Automated Assessment of Ventricular Volume and Function by Echocardiography: Validation of Automated Border Detection

Transcription

1 ORIGINAL ARTICLES Automated Assessment of Ventricular Volume and Function by Echocardiography: Validation of Automated Border Detection Richard L. Morrissey, MD, Samuel C. Siu, MD,]. Luis Guerrero, BS, John B. Newell, BS, Arthur E. Weyman, MD, and Michael H. Picard, MD, Boston, Massachusetts To determine the utility of a new on-line echocardiographic automated border detection (ABD) algorithm in assessing ventricular volume and ejection fraction, an optimal model was studied. This open-chest canine model allowed continuous measurement of actual left ventricular volume. In four dogs, true end-systolic and end-diastolic volume and ejection fraction were compared with those obtained by two-dimensional echocardiography with an automated method calculated from a border detection algorithm to define left ventricular endocardium and the single-plane Simpson method to calculate volume. Left ventricular volumes that used manual, off-line tracings of the left ventricle by two-dimensional echocardiograms and the single-plane Simpson method were compared. The automated echocardiographic volumes correlated with true volumes (y = 0.7x + 8.9; standard error of the estimate = 13.5 cc; r = 0.81). A significant mean underestimation of ll ± 15 cc was noted (p < ). Volumes obtained from the manual tracings of left ventricular endocardial contours also correlated well with true volumes (y = 0.89x + 4; standard error of the estimate = 6.7 cc; r = 0.96). However, the 3 ± 7 underestimation was significantly lower than the error of the ABD method (p = ). Both on-line ABD and off-line ejection fractions correlated well with true ejection fractions (r = 0.94 and 0.96, respectively). There was no statistically significant difference between the mean errors of the ABD or manually derived ejection fractions. In the setting of optimal left ventricular imaging, the on-line and rapid features of this automated method make it potentially useful for quickly obtaining left ventricular volumes and ejection fraction.(] AM Soc EcHOCARDIOGR 1994;7: ) Quantitation of left ventricular function is critical for many aspects of clinical decision making. 1-4 The two most widely used indexes of left ventricular function are chamber volume and ejection fraction. Twodimensional echocardiography provides dimensional and area measurements noninvasively that can be integrated into a variety of geometric models to calculate these parameters. 5 However, until recently even the simplest methods for analysis of two-di- From the Cardiac Unit of the Massachusetts General Hospital and Harvard Medical School. Presented in part at the Forty-second Annual Scientific Session of the American College of Cardiology, Anaheim, California, March 17, Reprint requests: Michael H. Picard, MD, Cardiac Ultrasound Laboratory, VBK 508, Massachusetts General Hospital, Fruit St., Boston, MA Copyright " 1994 by the American Society of Echocardiography /94$ / mensional images were dependent on off-line review and measurement by the operator. Because of the time required to perform these tasks, the full quantitative capabilities of echocardiography were not used routinely and it was not feasible to assess changes in ventricular volume and performance as they occurred. Recent technologic advances have made it possible to detect ventricular borders automatically from thresholded-integrated backscatter images. 6 This approach offers the potential for rapid, on-line calculation of cardiac areas and, through further computer processing, volumes and ejection fraction. 7 8 However, the accuracy of these measurements has yet to be fully determined. The aim of this study was to assess the accuracy of left ventricular volumes and ejection fractions calculated by a new on-line echocardiographic method for automated detection of ventricular borders. The optimal experimental setting was chosen to deter- 107

2 108 Morrissey eta!. Journal of the American Society of Echocardiography March-April 1994 LV Pressure Catheter Fitter, Oxygenator and Heat Exchonger Roller Pump Figure 1 Experimental model created to allow instantaneous quantification of left ventricular volume. Ao, Aorta, Col, column; IVC, inferior vena cava; LA, left atrium; LV, left ventricle; LV BAL, left ventricular cavity balloon; PA, pulmonary artery; RA, right atrium; RCA, right coronary artery; R V, right ventricle; SVC, superior vena cava. mine both the accuracy of the system and reasons for discrepancies, should they occur. METHODS Experimental Model To compare instantaneous left ventricular volume determined by the automatic method with actual volume, a previously described canine model was modified (Figure 1). 