Feasibility and reproducibility of semi-automated longitudinal strain analysis: a comparative study with conventional manual strain analysis

Introduction

Myocardial function measurement is the basis for thediagnosis of cardiac diseases. Today it can performusing two-dimensional speckle-tracking echocardiography (2D-STE), which is a technique designed to evaluate myocardial deformation. Left ventricular globallongitudinal strain (LVGLS) is the most mature andwidely used parameter obtained using 2D-STE. LVGLShas been shown to have additional value in risk stratifcation and outcome prediction compared to LV ejection fraction (LVEF) in a variety of clinical conditions[1–5]. Te 2015 guidelines of the European Societyof Cardiovascular Imaging (EACVI) and the American Society of Echocardiography (ASE) recommendedthe use of LVGLS as a supplement to the LVEF whenassessing LV function [6]. Although 2D-STE was originally used as a tool specifcally designed for LV strainmeasurement, researchers also applied it to the analysisof right ventricular (RV) and left atrial (LA) myocardial deformation [7–9]. Several studies have shown thatoutcome prediction can be improved in many heart diseases using LA strain [7,10–12] and RVGLS [13–16].

Challenges remain in the application of 2D-STE in clinical practice. First, because the shape, structure, and function of the LA and RV are diferent from those of the LV,application of the same 2D-STE technique to strain analysisof diferent cardiac chambers is controversial [7]. Second,tedious and time-consuming manual editing of non-automated myocardial strain evaluations limits clinical application of conventional echocardiographic techniques forstrain analysis [17]. Terefore, automatic, and chamberspecifc methods for myocardial strain assessment havebeen recently introduced and used in research [18].

At present, there are few comparative studies betweenmethods for automatic and conventional manual strainanalysis. Tus, the purpose of this study was to compareAutoStrain, a new automatic and chamber-specifc strainanalysis software (referred to as automatic strain analysis), with the conventional QLab software for LV strainanalysis (referred to as manual strain analysis).

Methods

Study cohort

From March 2021 to April 2021, a total of 159 consecutive healthy volunteers who underwent routine annualcheck-up in our hospital were recruited. For all the volunteers, the ECG and X-ray results were normal. All subjects provided written informed consent and study wasapproved by the Human Research Ethics Committee ofShenzhen People’s Hospital.

Echocardiography

A Philips Epic7C or Philips CVx cardiac ultrasound scanner (Philips®, Best, Te Netherlands) was used to acquirethe standard apical four-chamber, two-chamber, andlong-axis views as well as the RV-focused four-chamberview of the heart. At least 4 cardiac cycles were recordedfor each view. Images were saved in DICOM format. LVend-diastolic volume (LVEDV), end-systolic volume(LVESV), and LVEF were calculated by the biplane Simpson’s method.

Strain analysis

The LV strain analysis was performed in the standardapical four-chamber, two-chamber, and long-axis viewaccording to the guidelines of the ASE [19]. Te RVfocused four-chamber view was used for the RV strainanalysis, and the standard apical four-chamber view forthe LA analysis [20]. Te same image cine-loops wereused for online analysis by the AutoStrain and ofineanalysis by the Qlab 9.1 software (Philips Medical Systems). Two experienced observers (G.P. and S.L., withmore than 5 to 7 years-experience in echocardiography)performed the strain evaluations blinded to each other.

Automated strain analysis

Online strain analysis was performed by the AutoStrainsoftware. Te “LV strain”, “LA strain” and “RV strain”applications were selected for the analysis of the LV, LA,and RV, respectively. Te software automatically recognized the image, generated a region of interest (ROI),tracked the endocardium throughout the cardiac cycleand provided the strain values and curves for each myocardial segment (Figs. 1and2). After the automatic strainanalysis, the two observers reviewed the tracking qualityfor each myocardial segment. If the tracking of more thantwo cardiac segments in the same view was unsatisfactory, the case was considered inadequate for analysis [6].If necessary, each observer manually corrected the ROIto obtain a satisfactory fnal strain result. To assess thevariability between diferent cardiac cycles, a diferentcardiac cycle was analyzed and compared with the previous one two weeks later.

Manual strain analysis

Stored image cine-loops were assessed ofine by theQLAB 9.1 Philips workstation. After the cine-loops of the3 standard apical views were selected, observers manually defned three key points, two at the hinge points ofthe mitral valve annulus and one at the LV apex in eachimage (Fig. 1). Te software then aligned and trackedthe myocardium, providing segmental strain values andcurves. Te apical four-chamber mode used for LV strainanalysis was used for both RV and LA strain analysis.

