The International Journal of Cardiovascular Imaging

, Volume 28, Issue 6, pp 1357–1368

Sources of variation and bias in assessing left ventricular volumes and dyssynchrony using three-dimensional echocardiography

Authors

    • Department of Cardiac, Thoracic and Vascular SciencesUniversity of Padua
  • Luigi P. Badano
    • Department of Cardiac, Thoracic and Vascular SciencesUniversity of Padua
  • Davide Ermacora
    • Department of Cardiac, Thoracic and Vascular SciencesUniversity of Padua
  • Gianluca Piccoli
    • Department of Radiology“Santa Maria della Misericordia” University Hospital
  • Sabino Iliceto
    • Department of Cardiac, Thoracic and Vascular SciencesUniversity of Padua
Original Paper

DOI: 10.1007/s10554-011-9985-0

Cite this article as:
Muraru, D., Badano, L.P., Ermacora, D. et al. Int J Cardiovasc Imaging (2012) 28: 1357. doi:10.1007/s10554-011-9985-0

Abstract

Study aim To explore various sources of variability in the measurement of LV volumes and dyssynchrony by 3D echocardiography (3DE). Methods We studied 100 patients (58 ± 18 years, 51 men) to assess the impact of: (1) manual editing; (2) 3D data set temporal resolution; (3) LV 16- or 17-segmentation model; (4) software sensitivity for automated endocardial surface detection; and (5) image quality, on the measurement of LV end-diastolic (EDV) and end-systolic (ESV) volumes, sphericity indices (EDSI, ESSI), ejection fraction (EF) and dyssynchrony (SDI). Two- and 4-beat LV full-volume data sets were analyzed and compared. Cardiac magnetic resonance (CMR) was used as reference in 26 patients. Results Manual editing of endocardial surface improved the agreement of LV volumes with CMR, but increased SDI (SDI17: 5.6 ± 0.5% vs. 4.3 ± 0.3%; P < 0.0001). Data set temporal resolution had no significant impact on LV parameters. Adding the 17th to 16-segment LV model did not significantly increase SDI. Reducing software sensitivity in endocardial surface detection increased EDV (101 ± 46 ml vs. 118 ± 50 ml) and sphericity, decreased SDI (SDI 17: 6.7 ± 3.3% vs. 2.9 ± 3.7%) (P < 0.05 for all), and improved agreement of EDV and ESV with CMR. Impact of software sensitivity in LV endocardium detection on LV parameters was related to image quality: higher on SDI in pts with suboptimal quality (SDI 17 bias 4.5% vs. 3.2%, P < 0.05); higher on LV volumes in patients with optimal quality (EDV bias 14 ml vs. 19 ml, ESV bias 5 ml vs. 9 ml; P = 0.01). Conclusions Manual editing, software settings and image quality significantly impact on 3D LV volumes and dyssynchrony assessment.

Keywords

Three-dimensional echocardiographyLeft ventricular volumesLeft ventricular dyssynchronySemi-automated softwareTechnical factorsTemporal resolution

Abbreviations

2DE

Two-dimensional echocardiography

3DE

Three-dimensional echocardiography

CMR

Cardiac magnetic resonance

CRT

Cardiac resynchronization therapy

EDSI

End-diastolic sphericity index

EDV

End-diastolic volume

ESSI

End-systolic sphericity index

ESV

End-systolic volume

LV

Left ventricle/ventricular

LVMD

Left ventricular mechanical dyssynchrony

SDI

Systolic dyssynchrony index

Introduction

Assessment of left ventricular (LV) size and function is of paramount importance in various clinical settings and has direct implications for patient management [1]. Suitability for device implantation [2, 3], indication to cardiac surgery [4, 5] or treatment initiation in asymptomatic patients with LV systolic dysfunction [6] are among the most critical decisions that rely on accurate LV function assessment.

Three-dimensional echocardiography (3DE) enables an accurate and comprehensive LV quantitation [7]. Superiority of 3DE over conventional two-dimensional (2DE) approach for LV volume measurement in comparison to cardiac magnetic resonance (CMR) has been well documented [8]. However, despite eliminating LV apical foreshortening and geometric assumptions, 3DE still yields a systematic underestimation of LV volumes, as shown in a meta-analysis of 95 studies having CMR as reference [9]. This fact, along with some discrepancies in the reproducibility of LV 3D volume measurements [1013], has hindered the widespread acceptance of 3DE for clinical decision making in the single patient.

