Transthoracic 3-dimensional echocardiography

Transcription

1 Impact of On-Line Endocardial Border Detection on Determination of Left Ventricular Volume and Ejection Fraction by Transthoracic 3-Dimensional Echocardiography Michael L. Chuang, MS, Raymond A. Beaudin, MS, Marilyn F. Riley, BS, Matthew G. Mooney, MS, Warren J. Manning, MD, Mark G. Hibberd, MD, PhD, and Pamela S. Douglas, MD, Boston and Andover, Massachusetts This study was performed to determine whether use of on-line automated border detection (ABD) could reduce data analysis time for 3-dimensional echocardiography (3DE) while maintaining accuracy of 3DE in measures of left ventricular (LV) volumes and ejection fraction (EF). The study proceeded in 2 phases. In the validation phase, 20 subjects were examined with the use of 3DE and of monoplane 2- dimensional (2D) ABD. Results were compared with the reference standard of magnetic resonance imaging (MRI). In the test phase, 20 subjects underwent two 3DE studies (once with images optimized for visual border definition and once with images optimized for ABD border tracking) and a conventionally used 2D ABD study. For 3DE, volumes and EF were determined with the use of manually traced borders and ABD. Analysis times were recorded with a digital stopwatch. In the validation phase, 3DE and MRI results correlated very well (r = 0.99) without systematic differences. Comparison of 2D ABD with MRI showed good correlation for LV volumes (r 0.90) and EF (r = 0.85) despite significant underestimation. For the test phase, Acoustic Quantification optimized 3-dimensional datasets underestimated end-diastolic volume and EF relative to visually optimized 3-dimensional datasets regardless of whether borders were hand-traced or ABD was used. However, correlations ranged from r = 0.96 to r = 0.98 for LV volumes and 0.88 to 0.91 for LV EF and were superior to those for 2D ABD. Data analysis times decreased moderately with the use of ABD, but scan times increased; total study times were unchanged. Use of on-line ABD with 3DE reduces data analysis time and is more accurate than conventional monoplane 2D ABD but results in underestimation of LV volumes and EF. Additional automated postprocessing techniques may be required to obtain accurate measures, consistently using 3DE in conjunction with on-line ABD. (J Am Soc Echocardiogr 1999;12:551-8.) From the Charles A. Dana Research Institute and the Harvard- Thorndike Laboratory of the Department of Medicine, Cardiovascular Division, Department of Radiology, Beth Israel Deaconess Medical Center and Harvard Medical School, Boston, Massachusetts; and the Hewlett-Packard Company, Andover, Massachusetts. Reprint requests: Pamela S. Douglas, MD, Cardiovascular Division, Beth Israel Deaconess Medical Center, 330 Brookline Ave, Boston, MA pdouglas@caregroup.harvard.edu. Copyright 1999 by the American Society of Echocardiography /99 $ /1/98650 Transthoracic 3-dimensional echocardiography (3DE) has been shown to provide more accurate and reproducible measures of left ventricular (LV) size and systolic function than conventional 2-dimensional echocardiography (2DE). 1-4 However, scanning and data analysis times for 3DE are increased relative to those for 2DE. In particular, analysis generally requires manual delineation of ventricular borders from multiple images by an expert observer. 1-4 The time-consuming and labor-intensive aspects of 3DE limit its transition from research tool to clinical practice despite the substantially increased accuracy offered by 3DE. Methods for automated border detection (ABD), based on analysis of integrated ultrasonic backscatter, provide estimates of LV dimensions and function comparable to those by 2DE with manual analysis, 5-7 though some have reported systematic underestimation relative to cineventriculography 8 or anatomic measures. 9 It is unclear whether such volume underestimation is solely caused by the inherent limitations of transtho- 551

