Comparing HeartModel^AI and cardiac magnetic resonance imaging for left ventricular volume and function evaluation in patients with dilated cardiomyopathy

Abstract

Background

Integration of artificial intelligence enhances precision, yielding dependable evaluations of left ventricular volumes and ejection fraction despite image quality variations. Commercial software like HeartModel^AI provides fully automated 3DE quantification, simplifying the measurement of left chamber volumes and ejection fraction. In this manuscript, we present a cross-sectional study to assess and compare the diagnostic accuracy of automated 3D echocardiography (HeartModel^AI) to the standard Cardiac Magnetic Resonance Imaging in patients with dilated cardiomyopathy.

Methods

In this cross-sectional study, 30 patients with dilated cardiomyopathy referring to the Tehran Heart Center with cardiac magnetic resonance imaging and comprehensive 3D transthoracic echocardiography within 24 h were included. All 3D volume analysis was performed with fully automated quantification software (HeartModel^AI) using 3D images of 2,3, and 4-chamber views at the end of systole and diastole.

Results

Excellent Inter- and Intra-observer correlation coefficient was reported for HeartModel^AI software for all indexes. HeartModel^AI displayed a remarkable correlation with cardiac magnetic resonance for left ventricular end-systolic volume index (r = 0.918 and r = 0.911); nevertheless, it underestimated left ventricular end-systolic volume index and left ventricular end-diastolic volume index. Conversely, ejection fraction, stroke volume, and left ventricular mass were overestimated. It was found that manual contour correction can enhance the accuracy of automated model estimations, particularly concerning EF in participants needing correction.

Conclusion

HeartModel^AI software emerges as a rapid and viable imaging approach for evaluating the left ventricle’s structure and function. In our study, LV volumes assessed by HeartModel^AI demonstrated strong correlations with cardiac magnetic resonance imaging.

Introduction

Accurately measuring the volumes and ejection fraction (EF) of the left ventricle (LV) is crucial for diagnosing and predicting cardiac outcomes [1]. While cardiac magnetic resonance (CMR) imaging is considered the gold standard for this assessment [2], its use is limited due to accessibility challenges, cost, and contraindications in certain patients [3].

Transthoracic echocardiography is the primary method for LV assessment. Traditional two-dimensional echocardiography relies on the biplane Simpson formula, which has geometric assumptions and foreshortening limitations. In contrast, three-dimensional echocardiography provides direct evaluation without anatomical and wall thickening distribution assumptions, proving more accurate and reliable [4]; however, due to the high level of expertise needed and the time-consuming process of obtaining good-quality 3D images, its use has not been widely implemented.

HeartModel^AI (HM) is an artificial intelligence algorithm that uses adaptive analysis to automatically detect cardiac borders using three-dimensional echocardiography images, facilitating quick and accurate measurements of heart chambers [5]. The introduction of HM overcomes the laborious task of border delineation in 3D echocardiography [6] and deems a better diagnostic assessment of the heart [7]. Although HM measurements show promising correlations with CMR and minimal observer variability [8], occasional contour correction is needed, particularly in distinguishing endocardial borders.

HM uses a database comprised of a wide variety of populations as the training sample; however, studies have reported low ejection fraction to affect the capability of HM to detect cardiac borders correctly [9], and since dilated cardiomyopathy (DCM) patients have low ejection fraction with dilated LV volumes, ensuring diagnostic accuracy of HM would significantly impact their clinical care. Establishing the accuracy of diagnosis and agreement of HM with CMR would enable reliable, discrete measurements with reduced inter-observer variability.

Therefore, this study aims to assess the agreement between fully automated three-dimensional transthoracic echocardiography (HeartModel^AI) with and without manual editing compared to CMR imaging in DCM patients. Additionally, the study will evaluate the reliability and reproducibility of these methods.

Methods

Study design

This cross-sectional study included DCM patients referred to the Tehran Heart Center Hospital, Tehran, Iran, between March 2022 and January 2024. All procedures were conducted by the 1964 Declaration of Helsinki and its later extensions. Written informed consent was obtained. The ethics committee of Tehran University of Medical Sciences (TUMS) reviewed and approved this study (IR.TUMS.THC.REC.1402.065).