3DE has been also reported as a useful technique to assess LV mechanical dyssynchrony (LVMD) and to identify responders to cardiac resynchronization therapy (CRT) [1418]. Nevertheless, notably different abnormality thresholds and variable reproducibility values for systolic dyssynchrony index (SDI) have been documented [1, 14, 16, 19, 20]. The wide overlap of the SDI values in patients with left bundle branch block (LBBB) and normal LV function in comparison with SDI in controls [1] added some concerns on its accuracy to identify LVMD. Moreover, the reliability of SDI arouse major controversy among 3DE experts [21, 22] and it concerned particularly patients with severe LV dysfunction, those that would actually benefit the most from CRT. Confounding effects of several technical factors have been assumed, such as limited temporal or spatial resolution, use of various temporo-spatial smoothing settings or human interventions etc., but these currently remain controversial and largely unsupported by evidence [21].

Accordingly, our aim was to test 5 technical factors—(1) manual editing of automatically detected LV endocardial surface; (2) temporal resolution of the 3D data set; (3) choice of 16- or 17-segment LV model; (4) different settings of the software sensitivity for automated endocardial surface detection; (5) image quality of the endocardium—and to assess their relative impact on the measurement of LV geometry and dyssynchrony by 3DE.

Methods

Patients

We studied 138 consecutive patients with various heart diseases referred for routine echocardiographic assessment of LV size and function. Surface 12-lead ECG recordings were obtained to assess QRS morphology and width. Thirty-three patients showed complete LBBB [23]. All patients underwent a 3DE study at the end of the standard 2DE examination. Patients with poor endocardial border visualization requiring contrast echocardiography (n = 17.12%) [24], arrhythmias (n = 11.8%), paced rhythm (n = 7.5%) or inability to cooperate for breath holding (n = 3.2%) were excluded from the study. Therefore, the final study population comprised 100 patients. Among the enrolled patients, 26 underwent CMR based on clinical indications. LV parameters derived from CMR analysis in this patient subset served to test the accuracy of 3DE analyses in our study group. The study was approved by local Ethics Committee and all patients gave their informed consent for the study.

Three-dimensional echocardiography

3DE was performed using a commercially available Vivid E9 ultrasound machine (GE Healthcare, Horten, Norway) equipped with 3 V matrix-array transducer. All patients were examined in left lateral position from apical approach and 3DE acquisitions were done during end-expiratory apnea using second-harmonic imaging. In every patient, 2 consecutive full-volume data sets were generated from ECG-gated subvolumes of 2 and 4 consecutive cardiac cycles, in order to obtain low and high temporal resolution LV data sets, respectively. Each acquisition was immediately approved for digital storage if, using 9-slice display mode, complete visualization of LV endocardial border and lack of stitching artifacts were verified. All examinations were exported to a digital server.

Two- and 4-beat full-volume data sets were randomly analyzed offline by a single observer (D.M.) using a commercially available dedicated semi-automated software (TomTec 4D LV function™, version 2.6, TomTec Imaging Systems, Gmbh, Unterschleissheim, Germany).

Our previous experience with 3DE analysis using the same software [13] indicated that frame-by-frame manual editing of the LV endocardial border yielded a LV cast with a “twitching” irregular surface even in normal ventricles, presumably inducing or aggravating measured LVMD. Accordingly, we performed a pilot study (see Results section) to assess the impact of manual editing on SDI in 10 patients from CMR subgroup having normal QRS width, global and regional LV systolic function. The findings from this pilot study were used to design our final study protocol, in which we consequently excluded any manual editing during the 3DE data set analysis to assess SDI.

Image quality of 3DE data set was defined as optimal if a sharp endocardial contour was evident in all segments. Conversely, image quality was defined as suboptimal if LV endocardial border had a lower signal-to-noise ratio and a blurred blood-tissue interface, or ≥2 missing segments.

A standardized image analysis protocol was applied to minimize operator’s subjectiveness:
  • 3DE data set alignment was performed so that LV longitudinal axis was set to cross the center of mitral valve opening and LV apex in each view;

  • Manual initialization of endocardial border identification was done right at the black-white interface;

  • Systolic and diastolic timing were manually set as minimal and maximal LV size;

  • Keeping the same initialization contour, LV parameters were measured at 3 arbitrarily chosen settings of the software sensitivity in automated border identification (adjustments possible from 0 to 100%): at the default level (highS, 75%), at intermediate (midS, 50%) and at low levels (lowS, 25%).

After analysis completion, 3D LV surface-rendered dynamic casts were generated and optionally subdivided in 16 and 17 pyramidal-shaped volumetric segments.