2 552 Chuang et al July 1999 racic 2D imaging, such as ventricular foreshortening in apical views or limited acoustic windows, or whether some aspect of the ABD process (eg, the optimization of imaging for ABD or the border detection algorithm itself) additionally contributes to this effect. This study was performed to determine the effects of the use of ABD in conjunction with a previously validated transthoracic 3DE system for determination of LV volumes and ejection fraction (EF). We hypothesized that first, the previous findings of systematic volume underestimation were at least in part caused by inherent 2D imaging limitations and that the multiple spatially registered views provided by 3DE would improve the accuracy of volumes determined with the use of ABD methods. Second, we hypothesized that 3DE data analysis times could be reduced with use of ABD. Finally, we investigated whether the combination of 3DE and ABD maintains the accuracy of 3DE with manual border delineation. METHODS Study Design The study proceeded in 2 phases: 20 subjects were enrolled in each phase. In the (first) validation phase, LV volumes and EF were determined by use of the 3DE system with manually traced borders and by standard 2D ABD techniques. These results were compared with the reference standard of magnetic resonance imaging (MRI).The validation phase was performed to compare 3DE and conventional 2D ABD results, in the same subjects, against a widely accepted reference standard and to verify that ABD techniques in our hands can produce results in accord with those reported by previous investigators. In the second test phase, two 3DE datasets were collected for each subject, once with imaging optimized for overall visual image quality and once with imaging optimized for ABD endocardial border tracking, to determine whether the combination of 3DE with ABD would yield results comparable to previously validated 3DE with manually traced borders and to compare data analysis times for the two 3DE datasets. Subjects Forty adults were enrolled in the study after providing written informed consent.the study was approved by the hospital committee on human research. Subjects were not screened for echocardiographic image quality, though patients with contraindications to MRI (pacemaker, automatic implantable cardioverter-defibrillator, intracranial or ocular implants, claustrophobia) or non sinus rhythm were excluded.the 20 subjects (16 men, 4 women; ages 24 to 76 years, mean ± SD = 44 ± 16) in the validation phase included 13 patients with abnormal ventricles and 7 healthy volunteers. Of the 13 patients, 7 had coronary artery disease (confirmed by cardiac catheterization) and regional wall motion abnormalities; 2 had severe mitral valve disease and LV dilatation; 2 had significant pericardial effusions with globally impaired LV systolic function; 1 had patent ductus arteriosus, membranous ventricular septal defect, and dilated cardiomyopathy; and there was 1 postpartum dilated cardiomyopathy. Twenty different subjects (15 men, 5 women, ages 30 to 76 years; 53 ± 16) were examined in the test phase of the study and included 15 patients with abnormal ventricles and 5 healthy volunteers.the 15 patients included 6 with coronary artery disease and focal wall motion abnormalities; 1 with atrial septal defect, dilated left ventricle, and globally reduced LV systolic function; and 8 with moderate to severe valvular disease (5 mitral, 3 aortic). Of the patients with valvular disease, 5 had enlarged left ventricles (2 with reduced systolic function); for the other 3, who did not have enlarged left ventricles, 2 had reduced systolic function. Imaging Methods All echocardiographic imaging was performed with either a Sonos 1500 or a Sonos 2500 scanner with a 2.5-MHz phased array transducer (Hewlett Packard Medical Products, Andover, Mass). The commercially available Acoustic Quantification (AQ) feature was used for on-line real-time ABD as needed. Processing for AQ has been described in detail elsewhere, 5 but in brief, backscatter data from each ultrasound A-line are integrated over a 3.2-µs period (compared with conventional imaging with <1 µs integration time) before data are sent to the scan converter for image reconstruction. The longer integration time results in reduced speckle, which aids in discrimination between signals from the blood pool and endocardium.a thresholding procedure is used to separate low-power signals from the blood pool and high-power signals from the myocardium, after which the detected boundaries are displayed in real time as a color overlay on the gray scale B-mode image. Both time gain