Inclusion and exclusion criteria

All participants who underwent cardiac magnetic resonance (CMR) imaging and comprehensive 3D transthoracic echocardiography (TTE) within 24 h and a comparable hemodynamic status were included. DCM diagnosis was based on clinical and paraclinical evaluations, characterized by a significantly reduced left ventricular ejection fraction (LVEF < 35%) and significant left ventricular enlargement (end-diastolic diameter exceeding 117% or > 2 standard deviations above the predicted value of 112% adjusted for age and body surface area) by two-dimensional echocardiography in the absence of underlying valvular or ischemic heart diseases [10, 11].

Patients with congenital heart disease, contraindication for magnetic resonance imaging (MRI), and poor visualization on 2D echocardiography (defined as > 3 consecutive unspecific endocardial contours) were excluded.

Data Collection

Patient demographic data, including age, sex, body mass index (BMI), and body surface area (BSA), were collected using a data collection form. Left ventricular volume and functional measurements, including left ventricular end-systolic volume index (LVESVI), left ventricular end-diastolic volume index (LVEDVI), left ventricular ejection fraction (LVEF), left ventricular mass (LV mass), LV mass index, and stroke volume (SV) were measured with HeartModel^AI and CMR.

Echocardiography

3D TTE was performed using EPIQ 7 with an X5 phased array transducer (Phillips Medical Systems., Andover, MA, USA). The patients were in sinus rhythm and positioned in the left lateral decubitus position. Wide-angled acquisition was implemented using “full-volume” mode, covering four consecutive cardiac cycles during a single breath-hold. The optimal frame rate was achieved by imaging depth and sector width minimization.

The acquired digital 3D volumes were archived for analysis utilizing fully automated quantification software (HeartModel^AI). HeartModel^AI processes 3D images obtained from 2, 3, and 4-chamber views at both end-systole and end-diastole. Upon activation, HeartModel^AI identifies the heart chambers and establishes the end-diastolic and end-systolic condition through analysis of cardiac motion, subsequently generating 3D volumes automatically.

Regarding the positioning of the LV volume border, users could choose to place it near the blood-endocardial interface or deeper within the compacted myocardium, depending on their preference. The software’s algorithm detected both inner and outer myocardial borders during end-diastole and end-systole. As per vendor recommendations, the default settings for 3D TTE were 60 in end-diastole and 30 in end-systole. In cases necessitating contour correction, semi-automatic analysis was conducted, integrating any applied corrections.

Cardiac magnetic resonance (CMR) imaging

CMR imaging was undertaken with a 1.5 tesla (T) MRI system (Philips Medical Systems, The Netherlands) set at a maximum gradient strength of 45 mT/m and a maximum slew rate of 200 mT/m/ms. The system received a signal with a digital interface from a 32-channel torso coil.

Cardiac function assessment employed a balanced steady-state free-precession (bSSFP) pulse sequence. The bSSFP setting included an acquired pixel size and a reconstructed pixel slice of 1.8 × 2.3 mm and 1.5 × 1.5 mm, respectively. Moreover, a slice thickness of 8 mm at the end of the expiration. Cine imaging used the bSSFP pulse sequence with parameters set as follows: Repetition time = 3.5 ms, Time to echo = 1.4 ms, flip angle = 55 degrees, and temporal resolution of ≤ 45 ms.

Subsequently, an independent specialist processed CMR data offline with Philips extended MR workspace software (Philips Medical Systems, The Netherlands). The specialist conducting CMR data analysis was blinded to the results of the 3D TTE. LVEF, LVEDVI, LVESVI, LV mass, and stroke volume were calculated using the standard formula on the cine images. CMR volume analyses were performed with the inclusion of papillary muscles as the blood pool.

