JACC Vol . 22, No . 3September 1993 :839-47

Objectives. The purpose of this shady was to evaluate the accuracyof the recently developed echocardiographic on-line endocardialborder detection system using ulararast computed tomography, anindependent and proved tomographic imaging modality .

Background. The automated system for on-line endocardialborder detection ldentifieS the blood-tissue interface by acousticquantification of the

backscaller signal.Methods. Eighteen % lijects were screened by conventional

echocardiography and acoustic quantification . Ten of these, withhigh qualky echocardiographic images, were also examined byultrafast computed tomography . Comparable image planes at themidpapillary level were analyzed . Measurements of left ventricu-lar cavity area were compared at end-diastole and end-systole andtime course analyses of cavity area during the cardiac cycle wereperformed.

Results. There was good correlation between values for leftventricular end-diastolic area (r = 0 .99), end-systolic area (r =0.93) and fractional area change (r = 0 .91) using the two methods.The on-line backscatter system underestimated end-diastolicarea (p < 0.001), but the negative bias was small (®1 .6 cam) andthe 95% confidence intervals were narrow (-3.6 cam to+0.4 cam) . In contrast, the bac',scaller system overestimated

Perez et al . (1) recently described a system for on-linequantification of tissue acoustic properties that enables contin-uous measurements of ventricular function in real time . Theechocardiography-based system, which uses ultrasonic back-scatter technology to characterize tissue properties, incorpo-rates a border detection algorithm for delineating the endo-cardial-blood interface . This enables automatic tracking ofthe endocardial boundary in real time . A variety of prelimi-nary reports have endorsed the new technology (2-4) .

To date, on-line measurements of left ventricular cavityarea by acoustic quantification have been compared with

From the Noninvasive Cardiac Imaging Laboratory, Section of Cardiol-ogy, Departments of Medicine and Radiology, University of Chicago, Chi-cago, Illinois.

Manuscript received October 21, 1992 ; revised manuscript receivedFebruary 22, 1993, accepted March 2, 1993 .

Address for cormsnondeng : Richard H. Marcus, MD, University ofChicago Medical Center, Noninvasive Cardiac Imaging Laboratory, Sectionof Cardiology, Department of Medicine, 5841 South Maryland A' -nm,MC5084, Chicago, Illinois 60637 .

(91993 by the American College of Cardiology

NEW METHODS

RICHARD H . MARCUS, MD, JAMES IBEDNARZ, BS, RICHARD COULD N, 1 D,

SANJEEV SHROFF, PHD, MARTIN LIPTON, MD, FACC, ROBERTO M . LANG, MD, FACE,Chicago, Illinois

end-systolic area (p < 0 .02) ; the positive bias for this variable wasalso small (+2A can) but the confidence intervals were relativelywide (+7 .9 to -211 cam). The negative bias of bulaseatnu valuesFor cavity area was fairly constant during diastole and early systole(range -5% to -10%), but during the second half of systole,backscatter values increased progressively relative to computedtoniographic values. Real time values for fractional area changemeasured by the backscatter system were 13% smaller than thosedetermined by ullrafast computed tomography (p < 0 . 1), withwide confidence intervals (+3% to -30%) . Absolute peak rates ofarea change during systole and diastole were lower by 39% (p <0.001) and 41% (p < 0.01), respectively, using the cu-lineultrasonic backscatter system . Time course analyses revealed theerrors to be consistent with cardiac cycle-dependent alterations ingain sensitivity of the ultrasonic backscatter system .

Conclusions . The ultrasonic backscatter system is associatedwith cyclic cavity area measurement errors that need to beaddressed if its early promise for on-line assessment of ventricularfunction is to be fulfilled . Incorporation of an electe, ocardiograph-ically triggered time-varying gain control may improve accuracyfor on-line analysis of ventricular performance .

(J Am Coil Cardiol 1993 ;22.839-47)

off-line echocardiographic measurements of identical or se-quential two-dimensional echocardiographic images . Real-time cross-sectional area values have also been correlatedwith Doppler echocardiographic, angiographic and inva-sively acquired hemodynamic data (1,2-4) ; however, nostudies have attempted to validate the backscatter systemusing an independent tornogiaphic method . Ultrafast com-puted tomography is an independent, accurate, high res-olution imaging modality that can generate a series oftomographic slices analogous to those obtained using thebackscatter system (5-7) . In this study we assessed theaccuracy of the automated system for endocardial honk,detection using ultrasonic backscatter by comparing thereal-time echocardiographic data with measurements ob-tained using ultrafast computed tomography .