compensation and lateral gain compensation 10 were used to enhance AQ border tracking. Three-dimensional echocardiography was performed with Sonos 1500 or 2500 with custom software. A commercially available magnetic tracking system (Flock of Birds, Ascension Technologies, Burlington, Vt) was used to record the position and orientation of the ultrasound transducer during scanning. For each 3DE examination, 20 electrocardiographically triggered cineloops widely surveying the myocardium were acquired during brief endtidal breathholds to minimize bulk cardiac motion caused by respiration. After delineation of endocardial borders at end-diastole and end-systole, a minimum-energy deformable shell model was used to compute LV volumes. 11

3 Volume 12 Number 7 Chuang et al 553 Table 1 Comparison of 3DE, 2DE, and 2D AQ with MRI by linear regression and limits of agreement Regression equation r SEE MB ± 95% CI 3DE EDV MRI = 1.01 X ml 2.4 ± 13.5 ml 2D AQ EDV * MRI = 0.86 X ml 45.5 ± 63.8 ml 3DE ESV MRI = 1.00 X ml 0.6 ± 9.1 ml 2D AQ ESV MRI = 0.96 X ml 10.3 ± 58.4 ml 3DE EF MRI = 1.02 X % 0.2% ± 4.5% 2D AQ EF * MRI = 1.03 X % 10.1% ± 17.1% MB, Mean bias; 95% CI, 95% confidence intervals; SEE, standard error of the mean; 3DE and 2DE, 3- and 2-dimensional echocardiography; EDV and ESV, end-diastolic volume and end-systolic volume; EF, ejection fraction; AQ, Acoustic Quantification; MRI, magnetic resonance imaging. * Significant difference from MRI (P.05). Intercept differs significantly from zero. Table 2 Comparison of 3D AQ, H-3D AQ (n = 20), and 2D AQ (n = 17) with STD 3DE by linear regression and limits of agreement Regression equation r SEE MB ± 95% CI 3D AQ EDV * STD = 1.06 X ml 9.9 ± 29.0 H-3D AQ EDV * STD = 1.08 X ml 11.8 ± D AQ EDV * STD = 0.72 X ml 15.2 ± D AQ ESV STD = 0.93 X ml 3.1 ± 21.4 H-3D AQ ESV STD = 0.96 X ml 0.4 ± D AQ ESV STD = 0.94 X ml 3.0 ± D AQ EF * STD = 0.69 X % 5.5 ± 15.6 H-3D AQ EF * STD = 0.72 X % 3.9 ± D AQ EF * STD = 0.72 X % 6.8 ± 20.1 MB, Mean bias; 95% CI, 95% confidence intervals; SEE, standard error of the mean; 3DE and 2DE, 3- and 2-dimensional echocardiography; EDV and ESV, end-diastolic volume and end-systolic volume; EF, ejection fraction; AQ, Acoustic Quantification. * Significant difference from STD (P.05). Significant differences from 0.0 and 1.0 for intercept and slope, respectively. Table 3 Comparison of 3D AQ with H-3D AQ by linear regression and limits of agreement Regression equation r SEE MB ± 95% CI EDV H-3D AQ = D AQ ml 1.9 ± 16.0 ESV H-3D AQ = D AQ ml 2.8 ± 15.6 EF H-3D AQ = D AQ % 1.6 ± 13.2 MB, Mean bias; 95% CI, 95% confidence intervals; SEE, standard error of the mean; 3DE and H-3DE, 3-dimensional echocardiography and hand-traced 3DE; EDV and ESV, end-diastolic volume and end-systolic volume; EF, ejection fraction; AQ, Acoustic Quantification. MRI was performed on a 1.5 T whole-body scanner (Gyroscan ACS/NT, Philips Medical Systems, Best, The Netherlands) with either an anteriorly placed 20 cm circular surface coil or a 5-element cardiac array coil for radiofrequency signal reception. A breathhold cine-sequence 12 was used to acquire 16 slices covering the left ventricle in the cardiac short-axis orientation. Imaging parameters included TR = 11 ms,te = 5.2 ms, FA = 25 degrees, data matrix, mm 2 field of view, with slice thicknesses ranging from 5.0 to 7.0 mm as needed to encompass the left ventricle. Endocardial borders were manually traced, and volumes were determined by use of a summation of disks (sum of slice areas multiplied by slice thickness) method. Data Acquisition and Analysis Subjects in the validation phase were studied with the use of MRI and 3DE as described above and with conventional 2D AQ. Imaging in the 3DE examination was optimized for visual endocardial border definition. For 3DE and MRI, endocardial borders were manually traced. Scanning procedure for 2D AQ followed published recommendations. 13 LV volumes were determined and updated on-line from an apical 4-chamber view by use of the single-plane modified Simpson s rule algorithm (commercially supplied with the AQ software package) after the sonographer defined a region of interest (ROI) around the left ventricle. End-diastolic and end-systolic volumes (EDV and ESV) were determined as the average of at least 3 cardiac cycles.