The following parameters from quantitative data panel were obtained for every analysis (i.e. using highS, midS and lowS setting for each 2- and 4-beat data set): LV end-diastolic (EDV) and end-systolic (ESV) volumes, ejection fraction (EF), end-diastolic (EDSI) and end-systolic (ESSI) sphericity indices, 16- and 17-segment SDI (SDI 16 and SDI 17), respectively. SDI 16 and SDI 17 were computed as the standard deviation of the time interval needed to reach minimal regional systolic volume for each segment of the 16 and 17-segment model, respectively, normalized for the cardiac cycle length (expressed as percent of R–R time interval).

Cardiac magnetic resonance (CMR)

CMR was performed <1 h apart from echo examination with a 1.5-T system (Magnetom Avanto, Siemens Medical Systems, Erlangen, Germany). In addition to the stack of 8–12 short-axis images covering the complete LV using a previously published protocol [25], three long-axis images were acquired using similar frame-rate, corresponding to LV apical 4-, 2-chamber and long-axis views (60° between each 2).

The acquisitions were analyzed offline using the same algorithm as for 3DE data sets (4D LV-Analysis MR v 1.0, TomTec Imaging Systems GmbH, Unterschleissheim–Germany), recently customed for LV analysis using CMR [26].

Statistical analysis

Data were summarized as mean ± SD or percentage, as appropriate. Bland–Altman analysis was used to determine the agreement between CMR and 3DE for LV volumes and EF measurements (in the subset of 26 patients who underwent CMR), or to compare the impact of different technical factors on LV parameters (in all study patients). Paired t test was used to compare mean values or biases, and F test to compare standard deviations for statistical significance. Relationships between parameters obtained with different analysis methods were evaluated by linear regression analysis with Pearson’s correlation coefficient. To assess the reproducibility of 3DE LV parameters, the 4-beat data sets of 10 random patients were re-analyzed using default sensitivity level (HighS) by the same observer (D.M.) at least 1 month after the first measurement, as well as by a second observer (L.P.B.) blinded to the results of the first observer. Intra-observer and inter-observer reproducibility were reported as the absolute difference of the corresponding pair of repeated measurements normalized to their average value in each patient and expressed as mean ± SD for entire sample group. P value <0.05 was considered statistically significant. Data analysis was performed using SPSS version 13.0 (SPSS Inc, Chicago, IL, USA) and MedCalc for Windows, 8.1.1.0 release (Mariakerke, Belgium) statistical softwares.

Results

Clinical and demographic characteristics of enrolled patients are summarized in Table 1. There were 63 patients (54 ± 18 year old, 44% men) with LV EF ≥ 50% and 37 patients (67 ± 12 years, 63% men) with various degrees of LV dysfunction (EF = 38 ± 9%, range 17–49%). In the subgroup of 26 patients who underwent clinically indicated CMR, mean LV volumes and EF were similar to those averaged over the remaining 74 patients (108 + 31 vs. 106 + 49 ml for EDV; 49 + 17 vs. 47 + 41 ml for ESV; and 57 + 9 vs. 56 + 11% for EF; P = NS for all).
Table 1

Baseline characteristics of the study population (n = 100)

Age (years)

58 ± 18

Men (%)

51%

HR (bpm)

68 ± 11

LBBB, n (%)

33 (33%)

LBBB QRS duration (ms)

160 ± 19

Indications for echo study (%)

 Ischemic heart disease

23

 Hypertensive heart disease

15

 Non-ischemic dilated cardiomyopathy

14

 Valvular heart disease

10

 Aortic pathology

10

 Anthracycline toxicity monitoring

9

 Transplanted heart

4

 Other

15

HR heart rate, LBBB left bundle branch block

Impact of manual editing of automatically identified LV endocardial border

As part of study protocol design, a preliminary analysis in 10 random patients with presumably synchronous LV contraction (QRS width <100 ms, normal LV regional function and EF) was performed. As expected, the addition of manual editing of LV endocardial border on top of the described analysis workflow (identical data set, temporal resolution, manual initialization, automated identification of endocardium, High S sensitivity setting etc.) significantly improved the agreement of LV volumes with CMR (EDV bias −15 ml vs. −31 ml and ESV bias −8 ml vs. −14 ml, P < 0.01 for both). Conversely, frame-by-frame manual editing induced a systematic increase of both SDI 16 (5.1 ± 0.3% vs. 3.6 ± 0.5%, P < 0.0001; bias 1.5%, LOA 0.3–2.7%) and SDI 17 (5.6 ± 0.5% vs. 4.3 ± 0.3%, P < 0.0001; bias 1.3%, LOA 0.1–2.5%) in comparison with no manual editing approach. Therefore, all further analyses about SDI excluded manual refinements of LV endocardial contour after automated endocardial border identification.

Impact of temporal resolution of 3D data set

Temporal resolution of 2-beat 3D data sets (25 ± 6 vps, range 10–49) was significantly lower than that achieved in 4-beat data sets (52 ± 16 vps, range 19–68, P < 0.0001).

