Machine Learning–Based Three-Dimensional Echocardiographic Quantification of Right Ventricular Size and Function: Validation Against Cardiac Magnetic Resonance

Machine Learning–Based Three-Dimensional Echocardiographic Quantification of Right Ventricular Size and Function: Validation Against Cardiac Magnetic Resonance Davide Genovese, MD, Nina Rashedi, MD, Lynn Weinert, RDCS, Akhil Narang, MD, Karima Addetia, MD, Amit R. Patel, MD, David Prater, MS, Alexandra Gonc¸alves, MD, PhD, Victor Mor-Avi, PhD, and Roberto M. Lang, MD, Chicago, Illinois; Padua, Italy; and Andover, Massachusetts

Background: Three-dimensional echocardiography (3DE) allows accurate and reproducible measurements of right ventricular (RV) size and function. However, widespread implementation of 3DE in routine clinical practice is limited because the existing software packages are relatively time-consuming and skill demanding. The aim of this study was to test the accuracy and reproducibility of new machine learning– (ML-) based, fully automated software for three-dimensional quantification of RV size and function. Methods: Fifty-six unselected patients with a wide range of RV size and function and image quality, referred for clinically indicated cardiac magnetic resonance (CMR) imaging, underwent a transthoracic 3DE exam on the same day. End-systolic and end-diastolic RV volumes (ESV, EDV) and ejection fraction (EF) were measured using the ML-based algorithm and compared with CMR reference values using Bland-Altman and linear regression analyses. Results: RV function quantification by echocardiography was feasible in all patients. The automatic approach was accurate in 32% patients with analysis time of 15 6 1 seconds and 100% reproducible. Endocardial contour editing was necessary after the automated postprocessing in the remaining 68% patients, prolonging analysis time to 114 6 71 seconds. With these minimal adjustments, RV volumes and EF measurements were accurate in comparison with CMR reference (biases: EDV, 25.6 6 21.1 mL; ESV, 7.4 6 16 mL; EF, 3.3% 6 5.2%) and showed excellent reproducibility reflected by coefficients of variation <7% and intraclass correlations $0.95 for all measurements. Conclusions: The new ML-based 3DE algorithm provided accurate and completely reproducible RV volume and EF measurements in one-third of unselected patients without any boundary editing. In the remaining patients, quick minimal editing resulted in reasonably accurate measurements with excellent reproducibility. This approach provides a promising solution for fast three-dimensional quantification of RV size and function. (J Am Soc Echocardiogr 2019;-:---.) Keywords: Right ventricle, Right ventricular volume and ejection fraction, Three-dimensional echocardiography, Diagnostic techniques, Machine learning, Artificial intelligence

From the University of Chicago Medical Center (D.G., N.R., L.W., A.N., K.A., A.R.P., V.M-A., R.M.L.), Chicago, Illinois; Department of Cardiac, Thoracic and Vascular Sciences, University of Padua (D.G.), Padua, Italy; and Philips Healthcare (D.P., A.G.), Andover, Massachusetts. The study was supported by a research grant from Philips Healthcare, Andover, MA. D.P. and A.G. are full-time employees of Philips. Neil J. Weissman, MD, FASE, served as guest editor for this report. Reprint requests: Roberto M. Lang, MD, Section of Cardiology, University of Chicago Medical Center, 5758 South Maryland Avenue, MC 9067, Chicago, IL 60637 (E-mail: [email protected]). 0894-7317/$36.00 Copyright 2019 by the American Society of Echocardiography. https://doi.org/10.1016/j.echo.2019.04.001

Right ventricular (RV) size and function measurements are important for the diagnosis and prognostic evaluation of multiple cardiac diseases.1-5 The retrosternal location of the right ventricle, as well as its complex crescent shape, hampers two-dimensional echocardiography– (2DE-) based evaluation of this chamber. To circumvent these limitations, 2DE relies on the acquisition of multiple views from different transducer positions to quantify RV dimensions and function.6 Currently, cardiac magnetic resonance (CMR) is considered the gold standard for the assessment of RV volumes and ejection fraction (EF). However, this imaging modality is expensive, timeconsuming, not universally available, and not feasible in all patients. Three-dimensional echocardiography (3DE) has become an attractive alternative for RV quantification. This is because it avoids the drawbacks of 2DE, such as foreshortening and geometrical 1