MethodsSubjects. Eighteen normal volunteers were screened by

an experienced echocardiographer using conventional two-

0735-1097/93/56,00

840 MARCUS ET AL .AUTOMATED REAL .T(MP ENDOCARDIAL BORDER DETECTION

dimensional echocardiography. After optimizing the qualityof short-axis images of the left ventricle, the backscatter-based endocardial boundary detection system was activatedand the gain controls were adjusted to optimize tracking ofthe endocardial-blood interface . Subjects with high qualitytwo-dimensional short-axis images in whom the ultrasonicbackscatter system appeared to detect and accurately trackthe endocardial-blood interface were selected to undergo thefull study protocol . The study protocol was approved by theHuman Ethics Committee of the University of Chicago .

Backscatter Imaging system (Fig . 1) . The echocardio-graphic system used An quantative integrated backscatterimaging (Hewlett-Packard Sonos 1500) has been describedpreviously, and its application for myocardial tissue charac-terization has been validated QWY Briefly, backscatterdata from each ultrasound A line are integrated over a periodof 3.2 ps and arc relayed to a scan convertor for on-line,real-time image reconstruction (11,12) . The resulting two-dimensional backscatter image is smoothed and averLged toreduce speckling (13) . Time gain and lateral gain compensa-tion settings are optimized to improve visualization of endo-cardial borders (i4). The operator traces a region of interestaround the left ventricular cavity . A calculation and graphicssoftware package then computes and instantaneously dis-plays a trace of left ventricular cavity area versus time forthe region of interest (Fig . 1) .

UlWastcompuledlomography . An Imatron C-100 scan-ner was used to acquire analogous computed tomographicimages of the beating heart. The system uses an electronbeam that is swept across a series of tungsten target ringsthat surround the subject . Two tomographic sections, each8 mm thick, are obtained from each target ring . Scan time foreach image is 50 ms, with up to 80 scans from contiguoustargets being obtained during a single acquisition sequence .

tACC Vol . 22, No. 3September 1199k339-47

Figure 1. Representative ultrasonic backscatterdata. Upper panel, Short-axis echoc . ,rdiographicimage of the left ventricle after activation of thebackscatter system in a representative subject.The area of interest has been defined (the green-white line through the left ventricular myocardiumaround the left ventricular cavity), The automati--ally detected endocardial border is shown in red.Lower panel, Atialog trace of computed left ven-tricular cavity area versus time with accompany-ing electrocardiogram.

Scanning is initiated by the upstroke of the R wave of theelectrocardiogram (ECG), and images arc acquired seriallyat a rate of 17 .4 framesis (i .e., 58 ms/frame) . Subjects wereplaced supine on the scan table, which was then tilted 17° to18° to the horizontal ffeci down) and sleeved 17° to thesubject's right . This enabled imaging in the short-axis plane(I 5) . The desired scanning level was localized and thecirculation time established .

Flow mode scanning protocol. Circulation time was de-termined from a time-density analysis of left ventricularcavity opacification after injection of 25 ml of contrastmedium (Omnipaque 350) through an antecubital vein . Afterinjection of contrast medium at 10 mils, a 25-ml salinesolution -chaser" was infused at the same rate . Time-density data were acquired by repeatedly scanning the samelevel on alternate R waves for 25 s after the start of theinjection. The time to peak left ventricular opacification wasused as a measure of circulation time for subsequent cinemode examinations .

Cine mode scanning protocol (Fig . 2 and 3) . Sixty milli-liters of contrast medium was infused through an aittecubitalvein using a powered injector (Medrad Mk 5) at a rate of 2 to3 ml/s. Scanning was timed to coincide with peak leftventricular opacification and was performed in held midin-spiration. Four tomographic levels (spanning 4 cm of themid-left ventricular cavity) were scanned (Fig . 2) . To ensuredata acquisition throughout the cardiac cycle, 20 images(frames) were obtained at each tomographic level (Fig . 3, leftpanel), permitting acquisition of complete data sets forcardiac cycle lengths up to 1,352 ms (i .e., heart ratesX52 beats/min) (Fig. 3, right panel) .

Procedure . Ultrafast computed tomographic data wereacquired as described in the preceding section in 10 eligiblesubjects. Cine computed tomographic images were immedi-

JACC Vol . 22, No . 3September 1993 : 8 3 9-47

Figure 2. Schematic illustration of the plane and levels throughwhich the left ventricle was imaged . Cavity area was determined forthe tomographic section that transected the midportions of thepapillary muscles. Ao = aorta ; LA = left atrium; LV = leftventricle ; Ph4 zo papillary muscle .

ately reconstructed and visualized by three experiencedobservers . Particular attention was paid to the imaging planeand levels of the tomographic sections, using the papillarymuscles as anatomic landmarks. With the subject still on thescan table, and within 5 min of completing the computedtomographic study, the echocardiographic data were ac-quired . The echocardiographic image plane and level werematched to those of the corresponding midpapillary com-puted tomographic image to eliminate discrepancies thatmight result from regional variations in left ventricularfunction (16,17) . Once the desired two-dimensional imagehad been optimized, the backscatter system was activated .Gain settings were adjusted and a region of interest wasselected using a trackball to incorporate all portions of the

MARCUS ET A L,

84,AUTOMATED REAL-TIME ENDOCARDIAL BORDER DETECTION

left ventricular cavity throughout the cardiac cycle. Thegraphics software was then initiated to produce a real-timeanalog signal of left ventricular cavity area versus time . Thecardiac cycles used to acquire computed tomographic datawere different from those used for real time measurements,but cycles were matched for RR intervals .