Comparison of corresponding values of LV parameters obtained from 2- and 4-beat data sets (i.e. low versus high temporal resolution) while maintaining identical sensitivity setting for each comparison, did not show any significant impact on any of LV parameters (volumes, systolic function, sphericity, LVMD, P = NS for all) (Table 2). However, in comparison with patients with normal QRS width, those with LBBB had significantly wider limits of agreement of SDI when measured on 2- versus 4-beat acquisitions (SDI 16 and SDI 17, P < 0.01 for all), irrespective of sensitivity setting.
Table 2

Comparisons of left ventricular parameters derived from 4- versus 2-beat 3DE data sets analyzed using 3 different levels of sensitivity settings of automated border detection in the whole patient population and stratified according to left ventricular ejection fraction

N = 100

4-beat 3DE data set

2-beat 3DE data set

HighS

MidS

LowS

HighS

MidS

LowS

EDV (ml)

 All (n = 100)

101 ± 46

117 ± 49*

118 ± 50^

98 ± 38

116 ± 42*

120 ± 44^

 EF ≥ 50% (n = 63)

85 ± 28

100 ± 31*

102 ± 32^

84 ± 27

102 ± 32*

106 ± 33^

 EF < 50% (n = 37)

132 ± 45

148 ± 49*

148 ± 48^

133 ± 45

153 ± 51*

157 ± 55^

ESV (ml)

 All (n = 100)

52 ± 39

61 ± 42

60 ± 42

50 ± 31

58 ± 34

57 ± 34

 EF ≥ 50% (n = 63)

35 ± 12

42 ± 14

40 ± 14

35 ± 12

42 ± 14

40 ± 14

 EF < 50% (n = 37)

86 ± 40

97 ± 44

97 ± 44

88 ± 40

100 ± 43

99 ± 44

EF (%)

 All (n = 100)

52 ± 12

51 ± 12

53 ± 13

51 ± 13

52 ± 12

54 ± 13

 EF ≥ 50% (n = 63)

59 ± 5

58 ± 5

61 ± 5

58 ± 6

59 ± 5

62 ± 5

 EF < 50% (n = 37)

37 ± 9

36 ± 8

37 ± 9

35 ± 10

36 ± 9

38 ± 9

EDSI

 All (n = 100)

33 ± 10

38 ± 10*

39 ± 10^

33 ± 10

38 ± 10*

39 ± 10^

 EF ≥ 50% (n = 63)

30 ± 8

36 ± 8*

36 ± 8^

29 ± 7

35 ± 8*

36 ± 7^

 EF < 50% (n = 37)

38 ± 11

42 ± 11*

42 ± 10^

41 ± 11

46 ± 11*

48 ± 11^

ESSI

 All (n = 100)

27 ± 11

31 ± 10*

30 ± 10^

27 ± 10

31 ± 10*

30 ± 11^

 EF ≥ 50% (n = 63)

23 ± 7

27 ± 6*

26 ± 6^

23 ± 5

27 ± 6*

26 ± 6^

 EF < 50% (n = 37)

33 ± 13

38 ± 12*

38 ± 12^

36 ± 12

40 ± 12*

40 ± 12^

SDI 16 (%)

 All (n = 100)

6.0 ± 2.6

4.8 ± 2.5*

2.1 ± 2.8^§

5.9 ± 3.0

4.9 ± 3.0*

2.3 ± 3.6^§

 EF ≥ 50% (n = 63)

5.2 ± 1.7

4.1 ± 1.7*

1.3 ± 1.1^§

4.9 ± 2.1

3.8 ± 2.0

1.1 ± 0.7^§

 EF < 50% (n = 37)

8.1 ± 3.2

6.3 ± 3.4*

4.0 ± 4.4^§

8.6 ± 3.6

7.5 ± 4.0

5.5 ± 6.3^§

SDI 17 (%)

 All (n = 100)

6.7 ± 3.3

5.4 ± 3.2*

2.9 ± 3.7^§

6.6 ± 3.1

5.6 ± 3.5*

2.9 ± 4.0^§

 EF ≥ 50% (n = 63)

5.6 ± 2.0

4.4 ± 1.9*

1.7 ± 1.7^§

5.5 ± 2.3

4.4 ± 2.3

1.5 ± 2^§

 EF < 50% (n = 37)

9.4 ± 4.3

7.8 ± 4.7*

5.7 ± 5.2^§

9.4 ± 3.4

8.7 ± 4.3

6.5 ± 5.9^§

3DE three dimensional echocardiography, EDV end-diastolic volume, ESV end-systolic volume, EF ejection fraction, EDSI end-diastolic sphericity index, ESSI end-systolic sphericity index, SDI 16 LV 16-segment systolic dyssynchrony index, SDI 17 LV 17-segment systolic dyssynchrony index, HighS high sensitivity level (75%) of automated border detection, MidS medium sensitivity level (50%), LowS low sensitivity level (25%)