For RV strain analysis, the observer defned threepoints on the image, two at the lateral and septal side ofthe tricuspid valve annulus and one at the RV apex. Tesoftware tracked the myocardium, providing segmentaland average values and curves of the RV free wall strain(RVFWS) (Fig. 2).

For the LA strain analysis, the observer defnedthree points on the image, two at the hinge points ofthe mitral valve annulus and one at the LA roof. Tewidth of the ROI was set as 3 mm, according to guidelines [20,21]. Te LA peak strain, strain at the beginning of LA contraction, and strain at end-diastole wererecorded [20]. Te LA reservoir strain (LASr) wasdefned as LA peak strain – LA end-diastolic strain, theLA conduit strain (LAScd) was defned as the LA peakstrain - strain at the beginning of atrial contractionand the LA pump strain (LASct) was defned as LA atthe beginning of atrial contraction – LA end-diastolicstrain (Fig. 2).

All traces were reviewed by the observers and, if necessary, they were manually corrected. At the moment of themanual strain analysis, the observer was blinded to theresults of the semi-automatic strain analysis.

Strain analysis time

In 50 randomly selected subjects the time needed tomeasure LV, LA, and RV strains by the two analysismethods was calculated. Te strain measurement timewas defned as the time between the initial selection ofthe echocardiographic image to analyze and the completion of the strain calculation.

Intraobserver and interobserver variability

The intraclass correlation coefcient (ICC) and the BlandAltman analysis were used to determine the intra-, andinter-observer reproducibility in both groups of 20 randomly selected healthy volunteers. Intraobserver variabilitybetween the frst and second measurements (after 30 days)calculated by the same observer (G.P.), who was blindedto the previous measurements. Interobserver variabilityby two independent analysts (G.P and S.L) was calculated,with both observers were blinded to the result of the other.

Statistical analysis

Continuous variables are expressed as the mean ± standard deviation or median and interquartile range, andcategorical variables as numbers and percentages. Tepaired Student’s t test and linear correlation analysiswere used to compare and correlate the strains measured by the two diferent methods. Te Bland–Altmananalysis was used to evaluate the agreement between thetwo methods and the two cardiac cycles in semi-automatic analysis. Intra- and interobserver measurementvariability were assessed using the intraclass correlation coefcient (ICC) and the Bland-Altman analysis.At the Bland-Altman analysis the mean error and limitsof agreement (LOAs, mean error ± 1.96 standard deviations) were calculated. Te SPSS software (version 18.0)was used for statistical analysis. A p value <0.05 was considered statistically signifcant.

Results

Of the 159 volunteers recruited, 3 were excluded becauseof severe arrythmias or moderate aortic regurgitation,and 4 due to poor image quality. A total of 152 subjectswere fnally enrolled (mean age 40 years, range 20–69years). Eighty (53%) subjects were males. Te mean LVEFwas 62% (range 52%–73.5%). Subject characteristics areshown in Table 1.

Left ventricular strain analysis

For LVGLS measurement, the success rate of semi-automatic analysis was 95.4% (145/152) and that of manualanalysis was 98.0% (149/152). In the semi-automatic analysis, 14 cases (9.6%) required manual ROI adjustment.

There was no signifcant diference in mean valueof LVGLS in semi-automatic and manual analysis(-21.0%±2.5% vs. -20.8%±2.4%,p=0.230; Fig. 3andTable 2) and the correlation coefcient was 0.84 (p<0.001;

Fig. 4). Te Bland–Altman analysis showed the absenceof a bias (mean error -0.1%, LOAs -2.9, 2.6%%; Fig. 4andTable 2).

In semi-automatic strain analysis, there was no signifcant diference in LVGLS between the two cardiac cycles(-21.0%±2.5% vs. -21.1%±2.4%,p=0.450) (SupplementaryTable 1), and correlation was good (r=0.85,p<0.001). TeBland–Altman plots are shown in Fig. 5and reported inTable 2.

Right ventricular free wall strain analysis

In the analysis of RVFWS, the success rate of semi-automatic strain analysis was 96.7% (147/152) and that of manual strain analysis 90.8% (138/152). In the semi-automaticstrain analysis, 7 cases (4.9%) required ROI adjustment.

The semi-automatic analysis of RVFWS was signifcantly diferent from the manual strain analysis(-26.4%±4.8% vs. -31.3%±5.8%,p<0.001) and the absolutevalue of the manual strain was higher (Fig. 3and Table 2).Te correlation between the RVFWS obtained by thetwo methods was poor (r=0.248,p<0.005; Fig. 4). At theBland–Altman analysis the mean error was 4.9% and theLOAs were -8.1, 17.9% (Fig. 4and Table 2).