Data analysis (Fig . 4) . Single-frame cine computed tomo-graphic images were calibrated off-line according to framedimensions and pixel size. Endocardial boundaries of cinecomputed tomographic image slices, acquired at 58-ms in-tervals, were manually traced off-line using a mouse andcommercially available digitization software (Freeland Sys-tems) for calculation of left ventricular cavity area . Individ-ual computed tomographic area data points acquired at themidpapillary level were conjoined to produce a plot of leftventricular area versus time, similar to that generated in realtime by the backscatter imaging system (Fig . 4). The area-time plots for the two methods were digitized at 200 liz usingcustom-written software and a personal computer (EpsonEquity

Comparison of ultrasonic backscatter and ultraffist com-puted tomographic data . The following variables, deter-mined by the two imaging techniques, were compared :

Figure 3. Representative ultrafast computed tomographic data . Leftpanel, Electrocardiographically triggered time points at which im-ages were acquired. Twenty image slices were obtained at eachtomographic level. Right panel, Complete data set of computedtomographic images for a single level at 58-rns intervals throughoutthe cardiac cycle. Because the RR. interval in this instance was870 1 , the first 16 images from the R wave (arrows) were analyzed .

842 MARCUS ET ALAUTOMATED REAL-TIME ENDOCARDIAL BORDER DETECTION

September 1993 : 83 9-47

0

0

a

0

0 0 0 6 a aa 0

Figure 4. Computed tomographic image analysis . Upper panels,End-diastolic (left) and end-systolic (right) computed tomographicimage frames with superimposed off-line tracing of endocardialborder . Lower panel, Plot of left ventricular short-axis cavity areaversus time at the midpapillary level, constructed off-line by con-joining ultrafast computed tomographic data points in a representa-tive subject . Arrows indicate end-diastolic and end-systolic areavalues .

1) L10 ventricular end-diastolic cavity area (EDA). definedas the cavity area coinciding with the peak of the R wave onthe ECG; 2) Left ventricular end-systolic cavity area (ESA),defined as the smallest cavity area determined from thearea-time tam; 3) Left ventricular ,fractional area change(PAC), calculated as

i'AC (%) = 100 x [EDA - ESAYEDA ;

4) Peak rate of area change during systole [dAldtsl anddiastole (dAidid, determined by sifferentiation of the digi-tized area-time signals ; and 5) Area-time profiles during thecardiac cycle . For each analyzed beat, both systole anddiastole were divided into deciles to facilitate analysis ofgrouped data from individual subjects with different heartrules. Lea ventricular cavity areas for each of the 10equispaced corresponding time points during systolic anddiastolic phases of the cardiac cycle were determined by thetwo methods (i.e., computed tomography and echocardiog-raphy). The difference between backscatter and computedtomographic values for analogous beats was calculated ateach time point.

JACC Vol 22, No . 3

Table 1 . Cardiac Variables Measured by Acoustic Quantificationand Ultrafast Computed Tomography

Values are presented as mean value ± SD . AQ = acoustic quantification ;[dA/dt s] = rate of change of area during systole ; [dA/dt t)] = rate of change ofarea during diastole ; = end -diaslolic left ventricular cavity area ; ESA -.

end-systolic left v^ ,ltricular cavity area. FAC = fractional left ventricularcavity area change ; UFCT = ultrafast con ited tomography .

Statistical analysis. Three different analyses were per-formed to evaluate the validity of real-time ari'a data ,I) Linear regression, to identify a predictable relationbetween backscatter and computed tomographic data ;2) Bland-Altinan analysis, to quantitate bias between vari-able values obtained with the two techniques and to deter-mine the predictive value (population error) of backscattermeasurements in individual subjects (18) ; 3) Student t testfor matched pairs, to ascertain the statistical significance ofbias between values determined by the two methods .