2 Genovese et al

assumptions on the one hand and the cumbersome acquisition, 2DE = Two-dimensional postprocessing time, and high echocardiography costs of CMR on the other hand. Numerous studies have 3DE = Three-dimensional established the accuracy of 3DE echocardiography measurements of RV volumes 4Ch = Four-chamber and EF, compared with CMR reference,7-11 and have also CMR = Cardiovascular magnetic resonance demonstrated its additional diagnostic and prognostic value CoV = Coefficient of variation over conventional 2DE EDV = End-diastolic volume parameters.12-14 Due to these reasons, the recent American ESV = End-systolic volume Society of Echocardiography/ EF = Ejection fraction European Association of Cardiovascular Imaging chamber ICC = Intraclass correlation quantifications guidelines15 coefficient recommend using 3DE volume LV = Left ventricular, ventricle and EF measurements as the new standard for the assessment ML = Machine learning of RV size and function. RV = Right ventricular, However, this recommendation ventricle has not been fully incorporated into the clinical routine, because 3DE RV quantification remains relatively time-consuming and skill demanding, requiring specialized training and experience for both data acquisition and analysis. Recently, the use of artificial intelligence approaches, including machine learning (ML) algorithms,16,17 has enabled automated detection of left ventricular (LV) and left atrial endocardial boundaries throughout the cardiac cycle from 3DE data sets, allowing accurate measurements of LV and left atrial volumes and EF.18-21 Most recently, this ML technology has been further developed to allow automated detection of the RV endocardial borders. The aim of this initial validation study was to compare RV volumes and EF obtained by this new ML algorithm with CMR reference values. Abbreviations

METHODS Population and Study Protocol Fifty-six unselected patients with a wide range of RV sizes, function, and 3DE image quality were recruited from consecutive patients referred for a clinically indicated CMR exam who agreed to undergo transthoracic 3DE imaging on the same day. All patients were in normal sinus rhythm. RV volumes and EF were measured from 3DE images using the new ML-based automated software, with manual correction of endocardial boundaries when necessary, and compared with the corresponding CMR values obtained using standard methodology as a reference. The protocol was approved by the Institutional Review Board, and informed consent was obtained from each patient. CMR Acquisition and Analysis CMR was performed on a 1.5-T scanner (Philips Medical Systems, Best, Netherlands) with a five-channel cardiac coil. A steady-state freeprecision dynamic gradient-echo sequence was used to obtain cine loops, during approximately 5-second breath holds (repetition time, 2.9 msec; echo time, 1.5 msec; flip angle, 60 ; temporal resolution, 30-40 msec). In all patients, six to 10 short-axis slices were obtained