In semi-automatic strain analysis, there was no signifcant diference in RVFWS between the two cardiac cycles(-26.4%±4.8% vs. -26.5%±4.8%,p=0.886), and correlationwas moderate (r=0.74,p<0.001; Fig. 5and Table 2). TeBland–Altman plots are shown in Fig. 5and reported inTable 2.

Left atrial strain analysis

In the analysis of LA strain, the success rate of both semiautomatic and manual analysis was 99.3% (151/152).Eleven cases (7.3%) in semi-automatic strain analysisneeded manual adjustment.

There were signifcant diferences between semiautomatic and manual analysis for LASr and LAScd.In particular, the absolute value of semi-automatic

strain was higher (LASr: 48.0%±10.0% vs. 37.6%±9.9%,p<0.001; LAScd: -31.8%±8.9% vs. -22.2%±7.9%,p<0.001)(Fig. 3). For LASct, there was no signifcant diferencebetween the two methods (-16.3%±6.5% vs. -15.4%±6.0%,p=0.176).

All the LA strains obtained by the two methods werecorrelated. Te correlation of LAScd values was moderate (r=0.55,p<0.001), but that between LASr and LASctvalues was poor (LASr:r=0.33,p<0.001; LASct: r=0.21,p<0.005) (Fig. 4). Te Bland–Altman analysis showed alarge bias of 10.5% with wide LOAs (-12.3, 33.3%).

In semi-automatic strain analysis, there was no signifcant diference in LA strain between the two cardiaccycles (LASr: 48.0%±10.0% vs. 47.7%±10.2%,p=0.467;LAScd: -31.8% ±8.9% vs. -31.8%±8.7%,p=0.920; LASct:-16.3%±6.5% vs. -16.0%±7.0%,p=0.506). Correlation between the two cardiac cycles were good (LASr:r=0.88,p<0.001; LAScd:r=0.80,p<0.001; LASct:r=0.75,p<0.001) (Fig. 5). Te Bland–Altman plots are shown inFig. 5and reported in table 2.

Comparison of strain analysis time

The average times for LV, RV, and LA strain analysis in50 subjects were: 17 s, 9 s, and 7 s for semi-automaticanalysis and 91 s, 49 s, and 51 s for manual analysis (allp<0.001), respectively (Fig. 6).

Reproducibility analysis

For semi-automatic strain analysis, intra-observer andinter-observer ICCs for LVGLS, RVFWS, and LASr measurements were 0.96–0.98 and 0.90–0.95, respectively. Atthe Bland–Altman analysis, intra-observer mean errorand LOAs were -0.08% (-1.98, 1.82%) for LVGLS, 0% (-3.5,3.5%) for RVFWS, and 0.3% (-6.1, 6.7%) for LASr measurements; inter-observer mean error and LOAs were-0.08 (-2.68, 2.52%) for LVGLS, 0.23% (-4.67, 5.12%%) forRVFWS, and 0.5% (-7.3, 8.3%) for LASr measurements.

For manual strain analysis, intra-observer, and interobserver ICCs for LVGLS, RVFWS, and LASr measurements were 0.82–0.90 and 0.80–0.87 (SupplementaryTable 1).

Discussion

Myocardial strain can assess myocardial function and itsapplication is increasing in clinical practice. However, theconventional approach to myocardial strain measurement is complex, time-consuming, and dependent onthe observers’ experience, thus novel methods have beendeveloped which rely on a semi-automatic and chamberspecifc analysis. In this study a novel and conventionalsoftware for myocardial strain measurements have beencompared and reproducibility of the novel software hasbeen evaluated (Graphical Abstract).

Reproducibility analysis

Intra- and inter-observer variabilities of semi-automaticmeasurements of LVGLS, RVFWS and LASr were betterthan those of manual strain analysis of the same measurements, as indicated by higher ICC values. Also, cycleto-cycle variability with the semi-automatic method wasassessed, measuring images from two consecutive cycles with stable heart rate. At the Bland-Altman analysis theLOAs of cycle-to-cycle measurements were wider thanthose of intra-observer measurement, which may indicate the variability caused by diferent cardiac cycles.

Left ventricular strain analysis

In our study the semi-automated LV strain analysis had ahigh success rate (95.4%), comparable to the feasibility ofprevious studies on automated strain analysis [18,22]. Only9.6% of cases required manual adjustments, a signifcantlylower percentage compared to that reported by KawakamiH et al. [22], which was close to 40%. Tis may be explainedby the diferent study population (healthy volunteers vs.asymptomatic patients with heart failure risk factors).