ResultsPatient selection . Eighteen volunteer subjects were pro-

spectively screened by conventional two-dimensional echo-cardiography and backscatter imaging with automatedendocardial border detection . Ten subjects with high qualityimages were selected for admission to the computed tomo-graphic portion of the study . The remaining eight subjectswere excluded either because conventional two-dimensionalimages were suboptimal (n = 4) or because the backscatterautomated boundary detection algorithm appeared not totrack the endocardial-blood interface accurately throughoutthe cardiac cycle .

Left ventricular end-diastolic and end-systolic cavity areas(Table 1, Fig. 5, A and B). Values for end-diastolic andend-systolic areas obtained in real time using the backscatterimaging system with automated endocardial border detec-tion correlated well with values determined off-line fromultrafast computed tomographic images (r = 0 .99 for end-diastolic area and r = 0 .93 for end-systolic area) ; however,echocardiographically derived on-line end-diastolic area val-ues were smaller than the corresponding computed tomo-graphic values in all subjects . The difference was significant(p < 0 .00 1), but the negative bias was small (-1 .6 cm), andthe 95% confidence intervals for the difference were narrow(-3.6 to +0.4 cm). Conversely, real-time end-systolic areavalues were larger than corresponding computed tomo-graphic data in nine subjects, resulting in a significantdifference between the two groups (p < 0 .02). Although the

AQ UFCT p Value

EDA (cm2 ) 21 .9 t 113 23 .5 ± 7.7 < owlESA (cm'- ) 12 .2 ± 7.2 9 .6 ± 6.7 < 102FAC (cm2 ) 50 ± 19 63 ± 20 < 0 .001Peak [dA/dt s] (cm 2/s) -57 ± 20 -93 t 34 < 0.001Peak [dA/dt D ] (cm 2ls) 62 ± 29 106 ± 46 < 0.01Heart rate (beats/min) 65 ± 13 68 ± 12 NSDuration of systole (ms) 436 ± 53 372 ± 39 < 101

JACC Vol . 22, No . 3seMember a 3 :539 -47

W

Zo

at

Q

U

e9

8U

CF MA (01fi2)

positi-e bias was small (+2 .6 cm2 ), the confidence intervalsfor this difference were relatively large (+7 .9 to -2.8 CM2) ,

indicating a low predictive value for real-time end-systolicarea data .

Fractional area change (Table 1, Fig. 5Q . Values forfractional area change obtained by the two imaging modali-ties also correlated well (r in 0.91). Because the echocardio-graphic technique underestimated end-diastolic area andoverestimated end-systolic area in comparison with com-puted tomographic data, fractional area change values deter-mined in real time were smaller than those obtained by cinecomputed tomography (p < 0 .001, negative bias -13%) ; 95%confidence intervals for the difference were wide (+3% to-30%) .

Peak instantaneous rates of area change (Table 1). Al-though paired data were acquired sequentially, heart ratesduring acquisition of backscatter and computed tomographicdata were not significantly different . Because the total areachange was smaller for backscatter measurements, the meanrate of area change (average slope of the area-time plot) waslower than for computed tomographic measurements . Theabsolute values for both peak [dA/dtsl and peak tdA/dtDJdetermined by the on-line ultrasonic backscatter systemwere also lower than equivalent computed tomographicdata (-39%, p < 0 .001 for tdA/dtv,] ; -41%, p < 0 .01 for[dA/dtD]) . These differences raise the possibility that thefrequency response of the backscatter system is insufficientto track the most rapidly changing endocardial-blood inter-faces .

Asynchnnty between electrocardiogram and barkscatterarea analog signal (Table 1, Fig . 6). Although heart ratesduring acquisition, of computed tomographic and backscatlerimages were not significantly different, the duration of sys-tole, measured from the peak of the R wave of the ECG toend-systole (i .e ., minimal cavity area), was longer for the

&MCUS ET AL .AUTOMATED REALAIME ENDOCM(DLAL

_TFECTION

15 25

Time t660

00

84

111b.

CT PAC; (%)

843

-2: ff,

Figure 5. Comparisons of real time echocarftvraphic and off-linecomputed tomographic measurements of end-diastolic area (EDA)(A), end-systolic area (ESA) (B) and fractional area change (FAQ(C). Upper panels, Values for the two methods are plotted againsteach other. The line of identity has been superimposed . Thecorrelation coefficient is given for each relation. Lower panels, Thedifference between echocardiographic and computed tomographic(CT) values (vertical axis) is plotted against the computed qomo-graphic value (horizontal axis) . The central dashed line identifies themean difference between methods (i .e., the bias) . The limits ofagreement between methods are represented by the 95% confidence :intervals (upper and lower dashed lines) that identify the potentialerror of individual measurements, AQ = acoustic quantification .

real time data (436 ± 53 ms) than for ultrafast computedtomographic measurements (372 ± 39 ms, p < 0 .01). (hereis a delay between the ECG and the ultrasonic backscatter-based analog area waveform (manufacturer's specifications) .This delay, due to the asynchronous processing of FCC andultrasonic backscatter data, is variable, ranging from zero to[Ilacoustic frame rate] . The discrepancy of the systolicduration between the two methods could be attributable tothis time delay in the output of the ultrasonic data. Becausethe exact time delay in ultrasonic area signal for each subject

Figure 6. Area-time plots obtained using uhrafast computed lomog-raphy Washed curve) and acoustic quantification (solid curvet in arepresentative subject .