Journal of the American Society of Echocardiography - 2019

from the RV base to the apex (slice thickness, 6 mm; gap, 4 mm). These images were analyzed offline using commercial software (Medis, Leiden, Netherlands). In every slice, RV endocardial contours were manually traced at end diastole and end systole by an experienced operator blinded to the echocardiographic data. On the most basal slices, the right ventricle was differentiated from the right atrium by advancing the cine loop frame by frame throughout systole. If the cavity became smaller and the myocardium thicker, it was included within the RV volume. In contrast, when the cavity became larger and did not show wall thickening, it was considered as part of the atrium. Endocardial trabeculae were included within the RV cavity. Disk summation was used to calculate end-diastolic volume (EDV) and end-systolic volume (ESV), and EF was calculated using the standard formula: EF = (EDV ESV)/EDV. 3DE Imaging Echocardiographic imaging was performed from the apical transducer position using either an iE33 or EPIQ 7C system (Philips Healthcare, Andover, MA), equipped with an X5-1 phased-array transducer. Prior to acquisition, endocardial visualization was optimized by modifying the gain, compress, and time-gain compensation controls. During a single breath-hold, 3DE image acquisition included an apical four-chamber (4Ch) RV-focused data set, either wide-angled, single-beat, high frame rate (HeartModel mode on EPIQ system) or full-volume, electrocardiogram gated over a maximum of 6 beats (on IE33 system). Care was taken to include the entire RV cavity within the scan volume throughout the cardiac cycle. Imaging depth and sector width were optimized to obtain the highest possible frame rate. The 3DE data sets were stored digitally and used for offline analysis. 3DE Analysis The 3DE images were analyzed by an experienced echocardiographer blinded to CMR data. First, the quality of the data sets was systematically evaluated. Briefly, the multiplanar reconstruction tool from the three-dimensional (3D) Viewer software (QLAB, Philips Healthcare) was used to extract the RV basal short-axis, 4Ch, and inflow-outflow tract views. Each RV wall was divided into segments (Figure 1), and endocardial border visualization was scored as 0 (not visible), 1 (partially visible), or 2 (visible) in each of 17 endocardial segments defined as follows: 5 in the 2D RV basal short-axis view, 6 in the 4Ch, and 6 in the inflow-outflow tract views. The sum scores of all 17 segments determined the overall image quality, which thus potentially ranged from 0 to 34. A score >23 was arbitrarily defined as good quality, between 18 and 23 as fair, and <18 as poor. RV quantitative analysis was performed offline using the new ML approach (3D Auto RV, Philips Healthcare) on a commercial platform for data management (QLAB, Philips), installed on a standard personal computer. Briefly, two basic functions are performed to achieve the automatic segmentation: localization of the four cardiac chambers using a Hough transform and identifying the endocardial contours of the RV. The basis of the automatic segmentation process are deformable models. These models combine boundary detection with shape descriptions of the cardiac structures to optimally position the contours to separate the chambers from the surrounding tissue. Use of this combined approach produces a more robust estimate of the boundary position across a range of image quality and heart shapes. A large number of parameters control the deformable models, to optimize which a supervised ML approach was used. This approach takes as input a large number of clinical studies that

Journal of the American Society of Echocardiography Volume - Number -

HIGHLIGHTS  Machine-learning software for quantification of right ventricular function was tested.  Accurate, fully-automated measurements are possible in 32% of unselected patients.  With minimal editing, accurate measurements were obtained in the remaining patients.

Genovese et al 3

of the hand-traced contours, a consensus opinion from the multiple clinical experts was used. The parameters of the boundary detectors were adjusted to minimize the difference between the detected and ground truth contours. This study was designed to test the ability of this algorithm to perform a fully automated stepwise RV analysis to obtain EDV, ESV, and EF measurements on an independent exploratory set of echocardiographic images. Specifically, the software initially identified LV and RV long-axis landmarks in end-diastole in the apical two- and four-chamber views. Subsequently, the tricuspid valve annulus hinge

Figure 1 Each 3DE data set was assessed for image quality, which was defined as poor, fair, or good. The RV basal short axis (bottom left), 4Ch focused view (top left), and inflow-outflow tract view (top right) are obtained using the 2D multiplanar reconstruction. The RV walls in each view were divided into segments subsequently scored from 0 to 2 based on the visualization quality. A ap, Anterior apical wall; A bas, anterior basal wall; AL, anterolateral wall; A mid, anterior midwall; AS, anteroseptal wall; I, inferior wall; I ap, inferior apical wall; I bas, inferior basal wall; I mid, inferior midwall; IS, inferoseptal wall; L ap, lateral apical wall; L bas, lateral basal wall; L mid, lateral midwall; PL, posterolateral wall; S ap, septal apical wall; S bas, septal basal wall; S mid, septal midwall.

have been annotated by clinical experts and used as ground truth contours. These studies comprised the training set and were selected to span a wide range of image quality and of anatomical variations expected in daily clinical practice. To provide the most robust estimate

points and the RV endocardial surfaces were automatically defined and tracked throughout the cardiac cycle. The results were displayed side by side as dynamic RV short-axis slices of the base and mid-RV levels and RV long-axis views (Figure 2). Finally, a 3D dynamic surface

4 Genovese et al

Journal of the American Society of Echocardiography - 2019

Figure 2 Layout of the final phase of the automated workflow: end-diastolic and end-systolic RV endocardial contouring along the RV long-axis (bottom) and two RV short-axis views (top and middle).