4 554 Chuang et al July 1999 Table 4 Mean (±SD) scanning and data analysis times (in minutes) for STD and 3D AQ data sets Scan time * EDV analysis * ESV analysis STD 7.6 ± ± ± 3.0 3D AQ 13.1 ± ± ± 3.2 STD, Standard echocardiography; 3DE, 3-dimensional echocardiography; EDV and ESV, end-diastolic volume and end-systolic volume; AQ, Acoustic Quantification. * Significant difference between STD and 3D AQ. Figure 1 Dilated left ventricle represented by wireframe display of endocardial contours obtained with the use of 3- dimensional echocardiography and by reconstructed ventricular endocardial surface. Left, Manually delineated endocardial contours (top) and corresponding reconstructed surface (bottom). Calculated end-diastolic volume is ml. Right, Endocardial contours and reconstructed surface for the same ventricle with automated border detection. Calculated end-diastolic volume = ml. In the test phase of the study,subjects underwent 2 consecutive 3DE examinations and a conventional 2D AQ study as described above.for 3DE,a standard study (STD) was performed with imaging optimized for visual endocardial border definition. Data from the STD study were analyzed by manually tracing endocardial borders after identification of end-diastolic (ED) and end-systolic (ES) frames. The operator first traced all visualized ED borders from all 20 component images before tracing any ES borders.the ES borders were then traced without reference to ED borders.a second 3DE study was performed with imaging optimized for AQ border tracking, though ROIs were not used to define the region of the left ventricle.total data acquisition times for STD and AQ-optimized studies were recorded with a digital stopwatch and spanned the time from first placement of transducer on the patient s chest to recording of the final cineloop.the AQ borders were digitally encoded along with the image data and could either be displayed as a color overlay or hidden during analysis. Data from the AQ-optimized 3DE study were analyzed in 2 ways. First, the AQ overlay was displayed and the user selected those segments of the AQ-determined contours visually judged as overlaying LV endocardial borders (the user did not draw any borders).volumes and EF were computed after selection of ED and ES contours and denoted 3D AQ. Later, at a separate time point, after coding image files to mask patient identifiers and previous 3D AQ results, the AQ-optimized datasets were analyzed with AQ borders hidden (not displayed), and endocardial borders were traced by hand as for the STD study.volumes and EF thus determined were denoted H-3D AQ. All 3DE analysis was performed off-line on a computer workstation. Data analysis times for STD and 3D AQ studies were recorded with a digital stopwatch. Statistical Analysis In the validation phase of the study, MRI was used as the reference standard for LV volumes and EF.In the test phase, STD 3DE was used as the reference standard, as it was previously shown 14 to be an unbiased predictor of MRI results with narrow 95% confidence intervals.the various results obtained with the use of AQ were compared with the appropriate reference standard by simple linear regression. Standard (Pearson) correlation was used to summarize the association between AQ results and the reference standard. The F test was used to test the null hypothesis that the AQ methods yield measures of LV volume and EF identical to those obtained by the reference standard. Paired Student-Newman-Keuls tests, to account for multiple comparisons, were used to compare the various AQbased methods to the reference standard and with one another. Bland-Altman analysis 15 was performed to determine systematic bias and limits of agreement between the various echo-aq methods and MRI or STD 3DE. Data collection and analysis times for STD and 3D AQ studies were compared by use of the Student t test. Significance was assessed (at the level of P.05), and all tests were 2-tailed. RESULTS Comparison of Echocardiographic with MRI Measures (Validation Study) Echocardiographic and MRI studies were performed successfully in all 20 subjects.ed LV volumes by MRI ranged from 120 to 375 ml and EFs ranged from 18% to 70%.There was excellent correspondence between 3DE and the reference standard of MRI and good but lesser correspondence between 2D AQ and MRI (Table 1). Regression analysis indicated that 3DE pro-