Semi-automatic strain analysis shortens the analysis time. Te most widely-used speckle tracking methodcurrently requires several steps performed by the operator, with an execution time ranging from 5 to 10 minutes[23,24]. Artifcial intelligence based on deep learning byZhang et al. [25] achieved semi-automatic LV strain analysis, but it still took a long time to calculate GLS of eachview, which was 1 to 4 minutes. In the study of Kawakamiet al. [22], the AutoStrain software was applied for automatic strain analysis of the LV, and the analysis time alsorequired 0.5 min/patient. In this study, the time requiredfor strain analysis of the LV was signifcantly shorter thanthat of conventional strain analysis, and it was similar tothe time required for the recent automatic strain analysis of the LV based on deep learning artifcial intelligence(17.4s vs 13s) [26]. Tis analysis time can meet the needfor a rapid point-of care evaluation of critically ill patients.Compared with the manual strain analysis method (thereference method recommended by the guidelines), theresults of semi-automatic strain analysis and manual strainanalysis showed good reproducibility and agreement.

Strain analysis of the right ventricular free wall

In this study we showed that RVFWS analysis had a highsuccess rate, with a feasibility of 96.7%, which is comparable to the feasibility of previous studies [27,28]. In somecases endocardial tracking was limited by poor imagequality or by cardiac and respiratory motion, determining the RVFW to be “out of volume” in some frames(mostly in diastole).

We found that the semi-automatic analysis of theRVFWS provided diferent results in comparison withthe manual analysis. Mirea et al. showed that the RVFWSobtained by a RV-specifc tracking software was slightlylower than that based on a non-dedicated software,although the diference was not statistically signifcant(p=0.05) [29]. Li et al. showed that automated RVFWSwas signifcantly higher than manual RVFWS in thewhole study population but not in the subgroup withnormal RV function [28]. These controversial resultsneed further investigations.

The semi-automatic analysis of RVFWS required, onaverage, only 9 s to be completed, which is consistent withrecent research [28]. Other similar RV-dedicated softwareneeded several key points to be manually defned in theanalysis process, so the average analysis time was 58 s[29]. Therefore, the automated analysis software evaluatedin our study has the potential to facilitate the assessmentof RV function.

Strain analysis of the left atrium

The feasibility of the semi-automatic strain analysis ofthe LA in this study is relatively high, comparable to thefeasibility of previous studies [18]. Te LASr obtainedby semi-automatic strain was signifcantly higher thanthat evaluated by manual strain: this might be the consequence of a diferent ROI setting. In manual LA strainanalysis, the ROI width was adjusted according to guidelines, but the overall ROI was a uniform elliptic curvewith a width of 3 mm, which may exceed the actual thickness of the LA wall. In addition, the wall of the LA is notcompletely uniform in thickness. Te wall of the interatrial septum is very thin, and the motion range is large.Terefore, the movement of the interatrial septum maynot be well tracked in manual method, which may causeunderestimation of the overall longitudinal strain. A largeROI may lead to the inclusion of adjacent pericardium,pulmonary vein wall, and other structures, thus underestimating the strain value [30].

Mirea et al. [29] compared the LA strain obtained usinga non-LA specifc and a LA-specifc software and foundthat LASr was slightly higher when using a LA-specifctool, although the diference was not statistically signifcant. Results of our investigation are in line with the conclusions of Mirea et al. In the study of Mirea et al., the LAdedicated tool was manual rather than semi-automated,and required to defne the position of the mitral valvering and the left atrial roof [29].

In this study, the semi-automatic analysis of LA strainimproved the time-efciency of the analysis. Te analysistime of LA strain was signifcantly shortened comparedto a conventional LA strain study [31] (7 s vs. 51 s) and toa LA specifc strain analysis study [29] (7 s vs. 45 s).

Limitations

(1) The sample size is relatively small and only healthypeople with normal LVEF were included. (2) We did notprovide a comparison with an external reference technique for strain measure, for example cardiac magneticresonance imaging. Tis study only compared two diferent echocardiographic approaches to myocardial strainevaluation, thus it lacks assessment of accuracy for each method. (3) Te detection of myocardial strain was basedon 2D-STE and therefore has the limitations of a 2Dapproach.

Conclusions

Semi-automatic strain analysis has the potential toimprove efciency in measurement of myocardial strain.It shows good agreement with the manual analysis for LVstrain measurement.