Figure 3 Dynamic 3D surface rendering of the RV endocardial surface (left panel) is obtained along with the RV volume-time curve, from which RV size and function parameters are obtained (right). Abbreviations as in Figure 2.

model of the RV was generated, from which EDV and ESV were obtained from the RV volume-time curve and EF calculated using the above formula (Figure 3 and Video 1, available at www. onlinejase.com). Importantly, once the automated analysis was completed, the accuracy of the boundary detection was reviewed by carefully reviewing the RV contours tracking throughout the cardiac cycle in multiple 2D short-axis and long-axis views. It was judged accurate when no manual editing was deemed necessary to improve tracking. In

contrast, when the automated analysis was judged as inaccurate, manual editing was performed as needed. To this effect, the operator could edit the initial landmarks, the tricuspid annulus hinge points, and the RV endocardial borders as previously described.9,11 Briefly, endocardial border editing could be achieved by moving the 2D RV short-axis planes along the LV long axis from base to apex, while the 2D RV long-axis plane could be freely rotated around the LV long axis, from the RV inferior wall to the outflow tract, thereby encompassing the entire RV endocardial surface.

Genovese et al 5

Journal of the American Society of Echocardiography Volume - Number -

Table 1 Population and echocardiographic data Population, N = 56

Mean 6 SD or n (%)

Gender, male

31 (52.5)

Age, years

52 6 16

Height, cm

168 6 24

Weight, kg

81 6 19

BSA, m2

1.9 6 0.4

SBP, mm Hg

119 6 20

DBP, mm Hg

74 6 14

HR, bpm

73 6 16

Frame rate, vps

25 6 8

Table 2 RV volume and function indices measured using the ML-Based 3DE analysis RV HM, mL

CMR, mL

P

Bias, LOA, mL mL

EDV, 151.0 6 50.0 176.6 6 50.3 <.001 mL

r

P

25.6 41.3 0.91 <.001

ESV, mL

80.5 6 37.4

88.0 6 38.5 <.001

7.4 31.2 0.92 <.001

EF, %

48.0 6 7.8

51.3 6 10.1 <.001

3.3 10.2 0.87 <.001

Comparison with CMR reference (paired t-test, Bland-Altman, and linear regression analyses). CC, Correlation coefficient; LOA, limits of agreement.

Diagnosis Normal

17 (31)

HFREF

12 (21)

Ischemic heart disease

8 (14)

Heart transplant

5 (9)

HFPEF

5 (9)

Others

9 (16)

BSA, Body surface area; HFREF/HFPEF, heart failure with reduced/ preserved EF; HR, heart rate; SBP/DBP, systolic/diastolic blood pressure; VPS, volumes per ssecond.

Table 3 Distribution of the image quality, visualization scores in the study group, and accuracy of the ML-Based automated analysis Data set quality

Total

Patients (%)

Quality score (SD)

Accurate automated analysis (%)

56

21.3 6 5.5

18/56 (32)

0/13 (0)

Image quality subgroups by total score Poor, <18

13 (23)

13.5 6 3.6

Reproducibility Analysis

Fair, 18-23

23 (41)

21.3 6 1.6

4/23 (17)

Because of the deterministic nature of the ML algorithm, repeated measurements without manual corrections of the RV endocardial boundaries always result in exactly the same values. In other words, the automated analysis is 100% reproducible. It is the manual adjustments that cause intermeasurement variability, since these corrections are never exactly the same. To assess the magnitude of the variability caused by the manual corrections, intra- and inter-reader variability analysis was conducted, including repeated measurements by the same reader, at least1 month later, as well as by a second independent reader, both blinded to all prior CMR and 3DE measurements. Variability was expressed in terms of coefficients of variation (CoVs), calculated as the absolute difference between the corresponding pairs of repeated measurements as a percentage of their mean, and intraclass correlation coefficients (ICCs).

Good, >23

20 (36)

26.4 6 2.9

14/20 (70)

Statistical Analysis Quantitative data are presented as the mean 6 SD. Categorical data are presented as absolute numbers with percentages. The intertechnique comparisons between 3DE and CMR for RV volumes and EF included paired two-tailed Student’s t-tests, linear regression with Pearson correlation coefficients, and Bland-Altman analyses. Values of P < .05 were considered significant. CIs were defined at the 95th percentile around the mean. All analyses were performed using SPSS software (SPSS, Chicago, IL).

