Quantitative myocardial-perfusion SPECT: Comparison ... - Springer Link

ORIGINAL ARTICLES Quantitative myocardial-perfusion SPECT: Comparison of three state-of-the-art software packages Arik Wolak, MD,a Piotr J. Slomka, PhD,b Mathews B. Fish, MD,c Santiago Lorenzo, MS,d Wanda Acampa, MD,a Daniel S. Berman, MD,a and Guido Germano, PhDb Background. We aimed to compare the automation and diagnostic performance in the detection of coronary artery disease (CAD) of the 4DMSPECT (4DM), Emory Cardiac Toolbox (EMO), and QPS systems for automated quantification of myocardial perfusion. Methods and Results. We studied 328 patients referred for rest/stress Tc-99m sestamibi imaging, 140 low-likelihood patients and 188 with angiography. Contours were corrected when necessary. All other processing was fully automated. A 17-segment analysis was performed, and a summed stress score (SSS) >4 was considered abnormal. The average SSSs (ⴞSD) for 4DM, EMO, and QPS were 10.5 ⴞ 9.4, 11.1 ⴞ 8.3, and 10.1 ⴞ 8.9, respectively (P ⴝ .02 for QPS versus EMO). The receiver operator characteristics areas-under-the-curve for the detection of CAD (ⴞSEM) were 0.84 ⴞ 0.03, 0.76 ⴞ 0.04, and 0.88 ⴞ 0.03 for 4DM, EMO, and QPS, respectively (P ⴝ .001 for QPS versus EMO, and P ⴝ .03 for 4DM versus EMO). Normalcy rate was higher for QPS and 4DM versus EMO, at 91% and 94% versus 77%, respectively (P ⴝ .02). Sensitivity was higher for QPS (87%) versus 4DM (80%) (P ⴝ .045). Specificity was higher for QPS (71%) versus EMO (49%) (P ⴝ .01). The accuracy rate was higher for QPS versus 4DM and EMO, at 83% versus 77% and 76%, respectively (P ⴝ .05). Conclusions. There are differences in myocardial-perfusion quantification, diagnostic performance, and degree of automation of software packages. (J Nucl Cardiol 2008;15:27-34.) Key Words: Myocardial perfusion imaging • SPECT • automatic quantification • software • coronary artery disease The visual interpretation of myocardial-perfusion single-photon computerized tomography (MPS) studies is subjective and dependent on observer expertise.1 In an effort to standardize MPS interpretation, a segmental visual scoring approach was developed, initially using 20 segments,2 and later modified to the currently recommended 17 segments.3 However, the manual scoring of MPS is time-consuming and suffers from interobserver and intraobserver variability.4 Several clinically valiFrom the Department of Imaging and Medicine,a and Artificial Intelligence in Medicine/Department of Imaging,b Cedars-Sinai Medical Center, Los Angeles, Calif; Oregon Heart and Vascular Institute,c Sacred Heart Medical Center, Eugene, Ore; and Department of Human Physiology,d University of Oregon, Eugene, Ore. A.W. is a Fellow of the Save A Heart Foundation, Inc., at Cedars-Sinai Medical Center. Received for publication June 18, 2007; final revision accepted Sept 2, 2007. Reprint requests: Piotr J. Slomka, PhD, Artificial Intelligence in Medicine/Department of Imaging, Cedars-Sinai Medical Center, 8700 Beverly Blvd, A047, Los Angeles, CA 90048; [email protected]. 1071-3581/$34.00 Copyright © 2008 by the American Society of Nuclear Cardiology. doi:10.1016/j.nuclcard.2007.09.020

dated computer software packages were developed to provide a fully automated quantitative assessment of relative regional myocardial perfusion.5-8 The 4D-MSPECT (4DM), Emory Cardiac Toolbox (EMO), and Cedars Quantitative Perfusion SPECT (QPS) are three commercially available software packages in common use. Unlike the automatic assessment of ejection fraction from a gated MPS, the quantification of perfusion with these various approaches has not been extensively compared. In one study, considerable variability in the perfusion quantification was reported.9 We aimed to perform a detailed comparison of the latest versions of these three software tools with respect to automation and diagnostic performance in the detection of coronary artery disease (CAD) in a large group of patients with available coronary angiography results, and in patients with a low likelihood of CAD. MATERIALS AND METHODS Patients The subjects for this study were selected from 2450 patients referred to the Department of Nuclear Medicine, 27

28

Wolak et al Quantitative myocardial-perfusion SPECT

Journal of Nuclear Cardiology January/February 2008

Table 1. Patient characteristics

Parameter

Database group (N ⴝ 100)

Low-likelihood group (N ⴝ 140)

Angiography group (N ⴝ 188)

Age (y) Female Body mass index (kg/m2) Body mass index ⱕ30 Diabetes Hypertension Hypercholesterolemia Chest pain Adenosine

55.9 ⫾ 13.1 50 (50%) 30.9 ⫾ 6.3 50 (50%) 16 (16%) 50 (50%) 45 (45%) 14 (14%) 0 (0%)

52.5 ⫾ 11.9 97 (69%) 28.3 ⫾ 6.2 106 (76%) 11 (8%) 50 (36%) 53 (38%) 41 (30%) 0 (%)

63.5 ⫾ 11.6 58 (31%) 30.2 ⫾ 6.0 104 (55%) 49 (27%) 11 (63%) 91 (49%) 115 (61%) 143 (76%)

Data are presented as mean ⫾ SD.

Sacred Heart Medical Center (Eugene, OR), from March 1, 2003 to February 29, 2004, for rest/stress gated MPS. All studies were performed using a rest sestamibi/stress sestamibi protocol, in which MPS data were acquired with and without attenuation correction. For the purpose of this study, only data without attenuation correction were used, since only noncorrected normal limits were available for all three packages examined. The Institutional Review Board at Sacred Heart Medical Center approved the retrospective use of clinical data in this study. Clinical and imaging data with all personal identifiers removed were transmitted to Cedars-Sinai Medical Center (Los Angeles, CA) for analysis.

Low-likelihood group A low likelihood (LLK) of CAD (⬍5%) was defined based on age, sex, pretest symptoms, and electrocardiogram response to treadmill stress testing.10 Accordingly, low-likelihood patients were selected only from patients who underwent treadmill stress testing (n ⫽ 1102), and who had an adequate level of treadmill stress (ⱖ85% of predicted maximum heart rate). These patients had no history of CAD (a previous myocardial infarction or coronary revascularization) or other confounding cardiac conditions, including congestive heart failure, cardiomyopathy, significant valvular or congenital heart disease, left-bundle branch block, or paced rhythm. These patients did not undergo coronary angiography. Furthermore, these subjects had MPS studies of good to excellent quality, normal ventricular volumes, wall motion, and global systolic function, and no evidence of transient ischemic dilatation, as judged by the director of the MPS laboratory where the data were acquired (M.F.). The final group for the evaluation of normalcy rates consisted of 140 consecutive patients (97 females, 43 males) meeting the criteria.

with an adequate stress test and coronary angiography within 60 days without an intervening clinical cardiac event. Of these, 144 (76%) had adenosine stress. The 44 patients who had treadmill stress achieved a heart rate of 87% ⫾ 5.1% of the maximum predicted heart rate for age at the time of injection of the radiotracer. The consecutive catheterized population of this study excluded patients with bypass surgery, cardiomyopathy, valvular disease, left-bundle branch block, and paced rhythm, and also excluded seven additional patients with an insufficient quality of their MPS. The director of the MPS laboratory where the data were acquired (M.F.) made exclusion for image quality. The clinical characteristics of the database group, the LLK group, and the angiography group are summarized in Table 1.

Rest and stress imaging protocols Patient preparation included abstention from caffeinecontaining drinks, food, and medication for 24 hours, and from methylxanthine medications for 36 hours. Patients had to be free of short-acting nitrates for 2 hours, long-acting nitrates for 6 hours, calcium blockers for 24 hours, and beta-blockers for 48 hours before testing. Studies were performed using standard 99m Tc rest/99mTc stress protocols. A same-day rest/stress protocol was used for women who weighed ⬍91 kg (200 pounds) or whose body mass index (BMI) was ⬍35, and for men who weighed ⬍113 kg (250 pounds) or whose BMI was ⬍40. A 2-day rest/stress or stress/rest protocol was used for individuals whose weight or BMI levels were above these levels. The weight/BMI-related 99mTc-sestamibi dose schedule ranged from 314 to 429 MBq for rest MPS, and from 1091 to 1554 MBq for stress MPS. Two-day protocols used the “stress” dose for both the rest and stress portions of the study.

Stress testing Angiographic group The group for the evaluation of diagnostic sensitivity, specificity, and accuracy for the detection of CAD consisted of 188 consecutive patients (58 females, 130 males) who had MPS

Patients undergoing exercise stress underwent a symptomlimited treadmill test using the standard Bruce protocol. At near-maximum exercise, 99mTc-sestamibi was injected intravenously. Exercise was continued at the maximum workload for

Journal of Nuclear Cardiology Volume 15, Number 1;27-34

1.5 to 2.0 minutes when possible. When exercise testing was contraindicated or unsuitable, a pharmacologic stress test was performed by an infusion of adenosine at 140 ␮g · kg⫺1 · min⫺1 for 5 minutes. At the end of the second minute, 99m Tc-sestamibi was injected. If they were able, patients, during adenosine infusion, performed low-level treadmill exercise at 0% grade at 0.8 to 1.7 mph.

Image acquisition Image acquisition was started at 60 minutes after the administration of 99mTc-sestamibi at rest, or during adenosine infusion with the patient at rest, and at 15 to 45 minutes after radiopharmaceutical injection during treadmill testing, or adenosine infusion with low-level exercise. The patients drank 16 ounces of water immediately prior to imaging. The MPS was acquired using Phillips Vertex (Philips Medical Systems, Milpitas, Calif) dual-detector scintillation cameras with low-energy, high-resolution collimators. Images were acquired over a 180° noncircular orbit from 45° right anterior oblique to left posterior oblique, with a 64 ⫻ 64 matrix (pixel size, 0.64 cm) for emission images. The energy window was 140 keV ⫾ 20%. The time per projection used in this study was 45 to 50 seconds for rest MPS in the same-day protocol, and 30 to 40 seconds for the stress MPS and for the rest MPS in the 2-day protocol. The number of projections was 64.

Tomographic reconstruction Tomographic reconstruction was performed using AutoSPECT software on a Philips system (Philips Medical Systems). All emission images were automatically corrected for nonuniformity, radioactive decay, and motion during acquisition, and subjected to 3-point spatial smoothing. The mechanical center of rotation was determined to align the projection data to the reconstruction matrix. Readings were reconstructed by filtered back-projection (FBP) with Butterworth filters as follows: rest MPS order 10, cutoff 0.50; stress MPS order 5, cutoff 0.66. Short-axis images were generated in the DICOM format.

Quantitative software tools Quantitative analysis was performed on e.soft express cardiac, a comprehensive workstation supplied by Siemens Medical (e.soft™ Cardiac Workflow, Siemens Medical Solutions USA, Malvern, Pa) that allows for cardiac single-photon emission computerized tomography (SPECT) or positron emission tomography (PET) review and reporting. This workstation can include three popular state-of-the-art, quantitative cardiac software tools: 4D-MSPECT, version 3.1, updated July 2004, University of Michigan Medical Center (4DM). Emory Cardiac Toolbox, version 4, revision 01, October 2004, Emory University Medical Center (EMO). Cedars Cardiac Suite, e.soft version 4.0, July 2004, CedarsSinai Medical Center (QPS).


29

These three software packages were selected for comparison given their high popularity, automated quantitative features, and potential to process the same reconstructed data.

Quality control Short-axis reconstructed data were transferred to the e.soft express workstation, where 4DM, EMO, and QPS were installed. The same reconstructed short-axis images were loaded to all three packages. An experienced senior nuclear cardiology core laboratory technologist (James Gerlach) and a seniorphysician nuclear medicine specialist (W.A.), blind to all results including those of the patient group studies (normal database, normalcy rate, and accuracy groups), examined the short-axis images and manually corrected them for proper left-ventricular positioning, within each software package, when necessary. In three cases of disagreement, a consensus was achieved with an additional reader (A.W.). All corrections were made in strict adherence to the contour correction instructions within each specific package manual. Corrections were made only for major discrepancies, to avoid any unnecessary manipulation of the data. The process included insertion of the left ventricular (LV) center on axial, vertical long axis (VLA), and horizontal long axis (HLA) images, followed by masking of the region outside of the LV in case of extra cardiac activity. Whenever possible, the precorrection data were saved. All other processing was fully automated, to assure maximal reliability of the data. An additional analysis was performed without any contour corrections in the LLK group, to remove any possible bias due to contour adjustment, and to assess the performance of fully automated software processing (“handsoff”) for each package. To minimize the chances of data-collection error, the laboratory acquiring the studies employed a strict study regimen, which included meticulous attention to detail in following the patient preparation, stress testing and injection, image acquisition, and reconstruction study protocols. In addition, we confined manual intervention during processing to the adjustment of the LV boundaries, and this intervention was performed as infrequently as possible and in a blinded fashion by the operators. The possibility of unintended bias could have been introduced by the processing technologist employed in the CedarsSinai nuclear cardiology core laboratory; he does not, however, participate in Cedars-Sinai software development, and he was tasked with providing the best possible results with each software package. He otherwise did not participate in the analysis of our data or in the study in general. To minimize this kind of bias, the analysis was automatic, except for the alteration of obviously incorrect boundaries. Patient selection was sequential, and was controlled at the site by an investigator not affiliated with Cedars-Sinai.

Normal databases Vendor databases. Using the database menu in each package, databases that matched the described acquisition protocol were selected. For 4DM, we selected a high-dose

30


99m

Tc-sestamibi database specific to uncorrected (NC) female patients (V2-GSRD/TC/NC/F), or a high-dose 99mTc-sestamibi database specific to NC male patients (V2-GSRD/TC/NC/M), for females and males, respectively. For EMO, we chose the 1-day rest/stress sestamibi or 2-day sestamibi, and for QPS we used the standard Cedars-Sinai stress MIBI/rest MIBI genderspecific databases. Custom databases. In addition to the main analysis using the vendor-supplied databases, we aimed to compare the performances of 4DM and QPS by using custom databases built from scans of matching, normal, low-likelihood patients. The EMO does not provide such a custom database generator feature. To create the custom database group, 100 patients (50 women and 50 men) were selected from the 240 consecutive patients with a low likelihood of CAD (as previously described). All had to demonstrate no significant extra cardiac radioactivity overlapping with LV, and had to produce visually normal rest/stress MPS images. Low-likelihood criteria could only be determined for exercise studies (because the clinical and normal electrocardiogram responses to adenosine do not alter the likelihood of CAD). Therefore, only these studies were used to build normal databases. We previously showed, however, that stress exercise or adenosine can be interchanged and can be applied to the test population where adenosine stressing or exercise are performed.6 Using the Normal Database Generator feature in 4DM11 and the Database Editor function for QPS,12 gender-specific normal limits were created from exactly the same group of patients as described above.

Data analysis All results were saved automatically to an Excel-compatible format and then exported directly to an Excel file (Excel™, Microsoft Corporation, Redmond, Wash). No manual data entry of results was used at any point in this processing. Identical short-axis cases were used for these three packages.

Perfusion quantification The methods of quantification were previously described for the three systems.6,13,14 The following quantitative results were obtained with each software package: summed stress score (SSS), total defect extent (TDE), and regional defect extent (RGE) for the left anterior descending artery (LAD), left circumflex artery (LCX), and right coronary artery (RCA).


Table 2. Angiographic characteristics

Parameter

N (%)

No stenosis* Single-vessel disease Double-vessel disease Triple vessel disease LAD LCX RCA

45 (24) 61 (32) 41 (22) 41 (22) 102 (54) 75 (40) 89 (47)

LAD, Left anterior descending artery; LCX, left circumflex artery; RCA, right coronary artery. *A stenosis of ⱖ70% luminal-diameter narrowing was considered significant.

Coronary angiography Coronary angiography was performed with the standard Judkins approach, and experienced physicians interpreted all coronary angiograms visually. A stenosis of ⱖ70% luminaldiameter narrowing was considered significant. Two conventional cutoff points of CAD are used in many studies, ie, ⱖ70% and ⱖ50% diameter stenosis. However, because previous research found the best agreement between nuclear data and coronary angiography when ⬎⫽70% diameter stenosis was used,15 we chose to use that cutoff point in our study. Coronary angiographic findings are presented in Table 2.

Statistical analysis All continuous variables are expressed as means ⫾1 standard deviation. Paired t tests were used to compare differences in paired continuous data, and McNemar tests were used to compare differences in paired discrete data. All statistical tests were two-tailed, and a value of P ⬍ .05 was considered significant. For the assessment of agreement, correlation and bias plots were calculated. Receiver operator characteristics (ROC) curve analysis was performed to evaluate the three software packages for predicting ⱖ70% stenoses of coronary arteries. The ROC curves were created using all possible integer score values and a step of 0.1% for the defect extent values. The correlation, bias plots, and differences between ROC curve areas (area ⫾ standard error) were compared using version 1.71 of the Analyze-It statistical package from Analyze-It Software, Ltd. (Leeds, United Kingdom) and the pairedcomparison method for ROC curves using the method of Hanley and McNeil, which requires that all tests be performed on the same subjects.16

Abnormality threshold For the direct comparison of quantification results, we utilized any SSS ⱖ4 to define the quantitative threshold for MPS abnormality for all software packages, because this parameter is designed to mimic visual reading with each software package,2 and because this threshold was established previously in visual scoring2 and automatic scoring models.6

RESULTS Quality control The contours required adjustment in 12/328 (4.1%), 141/328 (43.9%), and 31/328 (9.5%) of the stress images using the 4DM, EMO, and QPS software, respectively



31

Figure 1. Correlations and bias plots (Bland-Altman analysis) for summed stress scores for the three software packages (4DM, EMO, and QPS).

(P ⫽ .001 for 4DM versus QPS, 4DM versus EMO, and QPS versus EMO). Quantitative results The mean SSS was 10.5 ⫾ 9.4, 11.1 ⫾ 8.3, and 10.1 ⫾ 8.9 for 4DM, EMO, and QPS, respectively (P ⫽ .02 for QPS versus EMO; P ⫽ NS for all other comparisons). Figure 1 shows the correlation and bias plots between the three packages. The correlation was lowest for EMO and 4DM (r ⫽ 0.68), and highest for QPS and 4DM (r ⫽ 0.84) (P ⬍ .01). The normalcy rate, sensitivity, specificity, and accuracy for all three packages obtained when applying the threshold of SSS ⱖ4 are shown in Figure 2. Normalcy was higher for QPS and 4DM versus EMO, the sensitivity was higher for QPS versus 4DM, the specificity was

Figure 2. Comparison of quantitative results of the three software packages (4DM, EMO, and QPS) for summed stress score abnormality threshold ⱖ4.

higher for QPS versus EMO, and the accuracy rate was higher for QPS versus 4DM and EMO. In the LLK group, when comparing analyses performed without and with contour correction, the normalcy rate was 66% versus 77%, 91% versus 91%, and

32


Figure 3. The ROC curves for detection of ⱖ70% stenosis, using summed stress score (SSS) (left) and total defect extent (TDE) (right) in vendor database.

Figure 4. The ROC curves for detection of ⱖ70% stenosis per coronary artery territory. LAD, left anterior descending artery; LCX, left circumflex artery; RCA, right coronary artery.

91% versus 94% for EMO, 4DM, and QPS, respectively (P ⫽ .05 for EMO comparison; P ⫽ NS for all other comparisons). The mean SSS without and with contour correction was 3.1 ⫾ 4.8 versus 1.8 ⫾ 2.9, 0.9 ⫾ 1.8 versus 0.9 ⫾ 1.8, and 1.2 ⫾ 1.7 versus 0.7 ⫾ 1.3 for EMO, 4DM, and QPS, respectively (P ⫽ .01 for EMO comparison, P ⫽ .01 for QPS comparison, and P ⫽ NS for 4DM comparison). Figure 3 shows the ROC curves for the detection of ⱖ70% stenosis by SSS and TDE. The areas-under-thecurve (AUC) for the SSS were significantly larger for QPS and 4DM versus EMO for the detection of ⱖ70% stenosis by both SSS and TDE. The regional performance for defect extent by coronary artery territory was also calculated, and is depicted in Figure 4. The ROC-AUC for the detection of CAD per vessel was similar for the three algorithms, except for the left anterior descending artery, which showed a larger AUC for QPS (0.81 ⫾ 0.03) versus EMO (0.73 ⫾ 0.04) (P ⫽ .02).


Figure 5. The ROC curves for detection of ⱖ70% stenosis using summed stress score (left) and total defect extent (right) in the custom database.

Figure 6. Case illustration of three patients. The polar map for the 4DM software is shown at right, for the EMO software at center, and for the QPS software at left. A, 79-year-old hypertensive woman with typical chest pain. The summed stress score (SSS) ⫽ 2, 10, and 10 for 4DM, EMO, and QPS, respectively. Angiography found multivessel disease (left anterior descending artery, 90%; left circumflex artery, 50%; and right coronary artery, 70%). B, 65-year-old woman with atypical chest pain and no risk factors. The SSS ⫽ 3, 17, and 1 for 4DM, EMO, and QPS, respectively. There was no significant coronary disease on angiography. C, 46-year-old woman with noncardiac chest pain and no risk factors. The SSS ⫽ 1, 13, and 0 for 4DM, EMO, and QPS, respectively.

We further analyzed the ROC curves for the detection of ⱖ70% stenosis by SSS and TDE, using custom databases for the 4DM and QPS packages. The EMO package was excluded due to its lack of a database creator feature. Figure 5 shows that the AUC for the SSS was significantly larger for QPS versus 4DM (0.89 ⫾ 0.03 versus 0.83 ⫾ 0.04, respectively, P ⫽ .03). The AUC for the TDE was statistically the same (0.88 ⫾ 0.03 versus 0.86 ⫾ 0.03, respectively, P ⫽ NS). Additional comparisons of the ROC curves of the 4DM vendor database versus the 4DM custom database, and of the QPS vendor database versus the QPS custom database, showed no significant differences in the AUC (0.83 ⫾


0.04 versus 0.84 ⫾ 0.03, and 0.89 ⫾ 0.03 versus 0.88 ⫾ 0.03, respectively, P ⫽ NS for both comparisons). Figure 6 provides polar maps of three examples that reflect our main findings, ie, a patient with CAD, a patient with no CAD, and a patient in the LLK group. DISCUSSION To the best of our knowledge, this is the first study of a head-to-head comparison of commercially available MPS software packages in terms of their ability to detect CAD, compared with invasive coronary angiography. Also, this is the largest study to compare the degree of automation and diagnostic performance of the abovementioned three software packages. The results show considerable differences in the quantification of perfusion using the three software packages. These results are in accordance with previous studies that found variation in both the automated quantification of reversible defects using the same three software packages9 and fixed defects using different software packages.17 Unlike previous studies, the present study used the same short-axis images for analysis by different software packages and automated batch mode processing, and had angiographic data for the verification of results. Differences in the quantification of perfusion were found for both the normalcy rates and the accuracy in detecting CAD. Furthermore, the pattern of differences was consistent, and showed the performance advantage of QPS and 4DM over EMO. The application of custom databases (for QPS and 4DM) did not significantly improve the performance of the software. The study also showed that the need for manual intervention in LV definition was more frequent with the EMO software than with the other two software packages. There are several possible explanations for the differences in results obtained with these three software tools. First, the algorithms used to generate the polar maps are different: QPS uses an elliptical sampling method for the whole LV,5 whereas the 4DM and EMO methods use cylindrical or spherical sampling schemes.18 In all cases of false-positive results by EMO, a significant myocardial perfusion defect in the apical region was detected, and the only difference between packages in the ROC-AUC was observed in the LAD territory. Thus the difference in performance could be related to the apical sampling techniques. Another reason for the discrepancies could involve the differences within the patient population used to derive the normal limits for the various methods with vendor databases. Furthermore, these packages use different approaches to generate the automatic segmental scores from the polar map data, to define the LV valve plane, and to normalize patient data to normal count distributions.6 The image


33

reconstruction method and the parameters applied in this study are similar to the reconstruction methods used in normal limit creation, as described in the respective software manuals. Therefore, the possibility that differences in reconstruction would significantly influence our results is low. Because the degree of automation was much greater for 4DM and QPS than for EMO, the manual constraint of borders was used more frequently with EMO, and may have interfered with the degree of detected perfusion defect. However, when we performed a fully automated “hands-off” analysis of the three software packages, the QC process significantly improved the normalcy rate for EMO, suggesting that manual correction has a positive effect on EMO performance. To minimize any possible bias, we used a fully automated method of analysis for comparison between the three software packages. However, in future studies, it might be interesting to compare other software packages to visual reading (similar to our recent study related to QPS in the Journal of Nuclear Cardiology6). The limitations of this study involve the use of angiography as a gold standard, which may not correspond to the physiologic significance of the defect and the need for some manual corrections (although this was addressed by including the results of a “hands-off” analysis). We could not compare the performance of EMO with the custom-made database, and therefore some differences may be attributable to the differences between normal limits. Also, the size of the study group may not have been large enough to show vessel-byvessel differences. CONCLUSIONS There are significant differences in the diagnostic performance and degree of automation of currently available software for the quantification of myocardial perfusion using MPS. Based on these differences, when patients have serial studies, or when results between different sites are compared, it appears advisable to use the same software for the quantification of myocardial perfusion.

Acknowledgment The authors thank James Gerlach, CNMT, who examined and corrected the MPS contours. Also, some of the authors (P.J.S., D.S.B., and G.G.) receive royalties from Cedars-Sinai Medical Center for licensure of the QPS software used in this analysis. The other authors indicate that they have no financial conflicts of interest. All raw data used in this work were acquired at the Sacred Heart Medical Center, and are available to any researchers for further analysis.

34


References 1. Akesson L, Svensson A, Edenbrandt L. Operator dependent variability in quantitative analysis of myocardial perfusion images. Clin Physiol Funct Imag 2004;24:374-9. 2. Berman DS, Kiat H, Friedman JD, Wang FP, van Train K, Matzer L, et al. Separate acquisition rest thallium-201/stress technetium99m sestamibi dual-isotope myocardial perfusion single-photon emission computed tomography: a clinical validation study. J Am Coll Cardiol 1993;22:1455-64. 3. Berman DS, Abidov A, Kang X, Hayes SW, Friedman JD, Sciammarella MG, et al. Prognostic validation of a 17-segment score derived from a 20-segment score for myocardial perfusion SPECT interpretation. J Nucl Cardiol 2004;11:414-23. 4. Verberne HJ, Habraken JB, van Royen EA, Tiel-van Buul MM, Piek JJ, van Eck-Smit BL. Quantitative analysis of 99Tcmsestamibi myocardial perfusion SPECT using a three-dimensional reference heart: a comparison with experienced observers. Nucl Med Commun 2001;22:155-63. 5. Germano G, Kavanagh PB, Waechter P, Areeda J, Van Kriekinge S, Sharir T, et al. A new algorithm for the quantitation of myocardial perfusion SPECT. I: technical principles and reproducibility. J Nucl Med 2000;41:712-9. 6. Slomka PJ, Nishina H, Berman DS, Akincioglu C, Abidov A, Friedman JD, et al. Automated quantification of myocardial perfusion SPECT using simplified normal limits. J Nucl Cardiol 2005;12:66-77. 7. Ficaro EP, Corbett JR. Advances in quantitative perfusion SPECT imaging. J Nucl Cardiol 2004;11:62-70. 8. Faber TL, Cooke CD, Folks RD, Vansant JP, Nichols KJ, DePuey EG, et al. Left ventricular function and perfusion from gated SPECT perfusion images: an integrated method. J Nucl Med 1999;40:650-9. 9. Svensson A, Akesson L, Edenbrandt L. Quantification of myocardial perfusion defects using three different software packages. Eur J Nucl Med Mol Imag 2004;31:229-32.


10. Diamond GA, Forrester JS. Analysis of probability as an aid in the clinical diagnosis of coronary-artery disease. N Engl J Med 1979;300:1350-8. 11. The Regents of the University of Michigan. Operating instructions, 4D-MSPECT with e.soft. Vol Publication #87 18 699 Revision 01 ed. Hoffman Estates, IL:Siemens;2004. 12. Germano G, Gerlach J. QPS & ARG— generic user manual. Report no. QPS-UM-GENERIC-001-rev3-02NL. Los Angeles: Artificial Intelligence in Medicine Program, Department of Medicine, Cedars-Sinai Medical Center, 2004. 13. Kritzman J, Ficaro E, Liu Y, Wackers F, Corbett J. Evaluation of 3D-MSPECT for quantification of Tc- 99m sestamibi defect size. J Nucl Med 1999;40:181P. 14. Vansant J, Krawczynska E, Shen Y, Folks RD, Kauppi AC, Garcia EV, et al. The prognostic value of quantitative indices of Tc- 99m sestamibi SPECT. J Nucl Med 1998;39:115P-6P. 15. Heller LI, Cates C, Popma J, Deckelbaum LI, Joye JD, Dahlberg ST, et al. Intracoronary Doppler assessment of moderate coronary artery disease: comparison with 201Tl imaging and coronary angiography. Circulation 1997;96:484-90. 16. Hanley JA, McNeil BJ. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology 1983;148:839-43. 17. De Sutter J, Van de Wiele C, D’Asseler Y, De Bondt P, De Backer G, Rigo P, et al. Automatic quantification of defect size using normal templates: a comparative clinical study of three commercially available algorithms. Eur J Nucl Med 2000;27: 1827-34. 18. Maddahi J, Van Train K, Prigent F, Garcia EV, Friedman J, Ostrzega E, et al. Quantitative single photon emission computed thallium-201 tomography for detection and localization of coronary artery disease: optimization and prospective validation of a new technique. J Am Coll Cardiol 1989;14: 1689-99.