P < 0.05 for MidS versus HighS comparison. ^ P < 0.05 for High versus LowS comparison. § P < 0.05 for MidS versus LowS comparison

Impact of sensitivity settings of automated border detection on LV volumes and EF

For both 2- and 4-beat data sets, lowering the sensitivity setting of the automated endocardial border detection algorithm from HighS to MidS or to LowS induced a significant increase in measured EDV and LV sphericity indices (EDSI, ESSI), while having no significant impact on ESV and EF (Table 2, Fig. 1). In 4-beat data sets, the bias in measured EDV induced by changing from HighS to MidS was −16 ml (LOA −32 to 1 ml), significantly higher than the bias from MidS to LowS −2 ml (LOA −9 to 6 ml) (P < 0.0001) (Fig. 2). The same impact was also shown for 2-beat acquisitions: EDV bias −18 ml (LOA −37 to 0 ml) from HighS to MidS versus bias −4 ml (LOA −17 to 9 ml) from MidS to LowS (P < 0.0001).
https://static-content.springer.com/image/art%3A10.1007%2Fs10554-011-9985-0/MediaObjects/10554_2011_9985_Fig1_HTML.gif
Fig. 1

Impact of lowering sensitivity level for the automated detection algorithm of LV endocardial contour in a patient with idiopathic cardiomyopathy, mild LV systolic dysfunction and left bundle branch block. For each of the three sensitivity settings (HighS, MidS, LowS), corresponding LV parameters and regional volumetric curves are shown. Note the major changes, especially for LV end-diastolic volume (EDV), end-diastolic (EDSI) and end-systolic (ESSI) sphericity indices, and 16- and 17- segment systolic dyssynchrony indices (SDI16, SDI17) when comparing HighS to LowS measurements

https://static-content.springer.com/image/art%3A10.1007%2Fs10554-011-9985-0/MediaObjects/10554_2011_9985_Fig2_HTML.gif
Fig. 2

Two-dimensional cut-plane obtained by transversal slicing a 3D data-set of the left ventricle at papillary muscle level showing the effect of the three sensitivity settings chosen (HighS; MidS; and LowS). The increase in end-diastolic area induced by changing from HighS to MidS was significantly larger than that induced by changing from MidS to LowS

The decrease in software sensitivity from HighS to MidS and from MidS to LowS (while maintaining all the other settings identical) led to a significant reduction of mean SDI values (Table 2, Fig. 3). Bland–Altman analysis of various sensitivity levels applied to 4-beat data sets showed a systematic decrease of both SDI 16 and SDI 17 when sensitivity settings were lowered: bias 1.2% for SDI 16 HighS versus MidS (LOA −3.0 to 5.5%) and bias 3.9% for SDI 16 HighS versus LowS (LOA −1.8 to 9.6%); bias 1.3% for SDI 17 HighS versus MidS (−3.5 to 6.1%) and bias 3.9% for SDI 17 HighS versus LowS (LOA −1.8 to 9.6%). The same effect was also demonstrated for 2-beat data set analysis: bias 1.0% for SDI 16 HighS versus MidS (LOA −3.2 to 5.2%) and bias 3.6% for SDI 16 HighS versus LowS (LOA −2.8 to 9.3%); bias 0.9% for SDI 17 HighS versus MidS (−3.8 to 5.6%) and bias 3.6% for SDI 17 HighS versus LowS (LOA −2.1 to 9.4%).
https://static-content.springer.com/image/art%3A10.1007%2Fs10554-011-9985-0/MediaObjects/10554_2011_9985_Fig3_HTML.gif
Fig. 3

Box-and-whisker plots showing the effect of lowering software sensitivity in automated border detection of LV endocardium on SDI 16 (left panel) and SDI 17 (right panel): high (75%, HighS), intermediate (50%, MidS), low (25%, LowS). In black patients with normal QRS width; in red patients with left bundle branch block. *P < 0.001 and ^P < 0.0001 compared with the corresponding value at high sensitivity setting

The impact of temporal resolution and sensitivity settings on LV parameters were not significantly different in patients with LV dysfunction in comparison to those with preserved LV EF (Table 2).