9 Under complete general anesthesia and after endotracheal intubation, a left thoracotomy was performed and a pericardia! cradle was created. The inferior vena cava, superior vena cava, right atrium, and right ventricle were cannulated and venous return was diverted to a pump oxygenator. Oxygenated blood was returned to the ascending aorta and the femoral arteries to maintain viability of the animal preparation. Right ventricular epicardial pacing electrodes were placed. The pulmonary artery and aortic root proximal to the coronary ostia were ligated. An incision was made in the left atrium and a latex balloon connected to a gradm1ted extracardiac reservoir through a mitral annular ring was introduced into the left ventricle through the mitral valve. Thus the left ventricular cavity was isolated from the remainder of the circulation. To ensure that the balloon conformed maximally to the contour of the left ventricle, chordae tendineae attaching the papillary muscles to the mitral valve were excised. The thebesian venous return into the left ventricle was drained by a small (24-gauge) cannula inserted at the ventricular apex. This ensured that in all cases the balloon-endocardial interface was contiguous. The balloon reservoir system was filled incrementally with known aliquots of colored water. By knowing the volume of the fluid in the column and the total volume of the fluid placed in the system, the instantaneous left ventricular volume could be determined. The end-systolic volume was taken as the total volume in the system minus the volume in the column at the point at which the column of water was at its maximum. Likewise, the end-diastolic volume was taken as the total volume in the system minus the

3 Journal of the American Society of Echocardiography Volume 7 Number 2 Morrissey et al. 109 column volume at the point at which the column of water was at its minimum. Experimental Protocol With the beating canine heart model, each animal was studied in a series of stages with left ventricular volume ranging from 12 to 127 cc (mean 67 cc). Fifty-four left ventricular volume stages in four dogs (weights ranging from 25 to 39 kg) were assessed. At each stage, two-dimensional echocardiographic imaging was performed simultaneously with recordings of systolic and diastolic column height. After each volume stage, known aliquots of fluid were added to the model and the process was repeated. On average, 14 stages were performed in each animal. Studies conformed to the guidelines of the "Position of the American Heart Association on Research Animal Use" and were approved by the Massachusetts General Hospital Subcommittee on Research Animal Care. Echocardiographic Data Acquisition At each stage, left ventricular apical images were obtained with a Hewlett-Packard Sonos 1500 equipped with acoustic-quantification hardware and software for automated border detection (ABD) and left ventricular volume calculation (Hewlett-Packard Co., Andover, Mass.). Echocardiographic imaging of the open-chest preparation was performed with the aid of a water bath for acoustic coupling and transducer standoff; 2.5, 3.5, and 5 MHz transducers were used, with the final choice of transducer frequency determined by optimal endocardial resolution. Care was taken to ensure that only apical views representing the true long axis were selected for analysis. When the ABD program was activated, the individual returning scan-line signals were processed to determine integrated backscatter amplitude. The signallevel was then thresholded to separate the highamplitude signals originating from tissue and the low-amplitude signals originating from the blood pool. The fluid-tissue/balloon interface was then automatically identified by a highlighted overlay that tracked this interface throughout the cardiac cycle. 7 8 A region of interest that included this border and extended from the medial aspect of the mitral annulus to its lateral aspect was defined by the operator. Echocardiographic Volume Calculation Once the region of interest was optimized by the operator, the algorithm then determined the long axis of this region as the length of a line extending from the bisector of the base to the most distal point of the region of interest. This length was assumed to traverse the ventricular apex (Figure 2) and could be modified by the operator. The long-axis line was divided into 20 equal segments. The length of the line drawn perpendicular to the long axis (medial to lateral) and extending to the contralateral highlighted border at each of these 20 points identified the ventricular diameter at that level. The length of each of these lines I diameters falling within the highlighted borders was used to calculate volume by the singleplane modified Simpson method. The left ventricle was assumed to be symmetric about the long axis and thus the product of the measured diameter times the distance between the origins of the 20 individual segments was used to calculate the volume of circular disks, which were