5 Volume 12 Number 7 Chuang et al 555 Figure 2 Limits of agreement for left ventricular ejection fraction (EF). Results from examinations with imaging optimized for Acoustic Quantification border tracking (3D AQ and H-3D AQ) are compared with results from standard (STD) 3-dimensional echocardiographic examinations in which imaging was optimized for visual endocardial border definition. The 3D AQ values (closed diamonds) were obtained with automated endocardial border detection, whereas borders were manually traced from the same datasets for H-3D AQ (open diamonds). duced standard errors of the estimate (SEE) approximately 3-fold smaller than those of 2D AQ. Regression lines did not differ significantly from the line of identity for 3DE, but there were differences for 2D AQ (EDV and EF), as shown in Table 1. Limits of agreement indicated small systematic differences, or mean biases, of negligible practical importance (<2.5 ml for LV volumes, <1% for LV EF) between 3DE and MRI with narrow 95% confidence intervals. These results suggest 3DE and MRI measures of LV volumes and EF are accurate, unbiased predictors of one another that may be used interchangeably. Conventional 2D AQ tended to underestimate LV volumes, and 95% confidence intervals were substantially wider than those for 3DE (Table 1). Performance of 3DE and 2DE with AQ Borders (Test Study) For the test study, 3DE studies were completed successfully in all 20 subjects. In 3 subjects, 2D AQ was judged unable to track endocardial borders adequately (<75% of endocardial borders visualized in the apical 4-chamber view), and these subjects were excluded from analysis. Results by 2D AQ, 3D AQ, and H-3D AQ were each compared with the reference standard of STD 3DE (Table 2). ED volumes by STD 3DE ranged from 69 to 298 ml and EFs ranged from 32% to 79%. An example of endocardial contours from corresponding STD and 3D AQ studies is shown in Figure 1 and demonstrates that 3DE with the use of AQ borders can accurately characterize ventricular shape and provide results that are qualitatively similar to those obtained with manual border delineation. The correlations between 3DE with AQoptimized images and STD were greater than the correlation between 2D AQ and STD for LV volumes and EF regardless of whether borders were semiautomatically (3D AQ) or manually (H-3D AQ) delineated. Limits of agreement indicated greater agreement between 3D AQ/H-3D AQ and STD than between 2D AQ and STD (Table 2) for LV EDV and EF. For LV ESV, 2D AQ showed a mean bias (systematic offset) comparable to that of 3D AQ, with somewhat narrower 95% confidence intervals, though the performance of H-3D AQ was superior to both 3D AQ and 2D AQ. Comparison of ABD and Manual Border Delineation on AQ-Optimized Images Volumes and EF determined with the use of AQ borders (3D AQ) were compared with results obtained with the use of hand-traced borders on the same images (H-3D AQ).The correspondence between 3D AQ and H-3D AQ was very good (Table 3), yielding similar results when compared with STD (Figure 2). EDV and EF were predicted with slightly greater accuracy by 3D AQ, whereas the performance of H- 3D AQ was slightly superior to 3D AQ for ESV (Table 2). However, there were no statistically significant differences between values for EDV, ESV, and EF regardless of whether endocardial borders were

6 556 Chuang et al July 1999 hand traced or determined with ABD methods on images optimized for AQ border tracking. Systematic differences between 3D AQ and H-3D AQ were small (< 3 ml for volumes and 2% for EF),though 95% confidence intervals for EF were wide at ±13.2%. Examination Times In a comparison of scanning and analysis times between STD and AQ-optimized studies (Table 4), mean analysis time was significantly reduced (P =.05) for end-diastolic images with use of selected AQ borders, but not for end-systolic images (P >.1). Mean scan times were increased with use of AQ (P <.01),but the total of scan and analysis times were not different for the 2 methods: 32.9 minutes (STD) versus 34.1 minutes (AQ) (P >.1). DISCUSSION On-line ABD methods provide rapid measures of LV volumes and EF with accuracies comparable to those achieved with conventional 2DE and manual off-line analysis. However, current, commercially available ABD methods as well as conventional 2DE show considerable variation for individual measures and tend to underestimate ventricular volumes relative to anatomic measures and other established methods.this study was performed to determine whether a combination of ABD with 3DE, which provides multiple, spatially registered views and allows use of partial endocardial border segments in volume calculation, would (1) improve accuracy of ABD-based techniques and (2) reduce data analysis times for 3DE. The validation phase of this study confirmed our previous results indicating that 3DE with manually traced borders and MRI are unbiased predictors of one another, with excellent agreement, and can be used interchangeably to determine LV volumes and EF on a wide range of patients. 