33 -~

224

1110 0.5

te

844

MARCUS ET AL.AUTOMATED REAL-TIME ENDOCARDIAL BORDER DETECTION

D

a.

d

a

s'

~bl1 .:I'i' ""' i

9 A

~~pNkMab

SYSTOLE

was net known, we followed a model-based procedure for itsestimation . Acoustically quantified ultrasonic area [A AQ(t))was related to computed tomographic area [Ac-r(t)) as fol-lows :

[Acr(t)] = S x [AAQ(I - r)] + I,

where 5 and I are the slope and y intercept, respectively, andr is the time delay between acoustically quantified andcomputed tomographic data and was estimated from exper-imental data (i .e ., [A cr(t)] and [A AQ(t)]) by using the previ-ous linear equation and the least squares estimation . Thisprocess can be alternatively stated as an attempt to find thevalue of r such that the highest possible linear correlationbetween [A c,(t)] and [AAQ(t-r)] is obtained. The estimatedvalues for r appeared to be appropriate in that they variedbetween zero and [Ilacoustic frame rate] . After time shifting[AAQ(t)) by the estimated value for r, the discrepancy in thesystolic duration between the two methods was abolished :372 ± 39 ass for computed tomographic data versus 41059 ms for time-shifted acoustically quantified backscatterdata .

T[me-course analysis of on-line area measurements (Fig. 7).The' ime course patterns of echocardiogaphic and ultrafastcomputed tomographic left ventricular cavity area measure-ments were compared by computing the differences betweenvalues for the two techniques at 10% time intervals through-out systole and diastole . During diastole and early systole, afairly constant negative bias of -5% to -1t% was found foron-line data relative to corresponding computed tome.graphic values. At midsystole, a sudden reversal of thisrelation occurred and backscatter values became progres-sively higher relative to computed tomographic data . Withthe onset of ventricular relaxation, this trend was abruptlyterminated, and the negative bias for echocardiographicvalues was restored .

Maximal on-fine area measurement difference was notcoincident with peak rates of area change (Fig . 7). Peak[dA/dt5] consistently occurred t-1 the first half of systole,before the maximal systolic backscatter measurement error

DIASTOLE

±ACC Vol . 22. No . 3September 1993 :539-d7

Figure 7. Average error of left ventricu-lar cavity area estimations by acousticquantification (AQI at incremental timedecdes of systole mud diastole before andafter time correction . The relatively con-stant negative bias of backscatter valuesduring diastole and early systole is re-placed by an exponentially increasingpositive bias during late systole . Timecorrection did not significantly influencethe magnitude or time course of the areameasurement error. The maximal errorwas not temporally related to peak de-celeration or acceleration of cavity area([dA/dt sI and [dA/dt„], respectively) .CT = ultrafast computed tornogaphy .

was observed . Peak tdA/dt ,] did coincide with a portion ofearly diastole when the backscatter overestimation of cavityarea was maximal, but the diastolic measurement error thenremained fairly constant despite the subsequent decrease in[dA/dtD] .

To examine whether the phasic pattern of error wasrelated to the time delay in acoustically quantified backscat-ter data, the errors were recalculated after time shifting thebackscatter area signal (described earlier). As illusuated inFigure 7, this time-delay correction did not impact signifi-cantly on the pattern or magnitude of the error during systoleor diastole .

Discussion

Ration* for current study. The sonographic systemrecently described by Perez et al . (1) uses backscattertechnology to characterize tissues according to the intensityof their acoustic reflv :tions . A computer-assisted algorithmis employed _o analyze unprocessed acoustic signals fromblood-tissue interfaces in real time . Perez et al . (1) comparedautomated cchocardiographic measurements of left ventric-ular cross-sectional area with off lire manual analyses ofidentical or sequentially acquired echocardiographic images .In an accompanying editorial, Martin (19) pointed out thatindependent tomographic methods are needed to assess theaccuracy of the backscatter system more meaningfully . Inthis study, ultrafast compu •_d tomography, a proved team-graphic technique for erdocardial border discrimination,was used to evaluate the accuracy of the real-time automatedechocardiographic system. Our results support the conclu-sions of Perez et al. that the backscatter system for auto-mated endocardial border detection has great promise butalso identify certain limitations in its current stage of devel-opment.