summed to determine total left ventricular voltime. This volume was tracked continuously through the cardiac cycle. In addition, updated end-diastolic (maximum) and end-systolic (minimum) volumes were displayed graphically along with the ejection fraction for each cycle (Figure 3). Once the image was optimized and stabilized, an average of 20 consecutive end-diastolic and end-systolic volumes was taken for each stage. At each stage, standard two-dimensional left ventricular images without ABD were also obtained for subsequent off-line digitized manual tracing and volume calculation according to the single-plane modified Simpson method Five off-line tracings were averaged for each stage. Statistical Analysis True left ventricular volumes and ejection fractions determined from the experimental model were compared directly by linear regression analysis to both the on-line values from the automated system and the off-line values determined from manual tracings. The differences of each calculated value from the true value were calculated as a measure of error. The mean errors of the automated and manual methods were compared by t tests. Correlation coefficients of the regression equations were compared by Z transformation. A p value < 0.05 was required for significance. RESULTS Left Ventricular Volume Left ventricular volumes were analyzed during 2 7 paired systolic and diastolic stages. The mean diastolic volume was 7 4 ± 25 cc and the mean systolic volume was 55 ± 26 cc. As observed in Figure 4,

4 llo Morrissey et al. Journal of the American Society of Echocardiography March-Aprill994 Figure 2 Apical view of left ventricle. ABD mode is engaged that highlights tissue-cavity interface. In this example, operator-defined region of interest (ROI) and machine-determined long-axis length (LAX) are also shown. Figure 3 Apical view of left ventricle with ABD determinations of left ventricular volume. Below image, continuous tracking of left ventricular volume is graphically displayed. At right of graph, updated maximum volume-end diastolic (EDV), minimum volume-end systolic (ESV), and ejection fraction (EF) are displayed.

5 Journal of the American Society of Echocardiography Volume 7 Number 2 Morrissey et al. Ill 12o y ===o:7x r = () 100 see = cc ()... w 80. ~ ~ :::::> > Q Ill <( :,...-< r.., ~ ~ TRUE VOLUME (cc) Figure 4 Correlation of left ventricular by on-line ABD and true volume. SEE, Standard error of estimate; --, line of identity; --regression line y = 0.89x + 4 () 100 r = 0.96 () - SEE= 6.7 cc w ~ :::::> > 40 w Q. :;;.o ~ C\J ~=~; ~ TRUE VOLUME (cc) Figure 5 Correlation of left ventricular volume by off-line two-dimensional echocardiographic (2DE) method and true volume. SEE, Standard error of estimate; --, line of identity; --, regression line. ABD volumes correlated with true volumes (Y = 0.7x + 8.9; standard error of the estimate = 13.5 cc; r = 0.81). A significant mean underestimation of ll ± 15 cc was noted (p < ), which corresponded to a mean percent error of -13.8% ± 24.7%. Comparison of left ventricular volumes determined by the off-line method with true volumes is shown in Figure 5. Off-line two-dimensional echocardiographic volumes correlated well with true volumes (Y = 0.89x + 4; standard error of the estimate = 6.7 cc; r = 0.96). This correlation of the manual, off-line two-dimensional method was significantly better than the ABD method (p = ). As in

6 112 Morrissey et a!. Journal of the American Society of Echocardiography March-Aprill ~ ~ LL w 0 (l) <( y = 0.76x / 80 r = 0: ,... / SEE= 4.3% ~ ~ ~ TRUE EF (%) Figure 6 Correlation of ejection fraction (EF) by on-line ABD and true ejection fraction. SEE, Standard error of estimate; --,line of identity; --,regression line. the ABD-derived volumes, off-line two-dimensional derived volume underestimated true volume. However, the 3 ± 7 cc of underestimation was significantly lower than the error of the ABD method (p = ). This error represented a mean underestimation of -3.2% ± 13.7%. Both the error between ABD volume and true volume and the error between off-line volume and true volume were analyzed as a function phase of the cardiac cycle. There was no significant correlation between error and phase of the cardiac cycle. Ejection Fraction As demonstrated in Figure 6, ABD ejection fraction correlated well with true ejection fraction (y = 0.76x + 5.2; standarderroroftheestimate = 4.3%; r = 0.94). Figure 7 shows a comparison of off-line ejection fraction with true ejection fraction. Again, an excellent correlation was noted (y = 0.89x + 2.3; standard error of the estimate 4.9%; r = 0.96). There was no statistically significant difference between the mean errors or the correlation coefficients of the ABD and manually derived ejection fractions. DISCUSSION At present, accurate methods to assess left ventricular volume rely on either invasive or time-consuming non-invasive techniques. Although echocardiography has proved to be an accurate method for calculating volume, it is operator dependent and timeconsuming because of the need to digitize the endocardial contours manually during systole and diastole. Previous methods that used computer-automated edge-detection algorithms have been described and validated; however, immediate and continuous on-line volume and ejection fraction data have not been available Echocardiographic online ABD is now possible by thresholding levels the received ultrasound signal that defines the blood-endocardial interface. Once the borders of the left ventricle are defined, a variety of algorithms can be used to assess left ventricular dimensions, area, volume, and ejection fraction. 