14 Additionally, conventional 2D AQ methods in our hands produce results with accuracy comparable to results by other investigators To our knowledge, the test phase of this study represents the first application of 3DE with the use of on-line ABD to determine LV volumes and systolic function in patients with a wide range of cardiovascular pathology. Our data indicate that the combination of ABD with 3DE significantly improves accuracy relative to conventional 2D ABD and reduces data analysis times compared with 3DE with manual border delineation.however,lv EDV and EF determined with the use of 3DE with ABD differed significantly from measures obtained with the use of standard 3DE with manual analysis, and scanning time increased with the use of ABD so that combined scanning and analysis times were similar for standard 3DE and 3DE using ABD. Comparison Between ABD and Manual Border Delineation in 3DE Use of 3DE in conjunction with on-line ABD (3D AQ and H-3D AQ) produced measures of LV volume and EF that correlated well with results by 3DE with hand-traced endocardial borders (STD), with correlation coefficients ranging from r = 0.88 to 0.91 for LV EF and r = 0.95 to 0.98 for LV volume.however,there was better correlation between 3D AQ and H-3D AQ (LV volume: r 0.98, LV EF: r = 0.93) than between either AQ-optimized method and STD 3DE. Similarly, limits of agreement indicated smaller systematic differences (mean bias) and narrower 95% confidence intervals for LV EDV and EF when comparing the two 3DE methods using AQ-optimized images with one another than either AQ-optimized method with 3DE using images optimized for gray scale visualization of endocardial borders. For ESV, pairwise differences among the 3 imaging/analysis methods were comparable. These results suggest that the AQ method of ABD displays endocardial borders that correspond well with those that would be manually drawn by an experienced echocardiographer on the same images in the absence of the AQ color overlay. However, the process of optimizing gains (time gain compensation and lateral gain compensation) for AQ border tracking may displace endocardial borders relative to borders on images obtained with gains optimized for visualization of gray scale B-mode scanning. Comparison with Previous 2-Dimensional ABD Studies Automated endocardial border detection, based on analysis of ultrasonic integrated backscatter, has been shown to produce measures of LV cavity area that agree well with those determined by human observers from conventional echocardiographic images. 5,7 Others have shown that changes in cavity area by ABD correlate well with changes in ventricular volume. 19 Comparison of LV volumes determined with the use of ABD with volumes determined with the use of ultrafast computed tomography, 20 cineventriculography, 8 MRI, and anatomic measures 9 have generally indicated good correlation but systematic underestimation of volumes.these results by previous investigators, which are in accord with and comparable to our 2D AQ results, serve to underscore the limitations of quantitative 2D echocardiography with a single tomographic imaging plane,

7 Volume 12 Number 7 Chuang et al 557 which include uncertainty of image plane position relative to the heart and reliance on assumptions about LV shape and uniform function, for determination of LV volumes and EF. Additionally, differences between conventional (2D) ABD results and the various reference methods may also stem in part from distortion caused by optimization of imaging for ABD. Finally, it should be noted that the current study compared a monoplane 2D ABD technique with MRI and 3DE; the results presented do not necessarily predict performance of biplane 2D echo techniques, with or without the use of ABD. Comparison with Previous Multiplane ABD Studies Other investigators have studied the use of multiple echo imaging planes in combination with ABD. Stewart and colleagues 21 compared LV volumes determined by use of AQ borders from apical views with volumes determined by cine-mri in 12 patients. Echo volumes were computed in 2 ways. First, volumes were determined by forming a solid of rotation: cross-sectional areas, determined from AQ borders, were rotated around a nonintersecting central axis to form wedge-shaped volume elements. Summation of these elements yielded LV volumes. From 1 to 6 views were used to determine