Accuracy of acoustic quantification for evaluation of overallsystolic performance. To date, protagonists of the ultrasonicbackscatter system with automated endocardial border de-tection have focused on the close correlations between

-ACC Vol. 22, Nn. 3

MARCUS Er AL.September 1993:839-47

AUTOMATED REAL-TIME ENDOCARDIAL SORDER DEtECTION

on-fine cchocarliographic data and measurements madeoff-line by manual digitization of corresponding two-dimensional echocardiographic (tomographic) or cineven-triculographic (nontomographic) images (1,20) . In this study,too, a high degree of correlation was found between echocar-diographic and computed tomographic measurements ofend-diastolic area, end-systolic area and fractional areachange . However, close correlation between backscatterdata and those obtained using proved measurement tech-niques does not necessarily imply that th° on-line echocar-diographic system is accurate . The correlation coefficient isnot an indicator of agreement between two methods ofmeasurement; it is an index of association (18) . To computethe "error" of one method of measurement relative toanother, it is important also to determine the relative biasbetween values obtained by the two methods and to assessthe variance of the data that accounts for the 95% confidenceintervals for measured values in individual subjects . Thestatistical techniques described by Bland and Altman (18)were used to address these issues .

The real-time ultrasonic backscatter system consistentlyunderestimated end-diastolic area and overestimated end-systolic area . The compounding effects of these errors onfractional area change values resulted in a significant under-estimation of fractional area change (p < 0 .001). The Bland-Altman analysis identified a relatively large bias and wide95% confidence intervals for individual fractional areachange values.

Accuracy of aearstie quantification far instantaneous areatreasurement . A time-area error analysis revealed a charac-teristic bimodal temporal pattern of error : I) a small constantunderestimation of cavity area by the backscatter systemfrom early diastole through early systole, resulting in under-estimation of end-diastolic area ; and 2) a sudden and pro-gressive increase in backscatter values relative to computedtomographic data during lace systole, resulting in a time-varying overestimation of end-systolic cavity area by acous-tic quantification (Fig. 7).

Underestimation of the end-diastolic area by the back-scatter system has been reported previously and has beenattributed to exclusion of the area occupied by the papillarymuscles from the acoustically quantified cavity area (1) .However, in this study the eudocardial boundary tracing forultrafast computed tomographic images also excluded thespace filled by the papillary muscles. A more likely expla-nation for this finding is that the image planes for computedtomographic sections were slightly oblique, resulting inmarginal overestimation of end-diastolic area by this tech-nique when compared with the more orthogonal echocardio-gaphic sections .

Overestimation of the end-systolic area by the backscat-ter system is more difficult to explain . Three possible mech-anisms were considered that could account for the phasicnature of the late systolic measurement error : 1) a low passfilter effect; 2) the time delay between the ECG and the

845

on-line area analog signal; and 3) time-varying gain sensitiv-ity of the backscatter system,

Law pass filter effect. Two observations could not bereconciled with low pass filter with fixed direct current (DC)gain . First, the ultrasonic system underestimated end-diastolic area but overestimated end-systolic area . Second,the ultrasonic area-time carves were in quasi-steady state(i.e., rate of area change close to z") at bath end-diastoleand end-systole (Fig. 6).

Time delay. Time correction of the on-line ecbocardio-graphic data to compensate for the delay between the ECGand the on-line analog area signal of the backscatter systemdid not have a significant impact on the cyclic disparitybetween echocardiographic and computed tomographic leftventricular cavity area measurements .

Gain sensitivity of acoustic quantification . Identificationof endocardial borders by the backscatter technology isdependent on the settings for transmitted power and time-gain compensation. The gain control settings are empiricallyselected by the operator such that the automatically detectedblood-endocardial interface is superimposed on the visual-ized endocardial border of the two-dimensional ecbocardio-graphic image. If tissue contrast is insufficient or the signal tonoise ratio is too low, the system will not detect the interfaceuntil the gain settings are increased . As gain settings areincreased, the recorded cavity area decreases . Accordingly,underestimation of end-diastolic area values by acousticquantification in this study may be attributable to an exces-sively high setting of the gain controls . Conversely, over,estimation of end-systolic area values by acoustic qualttif-cation reflects inappropriately low gain levels, resulting in adecreased signal to noise ratio . In other words, an optimalgain setting for end-diastole is inappropriate Mar end-systole,and vice versa . The phasic late systolic measurement errorcould be entirely accounted for by time- or magnitude-dependent direct current gain changes during the cardiaccycle. Magnitude-dependent alterations in gain sensitivityseem improbable because the temporal pattern of areadifference between the two methods was similar for largeand small ventricles (i .e., among different patients). It ismore likely that the periodic error is due to phasic variationsin myocardial acoustic properties that occur during anisotro-pie ventricular fiber shortening (21) . Such cyclic variations inacoustic intensity could account for the pattern of discor-dance bet .veen ecbocardiogaphic and computed tomo-graphic area values recorded in this study . Although discon-certing, these measurement errors do not constitute aninsurmountable problem. Because the gain controls areadjustable, it may be possible to incorporate a time-varyinggain adjustment, triggered by the ECG that would automat-ically adjust gain settings in a cyclic manner during eachcardiac cycle .