7 8 In this study an experimental model was used to assess the accuracy of ABD to calculate left ventricular volume and ejection fraction in an optimal, controlled fashion. This open-chest canine model provided ideal image quality along with a precise independent method for instantaneous continuous volume measurement. With this model, ABD-derived left ventricular volume correlated with true volume, although it underestimated true volume to a greater degree than an off-line method that used manually traced digitized borders entered into a Simpson rule algorithm. Echocardiographic Underestimation of Volume Both the ABD and manually digitized echocardigoraphic methods tested yielded underestimation of

7 Journal of the American Society of Echocardiography Volume 7 Number 2 Morrissey et al. 113 LL w w 0 C\J 100 ~ ~ y = 0.89x r-= 0;96 - SEE= 4.9% , ,; ;, :"" o ~ TRUE EF (%) Figure 7 Correlation of ejection fraction (EF) by off-line two-dimensional echocardiographic (2DE) method compared with true ejection fraction. SEE, Standard error of estimate; --,line of identity;--, regression line. left ventricular volume. This is not surprising and is in agreement with prior studies The echocardiographic underestimation of left ventricular volume is the result of several factors. The primary cause of underestimation has been proved to be caused by point-spread function and beam width. Especially when imaging from the apical window, there is significant beam-width spread that causes echoes from the endocardium to impinge on the left ventricular cavity, creating an artifactually smaller left ventricle.5 17 Underestimation of left ventricular volume can also be enhanced by foreshortening the left ventricular long axis. This underestimation occurs when the long-axis plane fails to pass through the long axis of the ventricle. Therefore one must ensure that the true long axis is obtained. 18 In this open-chest model the external tip of the cardiac apex was readily visualized and therefore imaging of the true long axis was easily achieved. However, in clinical practice, to ensure that the left ventricle is imaged along its long axis, multiple views should be assessed to find the longest ventricular length along the line that bisects the base.u Finally, both overestimation and underestimation can occur if inappropriate geometric assumptions are made. Use of a single-plane formula may not reflect the true three-dimensional shape of the ventricle. This inaccuracy occurs most commonly in the irregularly shaped ventricle. In this experimental model, only normally shaped ventricles were studied. On-line ABD Underestimation of Volume in This Model Because the foundation of the ABD on-line method relies on two-dimensional echocardiographic imaging, all the potential factors that cause volume underestimation with conventional two-dimensional echocardiography noted above will similarly cause underestimation with the on-line ABD method. In addition, further underestimation can be attributed to the ABD algorithm, which uses the inside of the imaged blood-tissue interface to define the border of the left ventricular cavity. This is in contrast to the standard off-line manual tracing technique that uses the middle of the bright echoes originating from the blood-endocardial interface. 19 The increased degree of underestimation noted in the on-line automated method may also be explained in part by the fact that the ABD algorithm identifies as an interface any portion of the image where a significant change in threshold signal has occurred. Real structures or artifacts within the left ventricle will be detected and decrease the true diameter I area at that level. Thus if papillary muscles, chords, or spontaneous echo contrast have a sufficient intensity threshold, they will be considered as borders and any volume of blood between them and the true left ven-