Inter- and intra-observer variability

The echocardiography evaluation was performed by two independent echocardiography specialists while blinded to each other’s results within less than 24 h. Intra-observer variability was assessed following an interval of one week, based on the recorded images in the echocardiographic system,

Statistical analysis

Baseline measurements of participants base on CMR, HeartModel^AI with and without contour correction are presented as mean and standard deviation (SD). In order to compare the measurements of, HeartModel^AI before and after contour correction, paired t-test was used. we took advantage of coefficient of determination and Pearson correlation coefficients to measurement agreement between HeartModel^AI and CMR and compare them with Fisher r-to-z transformation. Also, to visualize the degree of agreement between the two measurement and identify any systematic bias, Bland-Altman plot was drawn. In addition to evaluation the reliability and reproducibility, we report Inter- and Intra-observer correlation coefficients with 95% CI. Analyses were conducted using the R Statistical language (version 4.3.0; R Core Team, 2023).

Result

Study cohort

Thirty participants were included in the final evaluation. The mean age of the patients was 50.63 ± 12.47 years, with males comprising 21 (70.0%) cases. The mean BMI is 27.46 ± 4.64 Kg/m², and the median BSA was 1.9 (1.7, 2.0). Diabetes mellitus and hypertension were observed in 3 (10%) and 4 (13.3%) cases, respectively. Table 1 presents patients’ baseline characteristics.

Table 1.

Baseline characteristics among participants

Characteristic	Levels	Values
Age, years		50.63 ± 12.46
Sex	Male	21 (70.0)
	Female	9 (30.0)
Body Surface Area, m2		1.88 ± 0.25
Body Mass Index, Kg/m2		27.45 ± 4.63
Diabetes Mellitus		3 (10.0)
Smoking		4 (13.3)
Hypertension		4 (13.3)

Quantitative data are presented as mean with standard deviation, Qualitative data are presented as frequency with percentages

Measurement agreement

Table 2 presents the left ventricular mean volume and functional measurements from HeartModel^AI with and without contour correction, as well as CMR imaging.

Table 2.

Baseline measurements of participants

Characteristic	HeartModelAI without contour correction	HeartModelAI with contour correction	CMR
	N = 30#	N = 30#	N = 30#
LVESVI	87.73 ± 42.13	87.60 ± 43.70	104.83 ± 55.81
LVEDVI	124.47 ± 45.21	119.60 ± 47.44	138.20 ± 55.18
LVEF	31.93 ± 9.96	29.29 ± 9.86	27.47 ± 10.57
LV Mass	159.67 ± 52.44	160.27 ± 52.62	117.90 ± 41.62
LV Mass Index	85.04 ± 28.77	84.98 ± 27.43	61.65 ± 17.72
SV	66.63 ± 22.47	60.97 ± 21.73	63.87 ± 17.51

^#Mean ± SD

CMR Cardiac Magnetic Resonance, LVESVI Left Ventricular End-Systolic Volume Index, LVEDVI Left Ventricular End-Diastolic Volume Index, LVEF Left Ventricular Ejection Fraction, SV Stroke Volume

HeartModel^AI underestimated the mean value of LVESVI (−17.10 ± 23.90, p = 0.001) and LVESDI (−13.73 ± 23.35, p = 0.005) compared to the CMR. On the contrary, LVEF (4.46 ± 7.35, p < 0.001), LV Mass (41.76 ± 48.06, p < 0.001), LV Mass Index (23.39 ± 26.69, p < 0.001), and SV (2.76 ± 13.16, p < 0.001) were overestimated.

Contour correction was needed in 8 cases (26.6%). Comparing edited HeartModel^AI with CMR imaging showed an underestimation of LVESVI (−17.23 ± 22.05, p < 0.001), LVEDVI (−18.60 ± 23.61, p < 0.001), and SV (−2.90 ± 13.82, p = 0.168) compared to the CMR. In contrast, LVEF (1.82 ± 5.32, p < 0.001), LV Mass (42.36 ± 42.68, p < 0.001), and LV Mass Index (23.33 ± 23.80, p < 0.001) were overestimated. Table 3 demonstrates the average of corrections needed for each index. In cases needing contour editing, contour correction did not yield any statistically significant difference except for LVEF (p = 0.007) and SV (p = 0.004).

Table 3.