Accessibility of data. The time cost of off-line analysispackages that require manual tracing of still-frame imagesconstitutes the major limitation to quantitative analysis formost imaging modalities. The ultrasonic backscatter system

846

MARCUS ET AL .AUTOMATED REAL-TIME ENDOCARDIAL BORDER DETECTION

with automated endocardial border detection offers a poten-tial solution to this problem by providing on-line quantitativeinformation on left ventricular function that until now hasbeen inaccessible to the clinician ; however, technical con-straints may limit its application. Perez et al . (i) were able toacquire satisfactory data in only 72% of their patients usinglong-axis imaging planes . Short-axis data were even lessaccessible owing to echo dropout in the regions of the lateraland septal left ventricular walls (i .e., tangential to thesonographic signal) . To address this issue a lateral gaincompensation facility was incorporated that enhances back-scatter echoes in preselected radial segments of the field ofview, In our study all echocardiographic images were ac-quired in short-axis planes . Four (22%) of 18 subjects wereexcluded because of poor tracking of endocardial borders byacoustic quantification, even after deploying lateral gaincompensation .

Limitations of the study . Matching of the image planes forultrafast computed tomographic and echocardiogtaphic dataacquisition, although performed by three experienced ob-servers, was based entirely on visual landmarks . In addition,Rumberger (22) has identified potential sources of error inborder detection using ultrafast computed tomography, re-lating both to partial volume effects for 8-mm-thick sectionsand to the need for identification of the "full-width halfmaximum density value" of each border . Either or boththese factors might have accounted for the small baselinedifferences in end-diastolic area . However, because end-systolic area was consistently overestimated by acousticquantification despite consistently lower values for end •diastolic area, the underestimation of left ventricular systolicperformance (fractional area change) cannot be attributedeither to image plane heterogeneity or to problems withborder detection of computed tomogiaphic images .

Echocardiographic and ultrafast computed tom graphicdata were not acquired simultaneously; the backscatterimages were obtained within 5 min of the computed tomo-

hic study and were matched for RR intervals to mini-mize potential differences in cycle duration that might altercavity measurements .

The constraints placed on plane selection by the ultrafastcomputed tomo hic system dictated that the short-axisplane be used for data comparisons . Perez et al. (1) encoun-tered difficulties in detecting and tracking erdocardialboundaries in the sternal short-axis view, particularly inthe anterolateral and posteroseptal regions that are tangen-tial to the signal source . The advent of lateral gain compen-sation has enhanced short-axis image discrimination to theextent that this problem has been effectively resolved, asshown by this study,

Cz F ,>m,._i r ,, The acoustic quantification system for au-tomated real-time endocardial border detection is an excitingnew technology that could impact meaningfully on cardio-vascular diagnosis and management by providing on-lineserial analyses of left ventricular systolic and diastolic func-tion. Incorporation of an ECG-triggered time-varying gain

JACC Vol. 22, No . 3September 1993 :839-47

control may increase the accuracy and reproducibility ofautomated backscatter measurements and expedite the ap-plication of this technology in the clinical arena .

We thank Donna Barrett and Lynn Weinert for ccatributions to the prepara-tion of tois article .

ReferencesI . Perez JE, Waggoner AD, Barzilai B, Melton HE, Miller JG, Sobel BE .

On-line assessment of ventricular function by automatic boundary detec-tion and ultrasonic backscatter imaging . J Am Coll Cardiol 1992 ;19:313®20,

2 . Foley DA, Seward JB, Tatjik J . Assessment of left ventricular diastolicfunction with a new automated echocardiographic border detectionsystem : comparison with Doppler (abstr), J Am Coll Cardiol 1992 :19:261 A,

3 . Gorcsan J Ill, Deneault L,6, Pinsky MR . Echocardiographic automatedborder detection for on-line determination of left ventricular functionduring acute alterations in preload (abstr) . I Am Coll Cardiol 1992 ;19:262A .

4 . Sapin PM, Xic C, Smith MD . Kwan OL, DeMaria AN . Identification ofnormal and abrsrnral left ventricular diastolic filling using a real timeautomatic bordet detection algorithm (abstr) . J Am Coll Cardiol 1992 ;19 :19A .

