Journal of Cardiovascular Computed Tomography (2009) 3, 372–382

Original Research Article

Automated 3-dimensional quantification of noncalcifiedand calcified coronary plaque from coronary CTangiography

Damini Dey, PhDa,b,*, Victor Y. Cheng, MDa, Piotr J. Slomka, PhD, FCCPM, FACCa,b,Ryo Nakazato, MDa, Amit Ramesh, MSca, Swaminatha Gurudevan, MDa,Guido Germano, PhD, FACCa,b, Daniel S. Berman, MD, FACCa,b

aDepartments of Imaging and Medicine, Cedars-Sinai Medical Center, 8700 Beverly Boulevard, Taper Building, A238, LosAngeles, CA 90048, USA and bDepartment of Medicine, David Geffen School of Medicine at UCLA, Los Angeles, CA, USA

KEYWORDS:Coronary arterysegmentation;Coronary CTangiography;Coronary plaque;Noncalcified and calcifiedplaque;Plaque characterization;Plaque quantification

Conflict of interest: The authors re

This study was supported by Americ

Award 09GRNT2330000, by a grant fr

part by grant 6318 from the Glazer Fou

* Corresponding author.

E-mail address: deyd@cshs.org

Submitted April 17, 2009. Accepte

2009.

1934-5925/$ -see front matter � 2009

doi:10.1016/j.jcct.2009.09.004

INTRODUCTION: We aimed to develop an automated algorithm (APQ) for accurate volumetric quan-tification of non-calcified (NCP) and calcified plaque (CP) from Coronary CT angiography (CCTA).

METHODS: APQ determines scan-specific attenuation thresholds for lumen, NCP, CP and epicardialfat, and applies knowledge-based segmentation and modeling of coronary arteries, to define NCP and CPcomponents in 3D. We tested APQ in 29 plaques for 24 consecutive scans, acquired with dual-source CTscanner. APQ results were compared to volumes obtained by manual slice-by-slice NCP/CP definition andby interactive adjustment of plaque thresholds (ITA) by 2 independent experts.

RESULTS: APQ analysis time was ,2 sec per lesion. There was strong correlation between the 2readers for manual quantification (r 5 0.99, p , 0.0001 for NCP; r 5 0.85, p , 0.0001 for CP). Themean HU determined by APQ was 419 6 78 for luminal contrast at mid-lesion, 227 6 40 for NCP upperthreshold, and 511 6 80 for the CP lower threshold. APQ showed a significantly lower absolute difference(26.7 mm3 vs. 42.1 mm3, p 5 0.01), lower bias than ITA (32.6 mm3 vs 64.4 mm3, p 5 0.01) for NCP. Therewas strong correlation between APQ and readers (R 5 0.94, p , 0.0001 for NCP volumes; R 5 0.88, p ,

0.0001, for CP volumes; R 5 0.90, p , 0.0001 for NCP and CP composition).CONCLUSIONS: We developed a fast automated algorithm for quantification of NCP and CP from

CCTA, which is in close agreement with expert manual quantification.� 2009 Society of Cardiovascular Computed Tomography. All rights reserved.

port no conflicts of interest.

an Heart Association Grant-in-Aid

om the Lincy Foundation, and in

ndation, Los Angeles, CA.

d for publication September 16,

Society of Cardiovascular Computed

Introduction

Each year, .1 million people in the United States and.19 million people worldwide experience a sudden acutecoronary event (sudden cardiac death or myocardial infarc-tion).1 Early detection and accurate assessment of coronaryartery disease is crucial in the identification of patients at

Tomography. All rights reserved.

Table 1 Patient characteristics

Values

No. of patients 24Women/men, n/n 5/19Age, y, mean 6 SD (range) 64.5 6 10.1 (47–83)BMI, kg/m2, mean 6 SD (range) 26.5 6 3.6 (21–31)Hypertension 15/24Diabetes mellitus 2/24Current smoker 7/24Coronary calcium score,

mean 6 SD (range)478 6 519 (0–1792.0)

Symptomatic (chest pain,shortness of breath, or both)

14/24

BMI, body mass index.

Dey et al Automated plaque quantification from contrast CT 373

risk of these highly common yet usually preventable coro-nary events. Although the current ‘‘gold standard’’ for as-sessing coronary artery disease is the identification ofanatomically significant coronary luminal stenosis by inva-sive coronary angiography, it is known that most acute cor-onary syndromes arise from plaques that are not criticallyocclusive.2,3 Histopathologic analyses have shown that the‘‘vulnerable’’ plaques considered responsible for acutecoronary events have a large lipid pool, a thin cap, andmacrophage-dense inflammation on or beneath its surface,4

suggesting that plaque composition and structure may beassociated with future coronary events.5,6

Coronary CT angiography (CCTA) with the use of 64-slice CT scanners has recently become an increasinglyeffective clinical tool for noninvasive assessment of thecoronary arteries.7,8 CCTA has also shown substantial po-tential for in vivo plaque component characterization.9–12

Plaque attenuation thresholds have been shown to varysignificantly with intracoronary lumen attenuation13 andchoice of reconstruction kernels14 and are therefore patientand scan specific. Accurate and reproducible measurementof coronary plaque has been limited by the need to manu-ally trace contours separating epicardial fat from the vesselwall and enclosing plaque components, which is time con-suming and can be prone to intraobserver variability.15,16

Current commercial implementations for plaque character-ization require interactive modification of attenuationthresholds for each plaque and are also operator dependent.Although the quantification of calcified plaque (CP) bynoncontrast CT is standardized,17,18 no validated approachto quantify noncalcified plaque (NCP) or CP with CCTAis currently available.19 Toward this end, our objective inthis study was to develop automated computer software(APQ) for accurate, 3-dimensional (3D) plaque segmenta-tion and quantification of noncalcified and calcified coro-nary plaques. APQ combines derivation of scan-specificattenuation thresholds for lumen, NCP, CP, and epicardialfat from the individual CCTA scan; knowledge-based seg-mentation of coronary arteries and geometrical coronary ar-tery modeling, and connected voxel grouping in 3dimensions to define NCP and CP components. In this fea-sibility study, we describe the APQ algorithm and evaluateits initial performance compared with manual delineationof plaque volumes by 2 expert readers.