Without applying any manual editing to the automatically detected endocardial border, simply lowering the sensitivity setting from HighS to MidS or from HighS to LowS reduced significantly the systematic underestimation of LV volumes by 3DE in comparison with CMR (P < 0.0001 for both EDV and ESV) (Table 3). Despite similar agreement of 2-beat versus 4-beat 3D data sets with CMR for LV parameters, the correlations with CMR were weaker for the 3D indices obtained from lower temporal resolution data sets (p ≤ 0.001 for each corresponding sensitivity level of EDV, ESV and EF) (Table 3).
Table 3

Comparison against CMR values of left ventricular parameters derived from 2- and 4-beat 3DE data sets analyzed using 3 different levels of sensitivity settings of automated border detection

N = 26

3DE versus CMR

4-beat

2-beat

r

Bias

LOA

r

Bias

LOA

EDV

 HighS

0.954

−44

−6/−82

0.850

−45

−6/−83

 MidS

0.950

−25

16/−65

0.876

−22

15/−59

 LowS

0.944

−24

19/−68

0.848

−17

26/−60

ESV

 HighS

0.962

−20

13/−54

0.739

−19

12/−50

 MidS

0.962

−11

21/−43

0.640

−9

26/−44

 LowS

0.959

−13

20/−46

0.577

−9

29/−47

EF

 HighS

0.901

0

10/−10

0.785

2.1

17/−13

 MidS

0.877

−0.7

10/−11

0.593

1.9

19/−15

 LowS

0.861

0.9

11/−12

0.572

3.8

21/−13

SDI 16

 HighS

0.453

0

4.3/−4.3

0.613

−0.5

3.3/−4.3

 MidS

0.692

–1

2.5/−4.4

0.744

−1.4

1.7/−4.5

 LowS

0.581

−3.5

0.2/−7.2

0.455

−3.4

1.6/−8.4

CMR cardiac magnetic resonance, LOA level of agreement. Other abbreviations as in Table 2; P < 0.001 for all

Impact of LV segmentation model on LVMD (SDI 16 vs. SDI 17)

Irrespective of the sensitivity setting used and of whether the measurement were derived from a 2- or a 4-beat acquisition, mean SDI 17 was systematically higher than SDI 16, yet without reaching statistical significance. In the 4-beat data sets, systematic overestimation of SDI 17 in comparison with SDI 16 was similar among different sensitivity settings, however, the interval of agreement widened as the border sensitivity decreased: bias 0.6% (LOA −2.0 to 3.3%) for HighS; bias 0.6% (LOA −3.1 to 4.3%) for MidS; bias 0.8% (LOA −4.0 to 5.6%) for LowS. A similar effect was noted for the 2-beat data sets: SDI 17 versus SDI 16 bias 0.7% (LOA −2.5 to 3.8%) for HighS; bias 0.8% (LOA −3.1 to 4.6%) for MidS; bias 0.7% (LOA −4.1 to 5.4%) for LowS.

Impact of image quality

Applying a dichotomous separation of the patients according to the visually assessed quality of the LV endocardial border (optimal vs. suboptimal), 40 patients had a suboptimal definition of the endocardial border and 60 had an optimal one. There was no significant difference of LV parameters between the two subgroups, except for higher values of HighS LVEF (54 ± 12% vs. 49 ± 13%) and HighS SDI 16 (6.7 ± 2.7% vs. 5.6 ± 2.4%) in patients with suboptimal definition of the endocardium (P < 0.05 for both). When the impact of changing sensitivity setting (HighS vs. LowS) on LV parameters between the two subgroups was compared, SDI variation was significantly higher in patients with suboptimal image quality (bias 5% vs. 3.2% for SDI 16, 4.5% vs. 3.2% for SDI 17, P < 0.05 for both) (Fig. 4). In contrast, in patients with optimal image quality, there was a greater impact of software sensitivity on EDV and ESV (EDV bias −14 ml vs. −19 ml, ESV bias −5 ml vs. −9 ml, P = 0.01 for both), while its effect on LVEF was only marginally significant (EF bias −2.4% vs. −0.4%, P = 0.06).
https://static-content.springer.com/image/art%3A10.1007%2Fs10554-011-9985-0/MediaObjects/10554_2011_9985_Fig4_HTML.gif
Fig. 4

Box-and-whisker plots showing the effect of lowering software sensitivity in automated border detection of LV endocardium on SDI 16 (left panel) and SDI 17 (right panel): high (75%, HighS), intermediate (50%, MidS), low (25%, LowS). In black patients with optimal image quality; in blue patients with suboptimal image quality; *P < 0.001 and ^P < 0.0001 compared with the corresponding value at high sensitivity setting

Reproducibility of LV parameters measured using 3DE

Intra- and inter-observer variability were 7.0 ± 2.5% and 10.1 ± 6.3% for EDV, 8.9 ± 3.5% and 11.2 ± 8.0% for ESV, and 3.8 ± 2.5% and 4.7 ± 3.6% for EF. The reproducibility of SDI 16 on repeated measurements was 10.4 ± 9.6% for intra-observer and 16.2 ± 14.4% for inter-observer variability.