8 114 Morrissey et a!. Journal of the American Society of Echocardiography March-April 1994 tricular wall will be excluded from the volume calculation. In our experimental model this factor was reduced by the use of an intracavitary balloon that excluded subvalvular structures; however, spontaneous echo contrast was occasionally noted in the cavity. Potential Pitfalls Potential pitfalls that must be considered when using the ABD method to assess left ventricular volume include the dependence on image quality and clear border definition, operator-dependent factors including defining an appropriate region of interest, and appropriate use of instrument settings For example, when both increased gain and low gain have been examined systematically, underestimations and overestimations in volume on the order of 50% have been observed. 21 In clinical situations in which the endocardium may not be as well defined as in this open-chest model, overestimation of volume may occur in regions of echo dropout. This occurs because the ABD method is dependent on instrument gain for identification of structures. Thus received signals of low intensity from some walls may be falsely categorized as originating from the blood pool and the chambers will appear falsely expanded. However, this will be a problem for both on-line and off-line methods. Perez et al. 7 have reported that 72% of patients had adequate imaging to permit automatic tracking of endocardial-blood interfaces. In this experiment we used an optimal model in which only symmetrically shaped left ventricles were studied so that high-quality images could be uniformly obtained. Although a balloon was used to allow assessment of true instantaneous volume, this did not deter and may have enhanced the ABD system's ability to track the boundary. Imaging of the heart before balloon placement demonstrated excellent tracking and appropriate highlighting of the endocardium. Another operator-dependent factor that will influence the accuracy of the ABD method is positioning of the region of interest. This region is an operatordelineated search area in which the algorithm operates. Both underestimation and overestimation of volume can occur as a result of improper region of interest placement. If the blood pool to be measured moves outside the boundaries of the operator-defined region of interest during part of the cardiac cycle, this region is not included in the volume calculated and the volume will be underestimated. In this model this potential pitfall was avoided because of limited motion of the heart. Care was also taken to draw the region of interest large enough to ensure that the left ventricular cavity remained within the region of interest throughout the cardiac cycle. Likewise, if additional areas such as the left atrium are included within the region of interest, they will be included in the calculation and the volume will be overestimated. In this model the mitral valve was excised and replaced with an annular ring to which the balloon was attached. Therefore the potential for overestimation was reduced by our ability to identify the mitral annulus and to exclude the left atrium from the region of interest. In addition, the transmit gain settings were increased at the level of the left atrium to ensure that the left atrial cavity would not be detected by the ABD algorithm and included in the volume calculation. Where the portion of the region of interest bordered the right ventricular cavity, lateral gain compensation was used and adjusted selectively to ensure that the right ventricular cavity would not be detected. Ejection Fraction In this experiment, ejection fraction calculated by ABD showed an excellent correlation with true ejection fraction and no difference in accuracy from offline methods. An explanation for this is that any systematic errors of measurement present for both end-diastolic volume and end-systolic volume are decreased when a ratio such as the ejection fraction is determined. 11 This may also explain why other investigators have demonstrated a good correlation between ABD-derived ejection fraction and radionuclide-derived ejection fraction Our results suggest that, although caution needs to be applied when interpreting the on-line volume calculations because of the many potential sources of variability, the ejection fraction calculations may have more utility because of cancellation of any systematic errors. Limitations A potential limitation of the model might appear to be that on-line ABD volumes and off-line volumes were not calculated from the same cardiac cycle. To reduce or eliminate this variability, the on-line ABD measurements and the two-dimensional images recorded for subsequent off-line tracing were obtained sequentially and at steady state. In addition, the two methods were not compared with each other directly but with a known true volume. A second limitation when attempting to apply these results to clinical applications is that in this model most of the mitral apparatus was removed and replaced by a ring that improved the definition of the base of the left ven-

