artifacts compared to C-CINE, even in patients with severe cardiac amyloidosis, which is a particularly challenging group for CMR.

Cardiovascular disease patients often experience mild dyspnea, tachypnea, arrhythmia, and restlessness, making it difficult for technologists to shorten study duration and minimize breath holds. Significant efforts have been made to address these challenges, including the use of accelerated MR sequences and advanced reconstruction algorithms[13,14,25]. In this study, AI-CS-CINE consistently reduced image acquisition and reconstruction time, while minimizing the need for repeat scans due to poor image quality. This improvement enhanced CMR workflow efficiency, improved image quality, and increased diagnostic reliability. This was notably beneficial for patients with breath-hold instability, often eliminating the need for repeat acquisitions and ensuring more efficient, high-quality imaging.

AI-CS-CINE affords several practical advantages over C-CINE regarding improving the workflow of CMR procedures. First, AI-CS-CINE enables the efficient completion of functional CMR imaging within 5 minutes, using either breath-hold or rhythmic-breathing techniques. This expedited approach allows for seamless progression to subsequent tasks, such as strain analysis and native T1/T2 mapping. As a result, the entire CMR acquisition, including functional imaging, strain analysis, T1/T2 mapping, and late gadolinium enhancement, can be completed within 30 minutes. Second, multiple breath holds can lead to premature exhaustion of the patient in the early half of the study that contributes to suboptimal breath holds later during T1/T2 mapping and late gadolinium enhancement. Sometimes, patient tolerance is exceeded and premature termination ensues. AI-CS-CINE efficiently reduces the number and duration of breath holds and recently, images acquired with a continuous steady rhythmic breathing approach have been well tolerated and very successful. Third, AI-CS-CINE reduces the dependency on technologists’ experience in performing CMR. Whereas C-CINE requires technologists to adjust patient acquisition parameters based on breath-hold challenges and to continuously review image quality, AI-CS-CINE simplifies the scanning process, enabling technologists to easily complete the required scans with satisfactory image quality. Finally, AI-CINE reduces arrhythmia interference with the reduction of scan time to 1-2 seconds. The optimized implementation of AI-CS-CINE promises significant enhancement to CMR workflow, leading to improved image quality, increased study throughput, and easier patient compliance.

AI-CS-CINE reconstruction uses information from adjacent phases, which could introduce temporal blurring[26,27]. However, AI-CS-CINE images showed only slight reduction in sharpness compared to C-CINE, while maintaining overall image quality. The current analysis software, which automatically quantifies cardiac volume and function, was optimized for C-CINE. It is expected that other advanced models trained with AI-CS-CINE data will further improve the reliability of quantitative metrics.

AI-CS-CINE did not always outperform C-CINE on certain metrics, potentially due to variations in field strength, imaging sequences, technologist expertise, reconstruction algorithms, and image analysis software in previous studies[14,25,28]. In the present study, AI-CS-CINE provided superior image quality with very reliable results. When C-CINE image quality was compromised, AI-CS-CINE images salvaged otherwise non-diagnostic studies. A recent report using an AI tool to reconstruct low-resolution C-CINE images reduced acquisition time by 42%. However, that estimate did not include AI-enhancement during image acquisition[29]. In contrast, our study used AI for both acquisition and reconstruction, allowing technologists to improve workflow with image quality assessments in real-time and rapid repeat scans, if required.

Although multipurpose 3T scanners have been increasingly used for CMR, the 1.5-T scanner still remains the preferred choice for CMR due to its established performance and practical considerations[30,31]. Many hospitals worldwide have limited access to 3T scanners or only 1.5T scanners. CMR at 1.5T provides higher field homogeneity, fewer air-tissue interface artifacts, lower radiofrequency power deposition, broader availability, and comparable signal-to-noise ratios vs 3T. 1.5T MRI allows for more flexible imaging protocols with fewer SAR-related constraints. Additionally, SSFP sequences, which are widely used in CMR, suffer from increased banding artifacts at 3T, making 1.5T more reliable for CINE imaging. Furthermore, MRI-compatible medical devices, such as pacemakers and implants, are more commonly tested and approved for 1.5T, reinforcing its role as the clinical standard for CMR[32]. In comparison with two recent studies[14,29], the current study is more representative in terms of using 1.5T to better reflect CMR workflow worldwide, enabling AI to facilitate both image acquisition and inline reconstruction and encompassing clinical referrals of a broader age range, diverse body habitus, and a wider racial and disease spectrum.

LV volumes and functions with AI-CS-CINE agreed with C-CINE more closely than for the RV, which in part reflected the use of LV short-axis image stacks and system auto-calculation analysis software developed with C-CINE images. Specifically, the anatomy of the RV is more complicated than the LV with thinner walls, prominent trabeculations, and a complex LV wrap-around shape. These RV anatomical features complicate the software endocardial delineation, leading to potential errors. In addition, RV is more effected by diaphragm-mediated respiratory motion. Finally, the RV is more sensitive to field inhomogeneities arising from air-tissue interfaces and variations in tissue susceptibility. These inhomogeneities can distort the magnetic field, affecting image quality and contrast[33,34].

This study design has generalization limitations. The generalizability of our findings is limited by both the study population and imaging setup. The cohort included a high proportion of healthy volunteers and cardio-oncology patients, with limited representation of other cardiac conditions. All imaging was performed on a single 1.5-T scanner from one vendor. Future studies are warranted across broader patient populations, diverse disease types, and multicenter imaging platforms. Additional efforts should include protocol refinements to improve RV visualization and quantification, development of more advanced AI reconstruction algorithms trained on racially diverse datasets, and implementation of blinded image analysis to further validate these findings. In addition, this study focused on ventricular metrics, but future studies should include atrial metrics and feature-tracking strain analyses. Another limitation of this study is the lack of formal evaluation of patient comfort or satisfaction. While not systematically assessed, informal feedback indicated that rhythmic breathing during AI-CS-CINE was generally easier and better tolerated than breath holding in C-CINE.

CONCLUSION

AI-CS-CINE significantly enhances the CMR workflow by accelerating image acquisition and reconstruction, reducing motion artifacts, and minimizing the need for repeat scans. It preserves anatomical details while improving efficiency, particularly in challenging patient populations. The quantitative cardiac metrics from AI-CS-CINE closely align with C-CINE across diverse ventricular sizes and functions, including cardio-oncology referrals and healthy volunteers. By streamlining the imaging process without compromising diagnostic accuracy, AI-CS-CINE represents a major advancement in CMR, making high-quality CMR more efficient and accessible.

ACKNOWLEDGEMENTS

We would like to thank Elena Deych, MS, for reviewing the statistical data and offering valuable suggestions.