Discussion

The main findings of the present study can be summarized as follows: 1. Manual editing of endocardial border induced an artifactual LVMD in ventricles with normal function and normal QRS width; 2. The increase in temporal resolution per se did not have a significant impact on LVMD, nor on LV volumes, shape or EF measurements; 3. Changes in sensitivity settings of the algorithm for automated detection of LV endocardial borders dramatically impacted on measured LV volumes and LVMD; 4. Adding the 17th segment to the standard 16-segment LV model induced a slight, yet non-significant increase in LVMD; 5. When changing sensitivity settings from HighS to LowS, subsequent SDI variation was significantly higher in patients with suboptimal image quality of 3DE data sets, while the impact on LV volumes was larger in those with optimal LV endocardial border definition.

LV study using 3DE: current status

Since an accurate assessment of LV function is pivotal to clinical decision making, 3DE used for LV quantitation in terms of volumes, geometry, EF and SDI should be evaluated as rigorously as any therapeutic interventions before systematically applying it in everyday clinical practice [27]. For 3DE, the major limitations were assumed to be the lower spatial and temporal resolution of 3D data sets (average 20–25 vps acquired using second-generation scanners), the differences in operator expertise and 3D softwares used, and the inability of 3D software algorithms to detect end-ejection timing in patients with reduced LV EF and low-amplitude volume curves. However, objective evidence on these sources of bias for measuring LV parameters by 3DE (particularly SDI) is currently scarce [1, 9, 11]. Most 3DE studies on LV function did not describe in sufficient detail the manual interventions during offline post-processing of the data set or applied them inconsistently (i.e. “as needed”). The results of the PROSPECT study [28] underlined how measurement variability and analysis inconsistency may crucially impact on the power of echocardiographic parameters to predict CRT response. Identifying potentially avoidable pitfalls beyond inherent technique limitations is crucial for the emerging clinical applications of 3DE and future CRT study design.

Impact of technical factors: practical issues on 3DE software analysis

Our findings obtained in a relatively large study population with wide ranges of LV volumes, ejection fractions and SDIs do not support the lower temporal resolution as a major drawback of second-generation 3DE scanners for LVMD assessment. Since 3DE measures regional volume change and not regional velocities (as tissue Doppler imaging, which requires a high frame rate), it seems likely that echo systems using 20–30 volumes/s are adequate to sample the usual frequency of regional volume curves of <10 Hz [29]. It is also possible that our results may have been confounded by the analysis and comparison of different LV data sets (e.g. 2- and 4-beat consecutive LV full-volume data sets), since SDI reproducibility for same data set analysis was only fair. Moreover, achieving a superior temporal resolution in 4-beat data sets could be offset by stitching together a higher number of subvolumes, with possibly subtle artifacts that fail to catch the eye and lead to data set rejection accordingly.

The only significant influence of temporal resolution was observed as weaker correlations between 2- versus 4-beat volumes with CMR measurements, yet with no significant difference in agreement. Of note, using a high-performance third-generation 3DE scanner, the mean temporal resolution obtained from 4-beat LV full-volume data sets (52 ± 16 vps) was two-fold higher than in most previous studies on the value of SDI as predictor for CRT response. It is possible that increasing temporal resolution beyond a certain level does not add significant benefit to the predictive value of SDI. However, too low temporal resolution as in single-beat LV full-volume acquisitions (around 10–15 vps) is unlikely to capture regional end-systolic timing, especially in dilated LVs [30].

Our study demonstrated that several other technical factors (i.e. manual editing, sensitivity settings and image quality) may be relatively more important than temporal resolution for LV quantitation using 3DE.

We have previously reported that systematic border verification and manual editing are required to increase accuracy in LV volume measurements using 3DE softwares based on automated border detection [13]. However, in the present study, we found that manual editing induced a significant bias in SDI measurement in a small patient subgroup with normal LV function and QRS width, reaching closely to SDI cut-off values reported to predict acute response to CRT [15]. The software used in our study requires frame-by-frame manual displacement of the generated LV cavity contour to fit the actual endocardial border and this cannot be performed uniformly in each view by the operator. Therefore, the risk of inducing an artifactually different timing of minimal regional volumes is inevitable.