RESULTS Population data are summarized in Table 1, while RV measurements are reported in the first two columns of Table 2. Approximately one-third of the study group was composed of patients without cardiovascular disease, while the other two-thirds had a broad spectrum

of cardiac diagnoses. Accordingly, the study population encompassed a wide range of RV volumes (151 6 50 mL for EDV and 81 6 37 for ESV) and function (48% 6 8% for EF). The distribution of the image quality in the study group is reported in Table 3. The fully automated ML analysis was accurate in 18 patients (32%) without any corrections and required 15 6 1 seconds. In the remaining 38 patients (68%), manual editing was needed, prolonging the analysis time by 114 6 71 seconds. The accuracy of the automated ML analysis was strongly influenced by image quality, as all patients with poor image quality and most of those with fair image quality required manual editing, in contrast to the subgroup of patients with good image quality, where the automated analysis was accurate 70% of the time and therefore manual editing was necessary in only 30%. Most commonly, inadequate image quality resulted in incomplete visualization of parts of the endocardial boundary because of the ‘‘dropout’’ in the anterior RV free wall. Table 4 shows the details of the image quality results for each segment. Of note, the basal segment of the RV anterior wall, which is the region that is farthest away from the transducer, when imaging from the apical position, was the most difficult to visualize. With border corrections performed when needed, there was an excellent correlation between the ML-derived 3DE measurements and CMR reference values for all parameters (Figure 4 and Table 2), as reflected by r values of 0.91 for EDV, 0.92 for ESV, and 0.87 for EF; all P < .001. The 3DE measurements were reasonably accurate as reflected by biases of –26 6 21 mL for EDV (16% of the mean measured value), 7 6 16 mL for ESV (9%), and –3% 6 5% for EF (7%), even if the ML-based software algorithm slightly underestimated all three parameters compared with CMR.

6 Genovese et al

Journal of the American Society of Echocardiography - 2019

Table 4 Percentage of different endocardial visualization scores for each RV segment in the 56 study patients RV Views and Endocardial Segments

Not Visible

Partially Visible

Visible

Short axis Basal Anteroseptal

5

29

66

Inferoseptal

4

11

86

Inferior

4

27

70

Posterolateral

20

25

55

Anterolateral

54

36

11

Basal

4

11

86

Mid

5

32

63

14

50

36

4Ch Septum

Apex Lateral Basal

0

41

59

Mid

20

55

25

Apex

45

48

7

Basal

2

7

91

Mid

4

7

89

16

43

41

Basal

79

18

4

Mid

57

32

11

Apex

46

43

11

Inflow-outflow Inferior

Apex Anterior

The results of the intra- and interoperator variability caused by the manual corrections are summarized in Table 5, depicting excellent reproducibility for all measurements, as reflected by CoV < 7% and ICC $ 0.95 for all measurements.

DISCUSSION The main findings of our study are (1) the new ML-based algorithm provided reasonably accurate RV volumes and EF measurements and had an excellent reproducibility; (2) the fully automated analysis was accurate in 32% of the patients and required only 15 6 1 seconds, whereas manual editing was necessary in the remaining cases and required <2 minutes on the average; (3) the accuracy of the fully automated analysis was dependent on the image quality. The recent focus on the diagnostic and prognostic value of RV size and function has led to a better understanding and appreciation of the complexities of the 3D RV shape, which explains why modeling fails to yield satisfactory volume calculations by 2DE.22,23 This awareness has underscored the limitations of 2D assessment of RV size and function and has defined 3DE RV volumes and EF as the more promising parameters. Obtaining good-quality RV 3DE data sets remains considerably more difficult than for the LV. When the 3D acquisition is correctly