volume. Second, an ellipsoid long-axis formula was used to estimate LV volume from each of 6 apical views. Volume was determined as the average of the 6 estimates. Either of these methods using 6 views showed good correlation (r 0.90, with SEEs ranging from 20.3 to 23.4 ml) with MRI, with slightly better performance for ED volumes. There were no differences in accuracy with use of 3 views, but correlations diminished and SEEs increased with the use of only 2 views (r = 0.86, SEE = 28.1 ml) or 1 view (r = 0.75, SEE = 35.0).There did not appear to be systematic differences between multiplane ABD volumes and MRI, though manually analyzed biplane 2DE (V = 0.85 A 0 A 90 L) underestimated volumes relative to ABD and MRI. The results of Stewart et al indicate that combination of ABD with multiple views is promising and increases accuracy of measures of LV volumes, but both volume computation methods are dependent on acquisition of good apical views, which may limit application in some patients. Gorcsan and colleagues 16 were able to obtain AQ data from 3 (basal, midventricular, and apical) shortaxis slices in 66 (75%) of 88 unselected patients. Application of a modified Simpson s rule algorithm revealed good correlation with radionuclide EF (r = 0.91, SEE = 8%) and low interobserver and intraobserver variability ( 6% and 5%, respectively.) Using a custom 3DE system in conjunction with ABD, Jiang et al 22 observed significant and consistent underestimation relative to actual volumes of formalin-fixed animal hearts and to volumes determined by manual tracing, although reconstructed ventricular shapes were qualitatively similar.the 3DE algorithm used required continuous endocardial borders, and the process of adjusting gains to meet this requirement may have contributed to the observed displacement of AQ borders into the ventricular cavity, leading to volume underestimation. In this first study of 3DE combined with AQ border detection, Jiang et al noted that methods to accommodate noncontinuous borders with their 3DE system (eg, interpolating gaps by use of 3D information and prior knowledge) might allow use of lower gain settings and potentially better estimates of endocardial border location by AQ. In contrast, the 3DE system used in the current study does not require continuous borders, yet we also observed systematic underestimation of LV volumes in patients.a further difference in methods is that the study of Jiang et al required that AQ borders be manually traced for use by the 3DE algorithm because it was not possible then to obtain direct digital output of AQ borders, whereas the current study was able to use actual AQ borders obtained directly from the scanner. However, there does not appear to be any practical effect despite this methodological difference.(tracing over AQ borders, as was done by Jiang et al, would effectively smooth or low-pass filter the contours, potentially increasing accuracy. 23 ) Future Directions This study is a preliminary application of AQ borders to 3DE and as such did not incorporate a number of possibilities for improving the performance of AQ with regard to accuracy or potential time savings. Other investigators have shown that additional processing of raw AQ borders, including Fourier curve fitting and smoothing, 23 averaging over multiple cardiac cycles, 24 or model-based contour fitting, improve the accuracy of 2D AQ; such advantages should, in principle, extend to the use of AQ borders in 3 dimensions.with regard to time savings, we have shown elsewhere 25 that LV volumes and EF can be determined accurately with as few as 10 component views by use of the current 3DE system with manual border delineation.the relation between number of views and accuracy of 3D AQ remains to be investigated, but it is possible that fewer than the 20 views used in this study would suffice to maintain the current level of accuracy. Additionally, analysis time might be further reduced with scan-time use of ROIs