Manual tracing of LV border outwards the black-white interface in order to include endocardial trabeculations also proved to be more accurate for volume measurement in comparison with CMR [11]. However, its extent (i.e. how far outside the visible border to trace) relies on operator’s expertise in 3DE offline analysis and it is prone to a high subjectivity, therefore rather difficult to standardize. The same Authors demonstrated on phantom studies that a bearly visible 1 mm change in endocardial border position leads to a shift in LV volume measurement up to 11% of the initial volume [11]. In our study, a significant increase of LV EDV, EDSI, ESSI and LVMD was found when reducing sensitivity setting of automated border detection algorithm among 3 arbitrarily chosen levels (75, 50, 25%, thus excluding extreme limits of 0 and 100%). Volume change was found to be greater in patients with optimal image quality, in whom endocardial border is easier to track. In contrast, in patients with suboptimal quality of 3D data sets, the automated detection of endocardial borders presumably adopts a rougher tracking during sensitivity tuning between LowS and HighS, that may result in the lesser changes noted for LV volumes in comparison with the measurements on optimal image data sets.

Low sensitivity yielded the most accurate measurement of LV volumes in comparison with CMR data. Tracing right on the visible black–white interface and simply lowering the sensitivity border detection seems to be a good alternative to the recommended method (i.e. manually tracing 1–2 mm outside the endocardial border in order to include trabeculations [11]). The first approach could be easier to standardize across centers with various levels of expertise in 3DE.

The 17-segment model that includes the apical cap as a separate segment is currently recommended by the American Heart Association. In agreement with recent evidence [1], our results show that, by adding the 17th segment to the standard 16 model, SDI increases systematically, however, the differences did not reach statistical significance. SDI 17 was similarly affected as SDI 16 by manual editing, sensitivity settings or image quality. Therefore, the same precautions in the 3D analysis should be applied irrespective of the LV segmentation model used. In addition, mean SDI values significantly dropped when changing from HighS to LowS (2–3 times lower), below the normal cut-off values previously reported [1, 14] and this effect was significantly greater in patients with suboptimal image quality. It is possible that LV analysis aiming to assess LVMD may need a different approach than the one required for global LV volume and EF measurements (i.e. no manual editing, standardized settings to minimize operator’s subjectivity and/or avoid the risk of artifactual “synchronicity” noted for low sensitivity settings, standardization of LV segmentation model etc.), however, this remains to be demonstrated.

Limitations

Only 26 patients underwent CMR <1 h apart from 3DE data set acquisition for measuring LV volumes and EF. However, our study was not designed to test the accuracy of 3DE in comparison with CMR, which has been already well documented. Conversely, we used this subgroup to test how changes in sensitivity of the endocardial border detection algorithm could have affected the agreement of 3DE with CMR for LV volumes. Despite the CMR subgroup was rather small, the impact of sensitivity settings was high enough to reach statistical significance. Similarly, the differences induced by applying manual editing in the pilot study were significant and they confirm previous observations [3133]. The modest reproducibility found for SDI is in agreement with previous reports involving commercially available 3D softwares [1, 17]. Finally, a gold standard parameter for LVMD by echocardiography is lacking, therefore the most appropriate practical approach still needs to be defined by prospective CRT studies.

We used LV full-volume data sets obtained by stitching together subvolumes acquired from 2 to 4 consecutive cardiac cycles. Most recent 3D echo machines allow the acquisition of single-beat full-volume data sets of the LV, thus overcoming the potential limitation of creating subtle stitching artifacts, which may contribute to the increase of SDI. However, these echo machines are not widely available, the SDI reference values have been obtained from multi-beat ECG-gated full-volumes and the limited temporal resolution of current single-beat LV full-volumes prevents their use for quantitative purposes [34].

Our findings pertain to the specific ultrasound system and to the vendor-independent software used in the present study. The influence of manual editing and of sensitivity settings for endocardial border identification algorithm on SDI and LV volumes measured with other proprietary softwares provided by other vendors remain to be explored.

Conclusions

Our study underscores the significant impact of manual editing, software border detection sensitivity and image quality on the 3D assessment of LV geometry and function, including LVMD. Conversely, data set temporal resolution and LV segment model did not significantly affect LVMD. Extensive study of the confounders and development of high- performance 3DE softwares are mandatory before confidence in this method is gained and implementation of 3DE is done to manage clinical problems, as patient selection for CRT. Finally, the implementation of uniform technical standards for both acquisition and analysis of LV parameters is crucial for the appropriate design of future 3DE trials.

Acknowledgments

Dr Denisa Muraru was supported by a Research Grant Programme awarded by the European Association of Echocardiography.

Conflict of interest

Dr Denisa Muraru has received unrestricted research funding from GE Healthcare. Dr Luigi P. Badano has received equipment grants from TomTec Imaging Systems and GE Healthcare, and is on the Speakers’ Bureau of GE Healthcare.

Copyright information

© Springer Science+Business Media, B.V. 2011