performed in the 4Ch RV-focused apical view and the right ventricle is fully included in the 3D data set, the endocardial border of the basal segment of the anterior wall is often difficult to visualize, because it is the most distant from the probe and the closest to the sternum. Interestingly, it has been demonstrated that the use of ultrasoundenhancing agents improves RV endocardial visualization, allowing accurate 3D RV assessment to be performed also in patients with poor baseline quality.10 However, since the existing software packages rely mostly on cumbersome manual postprocessing, the analysis time required to obtain accurate RV volumes and EF might be judged as unacceptable in a busy echocardiography laboratory. The development of a near-automated postprocessing analysis tool has the potential to overcome these limitations, with respect to both time constraints and improvement in measurement accuracy and reproducibility, especially among operators with different experience and training. One of the advantages of 3DE methodology is that unlike 2DE imaging, where the choice of imaging plane during acquisition fully determines the view, with 3DE imaging, the term ‘‘view’’ is less meaningful, because imaging planes can be extracted offline from the 3D data sets to display an unlimited number of views. Importantly, however, to allow volume measurement of a given chamber, it is essential that the chamber be included in the scan volume in its entirety. Because scan volume is limited by the need to maximize spatial and temporal resolution, it is not always easy (or even possible) to include both left and right sides of the heart in a single scan, especially in patients with enlarged ventricles. For this reason, we believe that two separate acquisitions, one for each side, should be used to ensure complete acquisition of both ventricles, without having walls and/or parts of the cavity inadvertently move out of the scan volume during the cardiac cycle, precluding accurate measurements during postprocessing. The interest in computer-aided diagnosis has existed for decades, particularly in radiology, to assist with chest imaging. This ultimately paved the way for the use of artificial intelligence, ML, and deep learning techniques in medical imaging. ML is an application of artificial intelligence that enables a computer program to learn complex relationships or patterns from empirical data and generate accurate decisions. By definition, ML improves diagnostic performance by increasing exposure to data sets without the need for explicit programming. The development of ML algorithms that identify cardiac structures and chambers in 3DE data sets may allow accurate, near-automated chamber quantitation, which could potentially revolutionize the practice of echocardiography. While experienced readers are still needed to verify and amend the automated identification of endocardial borders when necessary, the use of ML algorithms would significantly reduce the time required to obtain these measurements. In 3DE, ML algorithms were first used for LV quantification likely because of the relative simplicity of the LV shape in most patients.24,25 Interestingly, image quality has been described as a pivotal determinant for accuracy also for the automated LV quantification algorithm.20 In contrast, because of the complex RV shape, the development of ML tools for the fully automated 3D RV assessment could not rely on a simple geometrical shape and therefore was a far more ambitious goal. In this study, we tested the performance of the first ML-based fully automated algorithm for RV quantification from 3DE data sets. Our study population encompassed patients with a wide range of RV sizes, function, and image quality, aimed at broad and unbiased evaluation of the performance of the new approach. In this group, with quick manual adjustments that were mostly needed in patients with

Journal of the American Society of Echocardiography Volume - Number -

Genovese et al 7

Figure 4 Comparison between ML-based and CMR measurements of RV volumes and EF: linear regression (left) and Bland-Altman analysis (right) for EDV (top), ESV (middle), and EF (bottom). suboptimal images, this algorithm resulted in reasonably accurate measurements of RV volumes and EF. The impressive feature of this tool from the point of view of an echocardiographer is that the starting point for the reader is no longer a nonsegmented 3D data set but rather a preprocessed output with endocardial borders in any desired cut plane that need to be confirmed or corrected by boundary adjustment. It is likely that with future technological refinements leading to improved image quality, as well as algorithmic improvements with additional training on larger data sets, the percentage of patients with accurate fully automated measurements without manual correction will increase over time. Our results in the subgroup of patients with good image quality provide evidence that this can indeed be achieved.

Regarding the nature of the intermodality differences reported in this study, the underestimation of 3DE RV volumes compared with CMR has been abundantly reported in the literature and is also larger with the worsening of image quality.10 The main reason has been imputed to the lower spatial definition of 3DE, leading to the lack of a clear identification of the interface between the compacted RV endocardium and trabeculations, resulting in smaller cavity contouring. Because of our study design that targeted consecutive patients unselected for the 3DE image quality, our study population included a high proportion of patients with fair and poor images, which may have accounted for the slightly larger RV volume underestimation, when compared with previous studies.11

8 Genovese et al

Journal of the American Society of Echocardiography - 2019

Table 5 Reproducibility of the ML-based software for 3DE measurements of RV volumes and function Intraobserver variability

Interobserver variability

CoV, %

ICC

CoV, %

ICC

EDV

3.4 6 4.1

0.99 [0.98-0.99]

6.1 6 7.7

0.97 [0.94-0.98]

ESV

4.3 6 4.9

0.99 [0.99-1.00]

6.5 6 8.2

0.98 [0.96-0.99]

EF

3.1 6 4.0

0.96 [0.94-0.98]

3.8 6 4.7

0.95 [0.91-0.97]