8 558 Chuang et al July 1999 to constrain border detection to the region of the LV, though this advantage might come at the expense of increased scanning time. Conclusions The combination of on-line ABD with transthoracic 3DE increases the accuracy of ABD-based methods and reduces the time needed for data analysis in 3DE. However, 3DE with ABD significantly underestimates LV volumes and increases scanning time. Although promising, further work is needed to realize the potential of using on-line ABD methods in conjunction with 3DE. REFERENCES 1. Nosir YF, Fioretti PM, Vletter WB, et al. Accurate measurement of left ventricular ejection fraction by 3-dimensional echocardiography: a comparison with radionuclide angiography. Circulation 1996;94: Siu SC, Rivera JM, Guerrero JL, et al. Three-dimensional echocardiography: in vivo validation for left ventricular volume and function. Circulation 1993;88: Leotta DF, Munt B, Bolson EF, et al. Quantitative 3-dimensional echocardiography by rapid imaging from multiple transthoracic windows: in vitro validation and initial in vivo studies. J Am Soc Echocardiogr 1997;10: Gopal AS, Keller AM, Rigling R, et al. Left ventricular volume and endocardial surface area by 3-dimensional echocardiography: comparison with 2-dimensional echocardiography and nuclear magnetic resonance imaging in normal subjects. J Am Coll Cardiol 1993;22: Perez JE, Waggoner AD, Barzilai B, et al. On-line assessment of ventricular function by automatic boundary detection and ultrasonic backscatter imaging. J Am Coll Cardiol 1992;19: Perrino AC, Luther MA, O Connor TZ, et al. Automated echocardiographic analysis. Anesthesiology 1995;83: Vandenberg BF, Rath LS, Stuhlmuller P, et al. Estimation of left ventricular cavity area with an on-line semiautomated echocardiographic edge detection system. Circulation 1992; 86: Vanoverschelde JL, Hanet C, Wijns W, et al. On-line quantification of left ventricular volumes and ejection fraction by automated backscatter imaging-assisted boundary detection: comparison with contrast cineventriculography. Am J Cardiol 1994;74: Morrissey RL, Siu SC, Guerrero JL, et al. Automated assessment of ventricular volume and function by echocardiography: validation of automated border detection. J Am Soc Echocardiogr 1994;7: Perez JE, Klein SC, Prater DM, et al. Automated, on-line quantification of left ventricular dimensions and function by echocardiography with backscatter imaging and lateral gain compensation. Am J Cardiol 1992;70: Pentland A, Sclaroff S. Closed-form solutions for physically based shape modeling and recognition. IEEE Trans Patt Anal Mach Intell 1991;13: Chuang ML, Chen MH, Khasgiwala VC, et al. Adaptive correction of imaging plane position in segmented K-space cine cardiac MR imaging. J Magn Reson Imaging 1997;7: Bednarz JE, Marcus RH, Lang RM. Technical guidelines for performing automated border detection studies. J Am Soc Echocardiogr 1995;8: Hibberd MG, Chuang ML, Beaudin RA, et al. Three-dimensional echocardiography improves reproducibility and accuracy of estimates of ventricular ejection fraction [abstract]. Circulation 1996;94(suppl I):I Bland JM, Altman DG. Statistical methods for assessing agreement between 2 methods of clinical measurement. Lancet 1986;1: Gorcsan JD, Lazar JM, Schulman DS, et al. Comparison of left ventricular function by echocardiographic automated border detection and by radionuclide ejection fraction. Am J Cardiol 1993;72: Lindower PD, Rath L, Preslar J, et al. Quantification of left ventricular function with an automated border detection system and comparison with radionuclide ventriculography. Am J Cardiol 1994;73: Yvorchuk KJ, Davies RA, Chan KL. Measurement of left ventricular ejection fraction by acoustic quantification and comparison with radionuclide angiography. Am J Cardiol 1994; 74: Gorcsan JD, Morita S, Mandarino WA, et al. Two-dimensional echocardiographic automated border detection accurately reflects changes in left ventricular volume. J Am Soc Echocardiogr 1993;6: Marcus RH, Bednarz J, Coulden R, et al. Ultrasonic backscatter system for automated on-line endocardial boundary detection: evaluation by ultrafast computed tomography [see comments]. J Am Coll Cardiol 1993;22: Stewart WJ, Rodkey SM, Gunawardena S, et al. Left ventricular volume calculation with integrated backscatter from echocardiography. J Am Soc Echocardiogr 1993;6: Jiang L, Morrissey R, Handschumacher MD, et al. Quantitative 3-dimensional reconstruction of left ventricular volume with complete borders detected by acoustic quantification underestimates volume. Am Heart J 1996;131: Chandra S, Garcia MJ, Morehead A, et al. Two-dimensional Fourier filtration of acoustic quantification echocardiographic images: improved reproducibility and accuracy of automated measurements of left ventricular performance. J Am Soc Echocardiogr 1997;10: Mor-Avi V, Gillesberg IE, Korcarz C, et al. Improved quantification of left ventricular function by applying signal averaging to echocardiographic acoustic quantification. J Am Soc Echocardiogr 1995;8: Danias PG, Chuang ML, Parker RA, et al. Relation between the number of image planes and the accuracy of 3-dimensional echocardiography for measuring ventricular ejection fraction. Am J Cardiol 1998;82: