automation of gated tomographic left ventricular ejection fraction

Automation of gated tomographic left ventricular ejection fraction

K e n n e t h N i c h o l s , PhD, E. G o r d o n D e P u e y , M D , a n d A l a n Rozansk i , M D

Background. The feasibility of determining left ventricular (LV) ejection fraction (EF) from 99mTc-labeled sestamibi gated tomography (GSPECT) is well established. To improve precision of measurement, rules used by observers in processing tomograms were encoded for auto- mation.

Methods and Results. LV centers were estimated from activity centroids of time-difference images exceeding 50% of maximum counts. End diastole and end systole were defined by time-varying maximum count extremes. Endocardial borders were generated by fitting maxi- mum locations with fifth-order two-dimensional harmonics, searching inward to predetermined thresholds, and reconciling endocardial with valve plane points. Regression analysis of GSPECT EF yielded r = 0.87 versus equilibrium gated blood pool in 75 patients and r = 0.87 versus gated first pass in 65 patients. GSPECT EF interobserver variability was r = 0.92 and intraobserver automatic versus manual linear correlation was r = 0.94. A subgroup of 25 studies was analyzed by six independent observers, for whom EF agreement with the core laboratory ranged from r = 0.93 to r = 0.96. Experienced observers judged it necessary to alter end-diastolic or end-systolic frames in 7% of patients, endocardial borders in 14%, and LV centers in 28%.

Conclusion. Results of automated GSPECT LV EF correlated well with those of manual GSPECT and gated first-pass and equilibrium blood pool values and were highly reproducible. (J Nucl Cardiol 1996;3:475-82.)

Key Words: radionuclide imaging.gated single-photon emission computed tomog- raphy • ejection fraction

The clinical importance of assessing the extent and severity of inducible myocardial hypoperfusion has been well established. As the magnitude of inducible myocar- dial hypoperfusion increases, the likelihood of a patient having cardiac events increases exponentially. 1 Other data indicate that resting left ventricular (LV) ejection fraction (EF) is also a powerful predictor of outcome in patients with coronary artery disease) Until recently, these two forms of information could be acquired only on different occasions, by performing two separate tests, such as thallium stress/rest myocardial perfusion single- photon emission computed tomography 3 and resting

From the Department of Radiology and the Division of Cardiology, St. Luke's-Roosevelt Hospital, and Columbia University College of Physicians and Surgeons, New York, N.Y.

Supported in part by a research grant from General Electric Medical Systems, Milwaukee, Wis.

Received for publication April 10, 1996; revision accepted June 28, 1996.

Reprint requests: Kenneth Nichols, PhD, Division of Cardiology, St. Luke's-Roosevelt Hospital, Amsterdam Ave. at ll4th St., New York, NY 10025.

Copyright © 1996 by American Society of Nuclear Cardiology. 1071-3581/96/$5.00+0 43/1/76176

radionuclide ventriculography. 4 With the advent of 99mTc-labeled sestamibi (MIBI) it has become possible to obtain high-count R wave-gated tomographic (GSPECT) images, s so it is now feasible to obtain both perfusion and functional information from the same set of images.

In 1993 we proposed a method to compute LV volumes and EF with a modified Simpson formula, using paired gated vertical long-axis (VLA) and hori- zontal long-axis (HLA) myocardial perfusion MIBI to- mograms. 6 Although accurate, this method used soft- ware requiring observers to make three kinds of choices: (1) estimation of the LV center, (2) determination of the end-diastolic (ED) and end-systolic (ES) frames, and (3) manual drawing of endocardial borders. The last task was the most demanding and required drawing edges on the midventricular VLA and HLA images separately for ED and ES frames. Whereas average GSPECT EF computed by these manual techniques agreed well by linear regression with equilibrium gated blood pool (GBP) EF, a range of correlation coefficients from r = 0.79 to r = 0.88 was found for manual computations of different observers. 6 To aid new observers in the application of these methods, we decided to automate

475

476 Nichols et al. JOURNAL OF NUCLEAR CARDIOLOGY Gated tomographic left ventricular ejection fraction November/December 1996, Part 1

Figure 1. Approximate ED frame (A) is subtracted from approximate ES frame (B) to form time-difference image (C). Starting from center of difference image, iterative searches are performed until convergence is achieved at center of ES frame, which is shown to observer (D).

the various steps required for calculations. This investi-

gation evaluates the validity of these steps for automa-

tion of EF and volume measurements versus those

obtained by other widely accepted techniques. Second, interobserver reproducibility is also evaluated. Third, we

assess how often a given set of automated algorithms

fails, under what circumstances, and to what extent operator intervention affects measurements.

METHODS

Data Acquisition and Preprocessing. Patients under- going a same-day rest/stress protocol were routinely injected at peak stress with 1.11 GBq (30 mCi) MIBI. Thirty minutes thereafter they were imaged supine for 64 projections for 20 seconds per projection with a high-resolution collimator and eight frames per R-R interval. All images were acquired as 64 x 64 matrixes, with a standardized pixel size of 6.4 ram. Alternatively, some patients underwent a 2-day protocol in- volving MIBI injections at both rest and stress of 814 MBq (22 mCi). 7 Data preprocessing consisted of an observer selecting LV anterior, inferior, septal, and lateral limits and identifying approximate LV symmetry axes. Preprocessing was performed with commercially available software, 8 which subsequently generated the dynamic R wave-gated midventricular VLA and HLA images.

Manual Computations. The actual "learning curve" for manual EF GSPECT processing was constructed to ascertain the number of patient studies required for new observers to reach a plateau of agreement with experienced users. LV volumes and EFs of an experienced observer and two new observers who used manual techniques described previously 6

were recorded for 75 sequential clinical studies. All observers' computations were performed independently.

Center Location Automation. The first stage of auto- mation was LV centering. Frames 1 and 4 were identified as "approximate ED" and "approximate ES" based on our previous experience with the vast l~ajority of clinical studies acquired with our equipment. A time-difference image was formed by subtracting the approximate ED frame (Figure 1, A) from the approximate ES frame (Figure 1, B) and computing the centroid for time-difference counts exceeding 50% of maximum myocardial counts (Figure 1, C), the rationale being that only myocardial counts should vary synchronously with R waves because of partial volume effects. 9'1° This was fol- lowed by two sequential searches of 10 x 10 neighborhoods of ES pixels, each time repositioning the center of the searching region to the local minimum count position. This estimated LV center was shown to the observer, who was free to modify it (Figure 1, D). This operation was performed separately for VLA and HLA midventricular projections.

Phase Identification Automation. VLA frames were rotated to be horizontal (Figure 2, A), and all HLA and VLA images were magnified to twice the original size, for a new pixel dimension of 3.2 mm. The next step was refinement of centered ED and ES frames, accomplished by tabulating maximum counts for the entire frame of each of the eight R wave-gated frames separately for VLA and HLA magnified cines. The greatest contrast among VLA and HLA maximum counts was then used to define ED and ES frames. The ED and ES VLA and HLA frames so identified were then shown to the observer, who compared the suggested images with frames before and after the chosen frames and altered these if necessary.

Endocardial Border Automation. The last processing

JOURNAL OF NUCLEAR CARDIOLOGY Nichols et al. 477 Volume 3, Number 6;475-482 Gated tomographic left ventficular ejection fraction

Figure 2. A, VLA images are normalized to uniform brightness and rotated to be horizontal for input to edge-detection algorithms. B, Initially identified radial maxima are depicted as bright isolated pixels within myocardium. C, Fifth-order Fourier fit connects these points and provides smooth, continuous region within which to begin searching for endocardial border. D, After inward threshold searching and valve plane estimation, final border is then displayed.

step was generation of endocardial borders. Count searching threshold values were defined as:

Percent ED threshold = c + 0.35 x (D - c) x ~/(] - c/D) (equation 1)

Percent ES threshold = c + 0.35 x (S - c) x (1 - c/S) 2 (equation 2)

where c = minimum LV cavitary counts, D = ED maximum myocardial counts, and S = ES maximum myocardial counts. These formulas were derived empirically to reproduce the aver- age of three cardiologists' manual endocardial drawings. A ra- dial search was performed, initially on the ED VLA image (Figure 2, A), and maximum counts and their radial locations were tabulated every 2 degrees. If maximum counts fell below the count thresholds of either equation, they were not included for initial estimation of midmyocardial radii (Figure 2, B). All candidate points were tested for consistency and discarded if radial changes exceeded 4 pixels. Radii were interpolated be- tween angles for which no radii were successfully identified. For this collection of radial points {r(0)} at 180 angles, the sets of Fourier coefficients {A} and {B} were computed as follows:

2re

A n = ~ cos(n x 0) x r(0);B n = 9 = I

2~t

sin(n x 0) x r(0) (equation 3) 0 = 1

The first five elements of each of the series of coefficients {A}

and {B } were then used to generate the fitted curve f(0) such that:

5

f(0) = ~ A n x cos(n x 0) + B n x sin(n x 0) (equation 4) n = l

This constitutes a two-dimensional fifth-order Fourier fit to the midmyocardial points (Figure 2, C).

Because the fitted midmyocardial curve is constructed from a low-order Fourier series, it is smooth and continuous and extends beyond the valve plane and into the outflow tract. "Average basal counts" were tabulated at all points along this curve for the back half of the heart. Then valve plane end points were estimated by finding the locations at which the fitted midmyocardial curve exited the heart base at 50% of the average basal counts. A search was initiated starting from the fitted curve toward the LV center for endocardial points at the count thresholds of equations 1 or 2. This was performed in polar coordinates for the front (apical) half of the heart and for Cartesian coordinates elsewhere, similar to the hybrid spheric-cylindric coordinate system employed by Van Train et al. 8 to construct polar perfusion maps. Thus from the middle of the heart to the valve plane, vertical searches were conducted along expected horizontal edges for VLA images, whereas horizontal searches were performed along vertical edges of HLA images. HLA points were tested for left/right discrepancies caused by detection of the right ventricle in- stead of the left ventricle, and VLA points were tested for upper/lower discrepancies caused by detection of intense

478 Nichols et al. JOURNAL OF NUCLEAR CARDIOLOGY Gated tomographic left ventricular ejection fraction NovembeffDecember 1996, Part 1

C o r r e l a t i o n c o e f f i c i e n t ( r ) 7

0.95

o.% ,o go 3'0 20 go ,'o , o

N u m b e r o f P a t i e n t s ( c u m m u l a t i v e )

(A)

C o r r e l a t i o n c o e f f i c i e n t ( r ) 1

o . o

0.85

0 8 o 1 o 2o 3o 4o 5 0

- 5 - 4 5 - 5 0 - 6 5 - 3 o 1 ~ 5 m b e r -75 o f P a t i e n t s

( i n c r e m e n t a l )

(a) Figure 3. A, Linear correlation coefficients are plotted for values of new users' EF, ED volume (EDV), and ES volume (ESV) against accumulated numbers of patients' values compared with those of experienced observer. B, More revealing trend is demonstrated when same data are plotted as moving average of most recent group of 25 patients processed.

splanchnic, hepatobiliary, or intestinal activity instead of inferior wall counts.

Algorithms ensured that sets of endocardial points and sets of independently derived valve plane points extended to meet, but did not overshoot, one another. An algorithm then determined whether the collection of endocardial points in combination with valve plane points formed a closed, unam- biguous region of interest. Nonessential attached tendrils and isolated islands of points were eliminated, and a final attempt was made to span remaining pixel gaps by fine mesh interpo- lations. The final border was displayed to the observer (Figure 2, D), who modified the edge if necessary. A parameter representing the ratio of final accepted points to suggested automatic points was stored as part of the output file contain- ing the region. With slight differences appropriate to the selected view, all of these steps were applied in turn to the ED and ES VLA and HLA frames.

Volume Computations. Endocardial VLA heights were combined with HLA widths at each location from apex to base to generate a series of short-axis ellipsoid cylinders for com- putation of LV volume. To account for the possibility of different apex-to-base central axis dimensions from the two VLA and HLA views, their common midpoint was computed and the minimum apex-to-base length was used for both VLA and HLA offsets. EF was subsequently computed from the difference between ED and ES volumes.

The four frames, with and without endocardial borders, were presented for the observer's review simultaneously with monochrome and standardized color tables, along with com- puted volumes and EF results. In addition, four quality assur- ance parameters were stored for each of the four images consisting of the number of pixets (if any) the observer moved the center, the automatically suggested ED and ES frame numbers, the final accepted ED and ES frame numbers, and the percent of automatically generated endocardial border loca- tions that were accepted as final points for computation of volumes and EE

Patient Validations. Equilibrium GBP studies were ac-

quired for 75 adult patients (age 65 + 11 years; male/female ratio 50/25) who subsequently also underwent GSPECT stud- ies for the evaluation of chest pain or known coronary artery disease. Average GBP administered activity was 740 MBq (20 mCi) red blood cells labeled by the in vitro technique. All GBP data were analyzed independently without regard to GSPECT EF determinations by previously validated semiautomated algorithms, a~,'2

Further comparisons of EF were accomplished with a single crystal Anger camera to image the gated first-pass (GFP) transit of the 740 MBq (20 mCi) MIBI bolus used for 65 patients (age 63 + 12 years; male/female ratio 38/27) un- dergoing the resting portion of a 2-day MIBI perfusion proto- col. 7 GFP studies were discarded if bolus transit time was greater than 2.0 seconds or lung transit time was greater than 10 beats/sec. GFP data were analyzed independently of GSPECT computations by software previously validated against both GBP and multicrystal first-pass gamma camera correlative studies. 13

Data Processing Reproducibility. All 140 GBP and GFP patient studies were analyzed independently by two observers at the core laboratory. From the total of 140 GBP and GFP studies were selected a subset of 25 patient studies for which ED volume occurred in equal numbers at 50 ml intervals spanning an ED volume range of 50 to 300 ml simultaneously with equal numbers of patients at 10% EF intervals ranging from 20% to 80%. These studies were recoded to remove all identifying patient information and sent to three separate laboratories, where a total of six observers analyzed data without knowledge of each others' or of the core laboratory' s results.

Statistical Methods. All numeric results are reported here as mean values +1 SD. Linear regression analysis was used to compare calculations of LV volumes and EF between modalities and results obtained among independent observers. Means and variances were input to t tests to compute the two-tailed probabilities that measurements were drawn from the same sample populations.

JOURNAL OF NUCLEAR CARDIOLOGY Nichols et al. 479 Volume 3, Number 6;475-482 Gated tomographic left ventricular ejection fraction

Table 1. Accuracy: GSPECT versus

equi l ibr ium GBP

n • Intercept Slope SEE (%)

Auto 75 0.87 0.00 1.01 7.7 Manual 67 0.85 0.01 0.97 8.1

Table 3. Reproducibility: GSPECT

interobserver a g r e e m e n t

n • Intercept Slope SEE

EF 140 0.92 0.07 0.86 6.4o/o ED volume 140 0.98 1.31 0.94 11.2 ml ES volume 140 0.98 2.29 0.92 8.6 ml

Table 2. Accuracy: GSPECTversus ga t ed

first pass

n • Intercept Slope SEE (%)

Auto 65 0.87 0.14 0.76 8.0 Manual 65 0.88 0.07 0.79 7.7

Table 4. Reproducibility: GSPECT

intraobserver a g r e e m e n t

n • Intercept Slope SEE

EF 126 0.94 0.03 0.99 5.8% ED volume 126 0.99 -4.9 1.01 10.9 ml ES volume 126 0.99 -4 .4 0.99 10.7 ml

Table 5. lnterinst i tut ion GSPECT EF variability

Observer • Intercept Slope SEE (%)

To determine the influence of manual interventions to automatic choices, the GBP data were subdivided according to whether alterations were made to LV center location or en- docardial borders, and the resulting EF values were compared with GBP EF by linear regression. To evaluate measurement sensitivity to LV size, GSPECT EF was computed for GBP data subdivided into groups according to ED volume greater or less than median ED volume and compared with GBP EE

R ~ U ~ S

Manual C o m p u t a t i o n s . Manual computations by

new users of LV volumes and EF were compared with an experienced observer's values by linear regression

analysis as cumulative groups in increments of 10

(Figure 3, A). These same data were also compared as

updated groups of the most recent 25 patients studied (Figure 3, B). The plateau for comparisons of EF was

r = 0.96 attained with 40 to 65 patients, at which point

the linear correlation coefficients also stabilized for ED

volume (r = 0.95) and ES volume (r = 0.98).

A u t o m a t i o n Accuracy . For equilibrium GBP stud-

ies, GBP EF = 49% + 13% (median = 49%; range = 22%

to 77%) and GSPECT EF -- 49% + 15% (median = 49%;

r a n g e = 1 5 % to 83%), so that t=0 .123 (p=0 .902) . Regression analysis yielded r = 0 . 8 7 for automatic

GSPECT and r = 0.85 for manual GSPECT versus GBP EF (Table 1).

For the GFP data, GFP E F = 5 4 % _ + 1 8 % (med ian=57%; r ange= 14% to 88%) and GSPECT EF = 56% _+ 16% (median = 59%; range = 18% to 79%), for which t = 0 . 6 4 0 (p =0.523). Regression analysis yielded r = 0.87 for automatic GSPECT and r = 0.88 for manual GSPECT versus GFP EF (Table 2).

Comparisons of linear regression analysis GSPECT

1 0.96 0.02 0.96 4.8 2 0.93 0.00 0.98 7.0 3 0.95 0.01 0.96 5.8 4 0.93 0.02 0.87 6.0 5 0.95 0.01 0.95 5.7 6 0.95 -0.04 1.05 6.1

to equilibrium GBP were essentially the same when

patients were subdivided according to whether changes

were needed for LV center location (r = 0.91 vs r = 0.88) or endocardial region generation (r = 0.87 vs r = 0.86)

and were identical (r = 0.89) for patients grouped ac-

cording to ED volume above or below a median ED volume of 87.5 ml.

A u t o m a t i o n Prec i s ion . Interobserver variability

for two observers independently processing all 140 GBP

and GFP patient data demonstrated r = 0.92 for EF, with

a standard error of the estimate (SEE) of 6.4% (i.e., 6.4 EF units). Interobserver linear correlation was r = 0.98

for both ED volume (SEE = 11.2 ml) and ES volume (SEE = 8.6 ml) (Table 3). For an entirely separate group

of 126 patients, automated and manual results were compared for one observer, such that r = 0.94 for EF (SEE = 5.8%) (Table 4).

All LV EF calculations combined from the six independent observers with the automated software to analyze the subgroup of 25 studies demonstrated overall agreement with the core laboratory of r = 0.92. Indi- vidual observers' linear correlations with the core labo-

480 Nichols et al. JOURNAL OF NUCLEAR CARDIOLOGY Gated tomographic left ventricular ejection fraction November/December 1996, Part 1

Figure 4. Incorrect starting center location in middle of horizontal long-axis lateral wall can confound rules used for locating endocardium on ED (A) and ES (B) images. Altering starting ventricular center can provide more amenable input to endocardial border generator for successful edge detection of ED (C) and ES (D) frames.

ratory ranged from r = 0.93 to r = 0.96 for LV EF and are given in detail in Table 5.

Automation Success Rate. While executing the automated software for processing the GSPECT data of the 140 GBP and GFP studies, experienced observers judged it necessary to alter (1) ED or ES frames in 7% of patients, for which differences in frame estimates never exceeded one out of eight frames; (2) endocardial borders in 14%, for which failure of edge generation occurred most frequently in conjunction with automated LV center failure; and (3) VLA or HLA LV cavitary center locations in 28%. Automated center failure usu- ally occurred in the presence of intense spleen, liver, or intestinal counts and less frequently was due to extreme myocardial count asymmetry associated with severe regional hypoperfusion but was unrelated to EF magni- tude. Altering the automated center at the beginning of program execution frequently resulted in subsequent successful endocardial border generation (Figure 4).

D I S C U S S I O N

This study establishes the validity of our automated methods. Agreement of automatic GSPECT to first-pass imaging EF was comparable to values reported by others. ~4'15 Agreement to the separate, widely accepted, equilibrium GBP "gold standard" for LV EF was also

high. Measurement reproducibility was excellent, with a high degree of interobserver and intraobserver agree- ment between manual and automated results. The learn- ing curve for agreement among observers who used manual techniques after 50 patients was consistent with that attained by six independent observers by the auto- mated method. Thus we conclude that automation en- abled novice operators to obtain the same level of reproducibility as that attained by users of manual techniques once they had gained experience. Automation also should aid others in the successful implementation of these techniques.

An automation success rate of endocardial border generation of 86% was found for these methods. Ger- mano et al.~5 reported a perfect automation success rate of 100%. Unfortunately, the only standards of success of edge automation in that analysis and in the one reported here were observers' judgments as to how well com- puted borders matched their visual impression of en- docardial wall locations. Because the methods reported here were designed to emulate clinicians' perceptions of wall locations from the outset, there was a high a priori probability of agreement with visual impressions of endocardial edges. Thereby the physician reading the studies has an immediately tangible display of endocar- dial borders against which to gauge his impression of wall locations and regional wall motion. Our hope in

JOURNAL OF NUCLEAR CARDIOLOGY Nichols et al. 481 Volume 3, Number 6;475-482 Gated tomographic left ventricular ejection fraction

striving to emulate physicians' drawings was to tap into the experience and knowledge of anatomy that cardiolo- gists bring to this process, so that the methods described in this article are in fact an "expert systems" approach to LV EF measurement with myocardial perfusion gated tomograms.

For the particular patient groups studied, a larger percentage of patients undergoing GFP (19%) had EFs greater than 70% than did members of the GBP group (5%). This is probably why linear regression analysis of GSPECT EFs yielded a slope of 0.76 versus GFP but a slope of 1.01, closer to unity, versus GBP. The highest hypercontractile EF values themselves may be expected to be relatively less precise, and this would have been more of a problem within the GFP group, for which EFs ranged up to 88%, than for the GBP group, for which the maximum EF was 77%.

It is likely that the methods reported here system- atically underestimate EF because of use of gating at only eight frames per cardiac cycle. Other groups have studied the influence of different time-sampling rates on accuracy of EF and found EF to be underestimated at eight frames per R-R interval but only by about 4% and consistently regardless of EF magnitude. 1°'~5 Previously it was assumed that gating at more than eight frames per R-R interval would too often provide inadequate myo- cardial counts per frame for gated tomographic EF analysis. However, recent studies have shown that SPECT LV EF computations with the methods described here are reproducible ( r= 0.93) for gated tomograms acquired with 333 MBq (9 mCi) of MIBI versus 1.11 GBq (30 mCi), '6 suggesting that gating with 16 frames per R-R interval will be feasible routinely.

An advantage of this method has been its success for patients with small or hypercontractile ventricles, achieved through variable-count thresholds. For patients with large, hypocontractile hearts a fixed 35% threshold, compared with maximum myocardial counts, consis- tently emulates the visual perception of the endocardial edge. However, the larger exponential power was needed in the ES formula (equation 2) to account for the greater tendency at ES for counts to spill into the LV cavity from bright, thickened myocardial walls.

Our preliminary experience was that there was a sufficiently large percentage of all patients studied by GSPECT who had large, severely hypoperfused apical territories, so that modeling the myocardium from the gated apical short-axis slices would likely prove futile. By fusing the HLA and VLA information, we are guaranteed that the apical cap is represented and should provide the highest likelihood of reliable input for modeling these regions. This also avoids problems of myocardial walls moving into and out of fixed short-axis imaging planes.

As for the question of whether geometry-based models are as accurate as count-based models, which rely on partial volume e f f e c t s , 9'1°'~4 a recent preliminary comparison demonstrated substantially better LV EF agreement with x-ray contrast angiography of a paired HLA and VLA geometric technique with image inver- sion than did a count-based method as applied to the short-axis cines. 17 That study's observation of failure of count-based techniques under circumstances of severe apical hypoperfusion was echoed by another recent report comparing percent regional myocardial thicken- ing to measurements of magnetic resonance imaging in canines, is Thus geometric models may provide more accurate global EFs and a more directly verifiable assessment of regional wall translation in cases of severe hypoperfusion than do the count-based approaches. The method presented in this article also lends itself readily to contrast-enhancement techniques specific to severe hypoperfusion, for the production of highly reproducible LV EF measurements. 19

Future Directions. Our results indicate that current technology already affords a high degree of accuracy and reproducibility of LV EF from MIBI gated tomog- raphy. Several lines of investigation are likely to im- prove measurement accuracy even further: (1) attenua- tion and scatter corrections2°; (2) image restoration techniques to provide better myocardial wall visualization21; (3) improvements in clinical computer capacity to allow incorporation of short-axis myocardial sampling with fused VLA and HLA information; (4) improved cardiac-specific collimator design to improve sensitivity and spatial resolution22; and (5) contrast enhancement to optimize identification of severely hy- poperfused myocardium. 19

References

1. Ladenheim ML, Pollock BH, Rozanski A, et al. Extent and severity of myocardial hypoperfusion as predictors of prognosis in patients with suspected coronary artery disease. J Am Coil Cardiol 1986;7:464-71.

2. Shah PK, Maddahi J, Staniloff HM, et al. Variable spectrum and prognostic implications of left and right ventricular ejection fractions in patients with and without clinical heart failure after acute myocardial infarction. Lancet 1988;6:255-65.

3. Maddahi J, Garcia EV, Berman DS, et al. Improved uoninvasive assessment of coronary artery disease by quantitative analysis of regional stress myocardial distribution and washout of TI-201. Circulation 1981;64:924-35.

4. Schelbert HR, Verba JW, Johnson AD, et al. Nontraumatic deter- mination of left ventricular ejection fraction by radionuclide angiocardiography. Circulation 1975;51:902-9.

5. Garcia E, Cooke CD, Van Train KF, et al. Technical aspects of myocardial SPECT imaging with technetium-99m sestamibi. Am J Cardiol 1990;66:23E-32E.

6. DePuey EG, Nichols K, Dobrinsky C. Left ventricular ejection fraction assessed from gated technetium-99m-sestamibi SPECT. J Nucl Ned 1993;34:1871-6.

482 Nichols et al. JOURNAL OF NUCLEAR CARDIOLOGY Gated tomographic left ventricular ejection fraction November/December 1996, Part 1

7. DePuey EG, Nichols KJ, Slowikowski JS, et al. Fast stress and rest acquisitions for technetium-99m-sestamibi separate-day SPECT. J Nucl Med 1995;36:569-74.

8. Van Train KF, Areeda J, Garcia EV, et al. Quantitative same-day rest-stress technetium-99m sestamibi SPECT: definition and vali- dation of stress normal limits and criteria for abnormality. J Nucl Med 1993;34:1494-502.

9. Hoffman EJ, Huang SC, Phelps ME. Quantitation in positron emission computed tomography, I: effects of object size. J Comput Assist Tomogr 1979;5:391-400.

10. Cooke CD, Garcia EV, Cullom SJ, Faber TL, Pettigrew RI. Determining the accuracy of calculating systolic wall thickening using a fast Fourier transform approximation: a simulation study based on canine and patient data. J Nucl Med 1994;35:1185- 92.

11. Reiber JHC, Lie SP, Simoons ML, et al. Clinical validation of fully automated computation of ejection fraction from gated tomographic equilibrium blood-pool scintigrams. J Nucl Med 1983;24:1099-107.

12. Cttristian PE, Nortmaan CA, Taylor A. Comparison of fully automated and manual ejection fraction calculations: validation and pitfalls. J Nucl Med 1985;26:775-82.

13. Nichols K, DePuey EG, Gooneratne N, Salensky H, Friedman M, Cochoff S. First pass ventricular ejection fraction using a single crystal nuclear camera. J Nucl Med 1994;35:1292-300.

14. Vansant JP, Faber TL, Folks RD, Rao JM, Garcia EV. Resting left ventricular volumes and ejection fraction from gated SPECT: correlation to first pass [abstract]. Circulation 1995;92:1-11.

15, Germano G, Kiat H, Kavanaugh PB, et al. Automatic quantifica- tion of ejection fraction from gated myocardial perfusion SPECT. J Nucl Med 1995;36:2138-47.

16, Nichols K, DePuey EG, Salensky H, Rozanski A. Reproducibility of ejection fractions from stress versus rest gated perfusion SPECT [abstract]. J Nucl Med 1996;37:115P.

17. Williams KA, Taillon LA. Five methods for evaluation of left ventricular function with Tc-99m-sestamibi: a comparison with contrast ventriculography [abstract]. J Nucl Med 1996;37:104P.

18. Gradel C, Staib LH, Heller EN, et al. Limitations of ECG-gated SPECT for assessment of regional thickening: experimental com- parison with ECG-gated MRI [abstract]. J Am Coil Cardiol 1996;27:241A.

19. Nichols K, DePuey EG, Salensky H, Rozanski A. Image enhance- ment of severely hypoperfused myocardia for computation of tomographic ejection fraction [abstract]. J Nucl Med 1996;37: 105P.

20. Ficato EP, Fessler JA, Shreve PD, et al. Simultaneous transmis- sion/emission myocardial perfusion tomography: diagnostic accu- racy of attenuation-corrected 99mTc-sestamibi single-photon emis- sion computed tomography. Circulation 1996;93:463-73.

21. Lange K, Carson R. EM reconstruction algorithms for emission and transmission tomography. J Comput Assist Tomogr 1984;8: 306-16.

22. Gullberg GT, Gengsheng LZ. A cone-beam filtered backprojection reconstruction algorithm for cardiac single photon emission com- puted tomography. IEEE Trans Med Imaging 1992;11:91-101.

AVAILABILITY OF JOURNAL BACK ISSUES As a service to our subscribers, copies of back issues of Journal of Nuclear Cardiology for the preceding 5 years are maintained and are available for purchase from Mosby until inventory is depleted at a cost of $20.00 per issue. The following quantity discounts are available: 25% off on quantities of 12 to 23, and one third off on quantities of 24 or more. Please write to Mosby-Year Book, Inc., Subscription Services, 11830 Westline Industrial Dr., St. Louis, MO 63146-3318, or call (800) 453-4351 or (314) 453-4351 for information on availability of particular issues. If unavailable from the publisher, photocopies of complete issues may be purchased from UMI, 300 N. Zeeb Rd., Ann Arbor, MI 48106 (313) 761-4700.