5 . Rumberger JA, Weiss RM . Feiring AJ, et al . Patterns of regional diastolicfunction in the normal human left ventricle : an ultrafast CT study . J AmColl Cardiol 1989 ;14:119-26.

6 . Reiter SJ, Rumberger JA, Felting AJ, : tanford W, Marcus ML . Precisionof measurements of right and left ventricular volume by ultrafast cinecomputed tomography . Circulation 1986 ;74 .990-

7 Marcus Mt, . Weis% RM . Evaluation of cardiac structure and function withultrafast computed tomography. In : Marcus ML, Schelbert HR, SkortonDJ, Wolf GI ., eds . Cardiac Imaging . Philadelphia : WB Saunders, 1991 :671-2

8 .. Masuyama T. Nellessen U . Schnittger I, Tye T, Haskell WI ., Pupp RI . .Ultrasonic tissue eharacteri ath~tr with a real time integrated backscatterimaging system in normal and aging human hearts . J Am Coll Cardiol1989J41702- -8

9 Vandenberg IJF Rath L Sh .,upTA, Ke-ber RE, Collins SM, Skorton DJ .Cyclic variatior. of ultrasonn--l bec'.scaiter in normal myocardium isview-dependent: clinical st ii-s with a real-time backscatter imapingsystem . J Am Sue Eehocararogr 1989 :2 :308-14.

10, Vandcnberg BF . Stuhlmuller JE, Rath L, et al . Diagnosis of recentmyocardial infarction with quantitative backscatter imaging : preliminarystudies . J Am See Echocardiogr 1991 ;4:10-8 .

It . Vered Z, Barzilai B, Mohr GA, et al . Quantitative ultrasonic tissuecharacterization with real-time integrated backscatter imaging in normalhuman subjects and patients with dilated cardiomynpaihy . Circulation1 7;76 :1067-73.

12 . Thomas L11II, Barzilai B, Perez JE, Sobel BE, Wickline SA, Miller JG .Quantitative real-time imaging of myocardium based on ultrasonic inte-grated backscatter . IEEE Trans Ultrason Ferroelecti Freq Control 1989 ;36:466-70.

13. Skorton DJ, Miller JG, Wickline SA, Barzilai B, Collins SM, Perez JE .Ultrasonic characterization of cardiovascular tissue . In : Marcus M,Schelbert HR, Skorton DJ, Wolf G, eds . Cardiac Imaging: Principles andPractice . Philadelphia: WB Saunders, 1990:538-56 .

14. Perez JE, Waggoner AD, Barzilai B, et al . Lateral gain compensation : anew method to improve two dimensional echocardiographic imaging andfacilitate on-line edge detection and quantification of ventricular function(abstr). J Am Coll Cardiol 1992;19 :262A.

15. Rees MR . Feiring AJ, Rumberger JA, MacMillan RM . Clark DL. Heartevaluation by cine CT: use of two new oblique views . Radiology 1986 ;159:804-6.

16. Smith MD, Gong-Yuan X, Sapin PM, et al . Heterogeneity of regionalcontractile function in normal subjects as assessed by acoustic quantifi-cation (abstr). J Am Coll Cardiol 1992 ;l9 :19A .

17. Feiring AJ, Rumberger JA, Reiter SJ, et al. Sectional and segmental

JACC Vol . 22, No . 3

M1,41!CUS LT AL .September 1993AW-47

AUTOMATED REAL-TME ENDOCARDIAL RnRDFR DETECTION

variability of left venlricalai function : exr,,-Amenial and clinical studiesusing uitFafmt computed tomography, 3 Am Coli Cardiol 1935 :12 ~t5 '25 .

18. Bland JM, Altman DG. Statistical methods for assessing agreementbetween two methods of clinical mcasurement. Lancet 19861 :307-W .

19 . Martin RP. Real time ultrasound quantification of ventricular function :has the eyeball been replaced or will the subjective become objective'?3 Aran Coll Caidiol 19920321-3 .

20. Sapin PM, Xie 0, Smith MD. Kwan CAL . DeMaria AN . The timing ofchanges in left ventricular area by echocardiography derived from a

847

real-infle automatic border detection algorithm (ahstr) . 3 Am Coll Cardiol1092 ;19 :261A .

2 1 . Madams El, Perez JE, Sobel BE, Miller JO, Anisotropy of idirastorlicbackscatwer of myocardial tissue . 11 . Measurements in vivo . 3 Acoust SucAm 1988:83 :762-9 .

22. Rumberger ]A . Hncipies avid practice of cardiac edge. detection andborder definition using ultrafast computed !omogrdpl .y, In : Stanford WS,Rumberger .1, eds. Ultrafast Computed Tcmography in Cardiac ; Imaging:Principles and Roac ice. New York : Futura, 3992 :25-35 .