Limitations This was a single-center study performed in unselected and consecutive patients with clinical indications for CMR. Despite the limited sample size, the number of patients was sufficient to reach high levels of statistical significance in comparisons with the reference technique. Nevertheless, future studies with larger numbers of patients and specific pathologies spanning a wider range of RV volumes and EF would be needed to further confirm our findings. Importantly, the ability of the ML-based approach to measure RV volume in patients with irregular heart rhythms needs to be evaluated in future studies. The use of CMR as a reference standard for RV volume quantification has its limits, and, although commonly used in validation studies, CMR is a tomographic method that requires manual tracings of multiple short-axis slices, rather than a truly volumetric technique. Also, the quantitative approach used for the evaluation of the 3DE image quality was arbitrary; however, it is unlikely that using a different visualization quality scale would have yielded substantially different findings.

CONCLUSION In summary, the ML-based software for the quantification of 3D RV volumes and EF provided an accurate and highly reproducible quantification in a completely automated fashion in one-third of patients with unselected image quality, while in the remaining patients it required minimal manual editing. Although future studies are needed to determine the impact of this algorithm on routine clinical evaluation of the right ventricle, this new software is the first step toward the automated 3DE-based RV quantification, which is likely to contribute to the widespread implementation of 3DE for RV assessment in clinical practice.

SUPPLEMENTARY DATA Supplementary data related to this article can be found at https://doi. org/10.1016/j.echo.2019.04.001.

REFERENCES 1. Ghio S, Gavazzi A, Campana C, Inserra C, Klersy C, Sebastiani R, et al. Independent and additive prognostic value of right ventricular systolic function and pulmonary artery pressure in patients with chronic heart failure. J Am Coll Cardiol 2001;37:183-8. 2. Mehta SR, Eikelboom JW, Natarajan MK, Diaz R, Yi C, Gibbons RJ, et al. Impact of right ventricular involvement on mortality and morbidity in patients with inferior myocardial infarction. J Am Coll Cardiol 2001; 37:37-43.

3. Zornoff LA, Skali H, Pfeffer MA, St John Sutton M, Rouleau JL, Lamas GA, et al. Right ventricular dysfunction and risk of heart failure and mortality after myocardial infarction. J Am Coll Cardiol 2002;39: 1450-5. 4. Burgess MI, Mogulkoc N, Bright-Thomas RJ, Bishop P, Egan JJ, Ray SG. Comparison of echocardiographic markers of right ventricular function in determining prognosis in chronic pulmonary disease. J Am Soc Echocardiogr 2002;15:633-9. 5. Pueschner A, Chattranukulchai P, Heitner JF, Shah DJ, Hayes B, Rehwald W, et al. The prevalence, correlates, and impact on cardiac mortality of right ventricular dysfunction in nonischemic cardiomyopathy. JACC Cardiovasc Imaging 2017;10:1225-36 (10 Part B). 6. Rudski LG, Lai WW, Afilalo J, Hua L, Handschumacher MD, Chandrasekaran K, et al. Guidelines for the echocardiographic assessment of the right heart in adults: a report from the American Society of Echocardiography endorsed by the European Association of Echocardiography, a registered branch of the European Society of Cardiology, and the Canadian Society of Echocardiography. J Am Soc Echocardiogr 2010;23:685-713. quiz 86-88. 7. Grewal J, Majdalany D, Syed I, Pellikka P, Warnes CA. Three-dimensional echocardiographic assessment of right ventricular volume and function in adult patients with congenital heart disease: comparison with magnetic resonance imaging. J Am Soc Echocardiogr 2010;23:127-33. 8. Leibundgut G, Rohner A, Grize L, Bernheim A, Kessel-Schaefer A, Bremerich J, et al. Dynamic assessment of right ventricular volumes and function by real-time three-dimensional echocardiography: a comparison study with magnetic resonance imaging in 100 adult patients. J Am Soc Echocardiogr 2010;23:116-26. 9. Medvedofsky D, Addetia K, Patel AR, Sedlmeier A, Baumann R, Mor-Avi V, et al. Novel approach to three-dimensional echocardiographic quantification of right ventricular volumes and function from focused views. J Am Soc Echocardiogr 2015;28:1222-31. 10. Medvedofsky D, Mor-Avi V, Kruse E, Guile B, Ciszek B, Weinert L, et al. Quantification of right ventricular size and function from contrastenhanced three-dimensional echocardiographic images. J Am Soc Echocardiogr 2017;30:1193-202. 11. Muraru D, Spadotto V, Cecchetto A, Romeo G, Aruta P, Ermacora D, et al. New speckle-tracking algorithm for right ventricular volume analysis from three-dimensional echocardiographic data sets: validation with cardiac magnetic resonance and comparison with the previous analysis tool. Eur Heart J Cardiovasc Imaging 2016;17:1279-89. 12. Nagata Y, Wu VC, Kado Y, Otani K, Lin FC, Otsuji Y, et al. Prognostic value of right ventricular ejection fraction assessed by transthoracic 3D echocardiography. Circ Cardiovasc Imaging 2017;10 https://doi.org/10.1161/CIRCIMAGING.116.005384. 13. Maffessanti F, Gripari P, Tamborini G, Muratori M, Fusini L, Alamanni F, et al. Evaluation of right ventricular systolic function after mitral valve repair: a two-dimensional Doppler, speckle-tracking, and threedimensional echocardiographic study. J Am Soc Echocardiogr 2012;25: 701-8. 14. Nochioka K, Querejeta Roca G, Claggett B, Biering-Sorensen T, Matsushita K, Hung CL, et al. Right ventricular function, right ventricular-pulmonary artery coupling, and heart failure risk in 4 US communities: The Atherosclerosis Risk in Communities (ARIC) Study. JAMA Cardiol 2018;3:939-48. 15. Lang RM, Badano LP, Mor-Avi V, Afilalo J, Armstrong A, Ernande L, et al. Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. J Am Soc Echocardiogr 2015;28:1-39.e14. 16. van Ginneken B. Fifty years of computer analysis in chest imaging: rulebased, machine learning, deep learning. Radiol Phys Technol 2017;10: 23-32. 17. Wang S, Summers RM. Machine learning and radiology. Med Image Anal 2012;16:933-51. 18. Tsang W, Salgo IS, Medvedofsky D, Takeuchi M, Prater D, Weinert L, et al. Transthoracic 3D echocardiographic left heart chamber quantification