9 Journal of the American Society of Echocardiography Volume 7 Number 2 Morrissey et a!. ll5 tricle. Additional variability in ABD volume may occur in clinical situations in which the motion and definition of the ventricular base are more variable. Conclusions Despite an additional loss of accuracy in volume calculation, the on-line and rapid features of this automated method make it potentially very useful, particularly in settings in which images of the left ventricle are optimal. In our experimental model, ABD permitted rapid assessment of ventricular volume and performance. Recording left ventricular volumes and ejection fractions on-line by ABD therefore has the potential to improve the serial noninvasive assessment of interventions aimed at influencing left ventricular volume and performance. It remains to be determined whether the degree of correlation and variability of the automated method, when applied to clinical images, will be within acceptable limits. We acknowledge the advice and technical assistance of Mr. David Domke, Dr. Stockton Miller-Jones, and Dr. S. Mark Nidor REFERENCES l. Ross J Jr. Left ventricular function and the timing of surgical treatment in valvular heart disease. Ann Intern Med 1981;94: Schuler G, Peterson KL, Johnson A, eta!. Temporal response of left ventricular performance to mitral valve surgery. Circulation 1979;59: Alderman EL, Fisher DL, Litwin P, eta!. Results of coronary artery surgery in patients with poor left ventricular function (CASS). Circulation 1983:68: Kirklin JW, Akins CW, Blackstone EH, et a!. ACC/ AHA Task Force Report: guidelines and indications for coronary artery bypass graft surgery. JAm Coli Cardiol1991;17: Weyman AE. Principles and practice of echocardiography. 2nd ed. Philadelphia: Lea & Febiger, 1994: Miller JG, Perez JE, Sobel BE. Ultrasonic characterization of myocardium. Prog Cardiovasc Dis 1985;28: Perez JE, Waggoner AD, Barzilai B, Melton HE Jr, Miller JG, Sobel BE. On-line assessment of ventricular function by automatic boundary detection and ultrasonic backscatter imaging. JAm Coli Cardiol1992;2: Vandenberg BF, Rath LS, Smhlmuller P, Melton HE, Skorton DJ. Estimation ofleft ventricular cavity area with an online, semiautomated echocardiographic edge detection system. Circulation 1992;86: Siu SC, Rivera JM, Guerrero BS, et a!. Three dimensional echocardiography: in vivo validation for left ventricular volume and function. Circulation 1993;88[part 1]: Schiller NB, Shah PM, Crawford M, eta!. Recommendations for quantitation of the left ventricle by two-dimensional echocardiography. JAM Soc ECHOCARDIOGR 1989;2: Feigenbaum H. Echocardiography. 4th ed. Philadelphia: Lea & Febiger, 1986: Geiser EA, Conerta DA, Limacher MC, Stockton VO, Oliver LH, Jones B. A second-generation computer-based edge detection algorithm for short-axis, two-dimensional echocardiographic images: accuracy and improvement in interobserver variability. JAM Soc ECHOCARDIOGR 1990;3: Collins SM, Skorton DJ, Geiser EA, eta!. Computer-assisted edge detection in two-dimensional echocardiography: comparison with anatomic data. Am J Cardiol1984;53: Garcia E, Gueret P, Bennett M, eta!. Real time computerization of two-dimensional echocardiography. Am Heart J 1981;101: Siu CS, Rivera JM, Xie SW, et a!. Three-dimensional echocardiography improves non-invasive left ventricular volume quantitation [abstract]. JAm Coli Cardiol1992;19:18A. 16. Schnittger I, Fitzgerald PJ, Daughters GT, et a!. Limitations of comparing left ventricular volumes by two-dimensional echocardiography, myocardial markers and cineangiography. Am J Cardiol1982;50: Geiser EA, Skorton DJ, Conetta DA. Quantification of left ventricular function by two-dimensional echocardiography: consideration of factors restricting image quality. Am Heart J 1982;103: Gordon EP, Schnittger I, Fitzgerald PJ, Williams P, Popp RL. Reproducibility of left ventricular volumes by two-dimensional echocardiography. JAm Coli Cardiol1983;2: Handshumacher MD, Lethor JP, Sui SC, eta!. A new integrated system for three-dimensional echocardiographic reconstruction: development and validation for ventricular volume with application in human subjects. J Am Coli Cardiol 1993;21: Smith MD, Xie GY, Sapin PM, et a!. Factors affecting the determination of left ventricular area by acoustic quantification [abstract]. JAm Coli Cardiol1992;19:299A. 21. Cao QL, Azevedo J, Snapper H, Schwartz S, Udelson J, Pandian N. Automated, on-line determination of left ventricular volumes and ejection fraction by acoustic quantification: (1) in-vitro validation, (2) in-vivo comparison with radionuclide method in patients, (3) examination of factors influencing its accuracy, and ( 4) clinical application [abstract]. JAm Coli Cardiol 1993;21:275A. 22. Gorcsan J III, Schulman DS, Koch L, Thornton J, Follansbee WP. Echocardiographic automated border detection derived left ventricular ejection fraction: comparison with radionuclide angiography [abstract]. Circulation 1991;84(suppl):II- 585.