Comparing differences between fully automated and semiautomated HeartModel^AI

Characteristic	Mean Difference	95% CI	P value
LVESVI	0.50	(−15.87, 16.87)	0.944
LVEDVI	18.25	(−2.33, 38.83)	0.074
LVEF	9.875	(3.72, 16.02)	0.007
LV Mass	−2.25	(−44.73, 40.23)	0.904
LV Mass Index	0.233	(−20.54, 21.02)	0.980
SV	21.25	(9.37, 33.13)	0.004

LVESVI Left Ventricular End-Systolic Volume Index, LVEDVI Left Ventricular End-Diastolic Volume Index, LVEF Left Ventricular Ejection Fraction, SV Stroke Volume

Correlation analysis demonstrated significant agreement between HeartModel^AI and CMR measurements for all indexes. Fully automated measurements exhibited excellent agreement for LVESVI (r = 0.918) and LVEDVI (r = 0.911), good correlation for SV (r = 0.811), and acceptable agreement for LVEF (r = 0.744). However, despite being significantly correlated, HeartModel^AI did not reveal acceptable agreement with CMR for LV Mass (r = 0.498) and LV mass index (r = 0.421). Fig. 1 illustrates the agreement of CMR and HeartModel^AI without contour correction.

Fig. 1.

Linear Correlation and Bland-Altman analysis for left heart volume and function between HeartModel^AI without contour correction with CMR

HeartModel^AI with the contour corrections revealed improved correlation for all indexes, except for LVEDVI (r = 0.905) and SV (r = 0.772). Nevertheless, no significant difference was observed between the CMR agreement correlations with automated and edited HeartModel^AI measurements. Fig. 2 exhibits the agreement of CMR and HeartModel^AI with contour corrections. Table 4 illustrates Pearson’s Correlation Coefficients for the agreement of CMR with automated and edited HeartModel^AI measurements.

Fig. 2.

Linear correlation and Bland-Altman analysis for left heart volume and function between HeartModel^AI with contour correction with CMR

Table 4.

Measurement agreement between HeartModel^AI and CMR

Characteristic	Pearson Correlation Coefficients			Z	p-value#
	HM&CMR	HME & CMR
LVESVI	0.918	0.93	0.305		0.760
LVEDVI	0.911	0.905	0.118		0.906
LVEF	0.744	0.866	1.315		0.188
LV Mass	0.498	0.612	0.608		0.543
LV Mass Index	0.421	0.514	0.440		0.660
SV	0.811	0.772	0.383		0.701

^#Fisher r-to-z transformation

HM HeartModel^AI, CMR Cardiac Magnetic Resonance, HME HeartModel^AI Edited, LVESVI Left Ventricular End-Systolic Volume Index, LVEDVI Left Ventricular End-Diastolic Volume Index, LVEF Left Ventricular Ejection Fraction, SV Stroke Volume

Reliability and reproducibility

Excellent reliability and reproducibility were observed, with less than 10% variability for all indexes. Table 5; Fig. 3 demonstrate the inter- and intra-observer correlation coefficients.

Table 5.

Inter- and intra-observer correlation coefficients

Characteristic	Interclass		Intraclass
	Correlation	95% CI	Correlation	95% CI
LVESVI	0.980	(0.960, 0.991)	0.992	(0.983, 0.996)
LVEDVI	0.977	(0.952, 0.989)	0.989	(0.977, 0.995)
LVEF	0.980	(0.959, 0.990)	0.930	(0.860, 0.966)
LV Mass	0.923	(0.846, 0.963)	0.941	(0.881, 0.971)
LV Mass Index	0.931	(0.862, 0.966)	0.928	(0.855, 0.965)
SV	0.929	(0.858, 0.966)	0.943	(0.884, 0.972)

LVESVI Left Ventricular End-Systolic Volume Index, LVEDVI Left Ventricular End-Diastolic Volume Index, LVEF Left Ventricular Ejection Fraction, SV Stroke Volume

Fig. 3.

Inter-observer and Intra-observer correlation coefficient comparison for left ventricular volume and function

Discussion

Our study investigates the efficacy of HeartModel^AI in assessing echocardiographic indices in patients with DCM. It demonstrates that the HeartModel^AI algorithm serves as a dependable assessment tool in this context. It exhibits a strong correlation in evaluating LV function with minimal variability between discrete measurements. Fig. 4 depicts HeartModel^AI models.

Fig. 4.

HeartModel^AI illustrations

Accurate assessment of cardiac function is pivotal for diagnosing and monitoring patients with DCM. Various cardiovascular imaging modalities have been introduced for this purpose, with CMR imaging as the gold standard [12]. However, its widespread application is hindered by challenges such as accessibility, the requirement for highly trained experts, and high costs [8]. TTE has emerged as the preferred initial assessment method due to its safety, relative simplicity, and efficiency [5]. Nonetheless, conventional 2D echocardiography suffers from inherent limitations, such as reliance on geometric assumptions [6]. The advent of 3D TTE in the early 2000s has enhanced diagnostic accuracy by overcoming the need for anatomical assumptions in cardiac function measurements [4, 13]. However, it remains time-consuming, demands significant expertise, and offers lower temporal resolution.

HeartModel^AI is an artificial intelligence-integrated software that employs an adaptive analytic algorithm, drawing from a vast library of echocardiography images, to assess endocardial and epicardial borders and measure echocardiographic indices [14]. With the HeartModel^AI, the software automates border assessment with a single button press, thus eliminating the time and expertise required for manual endocardial border delineation while still allowing for manual correction of regional or global contours [15, 16].

Previous studies have noted that 3D TTE tends to underestimate LVESV and LVEDV compared to CMR in numerous instances [17, 18]. Our findings align with these trends, as Tsang et al. [14] reported HeartModel^AI underestimating LVEDV while yielding similar results for LVESV. Similarly, Levy et al. [2] and Tamborini [1] reported underestimations in LVESV and LVEDV, although comparable results were observed for LVEF. Volpato et al. [5] also reported underestimations in LVESV and LVEDV, coupled with an overestimated LVEF. Furthermore, Barletta et al. [19] reported excellent correlation of automated HeartModel^AI and CMR. Consistent with prior research, our study demonstrates underestimations in LVESVI and LVEDVI and an overestimation of LVEF with HeartModel^AI compared to CMR, despite excellent correlations.

Regarding assessing left ventricular mass and left ventricular mass index, Volpato et al. [5] found no significant difference between HeartModel^AI and CMR. Our study observed an overestimation of both parameters compared to CMR measurements, with limited correlations. This discrepancy could be attributed to the low spatial resolution resulting from challenges in definitive endocardial and epicardial border delineation, particularly in dilated LV with difficulty distinguishing compact myocardium and LV trabeculae.

Regarding the need for manual contour correction, Tamborini et al. [1] reported favorable performance of HeartModel^AI in DCM patients, albeit often requiring contour correction. In Chen Ke-Pan et al. [6] study, contour correction was necessary in 42% of DCM patients while leading to favorable results. In our study, contour correction had to be done in 26% of cases. Although HeartModel^AI is quite good at automating the assessment process, the need for manual adjustments in a sizable subset does add time costs. Indeed, contour correction is an important means of increasing measurement accuracy, which is necessary in some patients, as pointed out by previous studies. For example, Ke-Pan et al. [6] reported significant improvements in diagnostic accuracy with manual correction, while our research found the improvements generally favorable but not statistically significant across the board. The major benefit of using automated 3DE is that it reduces both the time and expertise necessary for accurate assessment. To some extent, the need for manual correction in almost a quarter of cases potentially limits the time-saving capability, especially in a busy clinical setting. Further development in the algorithm of HeartModel^AI, reducing the need for manual adjustment, is helpful in fully realizing the intended efficiency benefits of this technology.

Evaluation of the accuracy of HeartModel^AI in patients with LV dysfunction has been mixed, but it holds significant potential. Beitner et al. [20] evaluated the agreement of 3DE with CMR imaging to estimate left ventricular volumes and LVEF in post-MI patients. The ICC for the LVEF was quite good between 3DE and CMR, indicating that HeartModel^AI could be a viable option in the field. However, for LVESV and LVEDV, the strength of association was comparatively weaker, with ICC values of 0.44 and 0.28, respectively, indicative of moderate and fair agreement. Automated 3DE showed good concordance with CMR measurement in a study by Wang et al. [21] for 53 patients with hypertrophic cardiomyopathy (HCM). While the initial correlations of the two modalities, before contour corrections, were suboptimal, the incorporation of manual contour adjustments significantly improved the agreement, indicating the crucial role of manual refinement in enhancing diagnostic accuracy. Once these discrepancies were corrected, the study confirmed that automated 3DE would work feasibly and reliably compared to manual 3DE in this patient population. Naser et al. [13] further evaluated performance in patients with NYHA I-III status LV dysfunction. In this study, 38 DCM patients were compared to 2DE. In the 166 subjects involved, no significant difference between 2DE and 3DE regarding LVEF and LV volume analyses could be found, reinforcing the reliability of HeartModel^AI. These findings further support HeartModel^AI clinical value as a reliable alternative for evaluating left ventricular function.

Consistent with prior studies, our research demonstrated excellent inter- and intra-observer correlation coefficients with less than 10% variability, affirming the reliability and reproducibility of HeartModel^AI in DCM patients.

Clinical implications

Our study’s novelty lies in evaluating automated 3D TTE in DCM patients. Low EF and dilated heart chambers, characteristic of DCM, have previously been reported to affect the diagnostic capability of 3D TTE. Therefore, validating HeartModel^AI for DCM assessment could significantly enhance clinical care by improving diagnosis and evaluation in these patients.

Limitations

While our study contributes to the limited body of research specifically assessing HeartModel^AI 3D TTE algorithm accuracy in DCM patients and utilizes adjusted values for BSA, several limitations must be acknowledged. Small sample size remains a primary constraint. Additionally, the use of high-quality images acquired by echocardiography experts with over 10 years of experience restricts the generalizability of our findings. Furthermore, exclusion criteria such as arrhythmia and pacemakers limit the applicability of HeartModel^AI in real-world scenarios. The use of multiple-beat breath-hold image acquisition could introduce artifacts, impacting results. The predominantly male participant cohort may also influence observations. Reproducibility assessment solely on saved images rather than test-retest analysis incorporating new image acquisition poses a limitation. Despite implementing wide sector minimization to improve temporal resolution, increased wide sector data acquisition in dilated cardiomyopathy patients may affect HeartModel^AI capability. More comprehensive multicenter studies including more patients, with image acquisition by more echocardiography experts, could yield more decisive results.

Conclusion

HeartModel^AI software emerges as a rapid and viable imaging approach for evaluating LV structure and function. Its high reproducibility and reliability make HeartModel^AI a promising tool for DCM patients’ evaluation.

Abbreviations

BMI

Body Mass Index

BSA

Body Surface Area

bSSFP

balanced steady-state free-precession

CMR

Cardiac Magnetic Resonance

DCM

Dilated Cardiomyopathy

EF

Ejection Fraction

HCM

Hypertrophic Cardiomyopathy

HM

HeartModel^AI

LVEDVI

Left Volume End-diastolic Volume Index

LVESVI

Left Volume End-systolic Volume Index

MRI

Magnetic Resonance Imaging

SV

Stroke Volume

TTE

Transthoracic Echocardiography

Authors’ contributions

M.S: Performed Data Curation and Wrote the Main manuscript text; S.F.A.: Supervised and administered the project; M.M: Wrote the main manuscript text; A.J., and A.V.B.: Performed Formal Analysis and Visualization; M.S: Conceptualized the project; H.A: Devised the study methodology; A.Z: Reviewed and Edited the manuscript text. All authors reviewed the manuscript.

Funding

This study did not receive any funds or grants.

Data availability

Data is available upon reasonable request from the corresponding author and will be provided after anonymization.

Declarations

Ethics approval and consent to participate

All procedures were conducted by the 1964 Declaration of Helsinki and its later extensions. Written informed consent was obtained for participation and publication. The ethics committee of Tehran University of Medical Sciences (TUMS) reviewed and approved this study (IR.TUMS.THC.REC.1402.065).

Consent for publication

Written informed consent was obtained for participation and publication.

Competing interests

The authors declare no competing interests.