Journal of the American Society of Echocardiography Volume - Number -

using an automated adaptive analytics algorithm. JACC Cardiovasc Imaging 2016;9:769-82. 19. Medvedofsky D, Mor-Avi V, Amzulescu M, Fernandez-Golfin C, Hinojar R, Monaghan MJ, et al. Three-dimensional echocardiographic quantification of the left-heart chambers using an automated adaptive analytics algorithm: multicentre validation study. Eur Heart J Cardiovasc Imaging 2018;19:47-58. 20. Medvedofsky D, Mor-Avi V, Byku I, Singh A, Weinert L, Yamat M, et al. Three-dimensional echocardiographic automated quantification of left heart chamber volumes using an adaptive analytics algorithm: feasibility and impact of image quality in nonselected patients. J Am Soc Echocardiogr 2017;30:879-85. 21. Narang A, Mor-Avi V, Prado A, Volpato V, Prater D, Tamborini G, et al. Machine learning based automated dynamic quantification of left heart chamber volumes. Eur Heart J Cardiovasc Imaging 2019; 20:541-9.

Genovese et al 9

22. Addetia K, Maffessanti F, Yamat M, Weinert L, Narang A, Freed BH, et al. Three-dimensional echocardiography-based analysis of right ventricular shape in pulmonary arterial hypertension. Eur Heart J Cardiovasc Imaging 2016;17:564-75. 23. Addetia K, Maffessanti F, Muraru D, Singh A, Surkova E, Mor-Avi V, et al. Morphologic analysis of the normal right ventricle using threedimensional echocardiography-derived curvature indices. J Am Soc Echocardiogr 2018;31:614-23. 24. Maffessanti F, Lang RM, Niel J, Steringer-Mascherbauer R, Caiani EG, Nesser HJ, et al. Three-dimensional analysis of regional left ventricular endocardial curvature from cardiac magnetic resonance images. Magn Reson Imaging 2011;29:516-24. 25. Salgo IS, Tsang W, Ackerman W, Ahmad H, Chandra S, Cardinale M, et al. Geometric assessment of regional left ventricular remodeling by threedimensional echocardiographic shape analysis correlates with left ventricular function. J Am Soc Echocardiogr 2012;25:80-8.