Methods

Patients and imaging protocol

From the database of CCTA studies acquired at ourinstitution, we retrospectively identified 24 consecutivepatients who underwent CCTA, with focal plaques in theproximal and mid segments of the coronary arteries, withminimum proximal-to-distal plaque length of 2 mm; a totalof 29 such focal plaques were identified. A summary of

patient characteristics is shown in Table 1. The patients un-derwent CCTA for clinical reasons on a Siemens DefinitionDual-Source 64-slice CT scanner (Siemens Medical Solu-tions, Forcheim, Germany) with gantry rotation time of330 milliseconds and detector collimation of 0.6 mm.The imaging protocol has previously been described indetail.20 Intravenous contrast (80 mL) was administeredduring each scan. Raw data were reconstructed from65%–80% of the cardiac cycle in 5% increments, using0.6-mm slice thickness, 0.3-mm slice increment, single-segment reconstruction, a medium-smooth reconstructionkernel (B26f), 512 ! 512 matrix size, and voxel size of0.4 ! 0.4 ! 0.3 mm. Reconstructed CCTA images wereclinically assessed on a Siemens Leonardo workstation aspreviously described.20 Image quality for all the scanswas excellent to good (5 and 4 on our 1–5 image qualityscale).20 The best phase of the cardiac cycle for visualiza-tion of coronary arteries was determined by an expertreader at the time of clinical assessment, and images instandard digital imaging and communications in medicine(DICOM) format corresponding to this phase were trans-ferred to a research workstation for plaque quantification.In addition, as part of our standard protocol, all patients un-derwent a noncontrast CT scan for coronary calcium scor-ing,20 and the Agatston coronary calcium score wasquantified on a SciImage workstation.17 The study was con-ducted according to the guidelines of the Cedars-SinaiMedical Center Institutional Review Board, and all patientsgave written informed consent for retrospective use of theirdata.

Plaque quantification algorithm

Previous studies have described contrast enhancement ofatherosclerotic plaque with the use of human coronaryartery specimens.13,21 By using multislice CT, intravascularultrasound (IVUS) scanning, and histology, Halliburton etal21 have demonstrated that intra-arterial injection of iodin-ated contrast agent results in enhancement of the arteriallumen as well as significant enhancement of atherosclerotic

374 Journal of Cardiovascular Computed Tomography, Vol 3, No 6, November/December 2009

plaque. Cademartiri et al13 have scanned ex vivo specimensof human coronary arteries injected with 4 different dilu-tions of contrast material with multislice CT. Attenuationin lumen and plaque was significantly different for each di-lution of contrast material. The investigators also concludedthat plaque attenuation varies with intracoronary attenua-tion and recommended that luminal attenuation be takeninto account for plaque characterization.13 Cademartiri etal14 have also demonstrated that plaque attenuation variessignificantly with the choice of reconstruction kernel. Inour approach, therefore, scan-specific attenuation thresh-olds for NCP and CP components are automatically deter-mined from luminal attenuation. A flowchart with anoverview of our algorithm is shown in Figure 1. The keysteps in our algorithm were initially tested visually andwere refined on a separate training group of 35 CCTA data-sets acquired previously as described in Dey et al20 andCheng et al.22 In addition to the CCTA data, the algorithmrequires as input a region of interest in the aorta definingthe ‘‘normal blood pool,’’ 5–7 control points in the arteriallumen along the plaque, from which luminal centerlines arederived by fitting a piecewise cubic Catmull-Rom splinefunction,23 and 2 points marking the start and the end ofthe artery to be analyzed.

To avoid possible bias caused by incorrect coronarylumen segmentation, we define a vessel neighborhood, withmaximum radius R from the luminal centerlines. Becausethe coronary artery diameter is typically %3 mm,24 weused a slightly higher value of 4 mm to make sure thewhole vessel is considered. APQ quantifies plaque in 3 di-mensions; however, 2-dimensionl (2D) coronary cross sec-tions of the coronary arteries are considered in several keysteps. By using geometry, we define a 3D–2D correspon-dence for the vessel neighborhood as described below.Given p! and q! to be successive points along a centerline,

Figure 1 Flowchart for the APQ algorithm.

an interpolated point r! is computed as the convexcombination:

r!5g p!1ð12gÞ q!; 0%g%1 ð1Þ

A 2D cross-sectional arterial slice, perpendicular to thedirection of vector ð p!2 q!Þ, is then created with r! as thecenter, and basis vectors ba and bb that span the slice. Everypoint s! on the slice can now be represented by polar rep-resentation ðr; qÞ:

s! 5 r!1rcosðqÞ,ba 1rsinðqÞ,bb; 0%r%R; 0%q%2p

ð2Þ

where R is the maximum vessel neighborhood radius.To calculate scan-specific epicardial fat (EF) thresholds

(EFTs), we search radially inward, toward the luminalcenterline, in the 2D cross-sectional vessel neighborhoodslices. For the preset upper EFT, a threshold of 210Hounsfield units (HUs) is used, only as an initial value; thisvalue was obtained by manual measurement from theCCTA datasets in the training group. From the meanattenuation (m) and standard deviation (SD) of EF area,the corrected upper EFT is then redefined as:

EFT 5 m13SD ð3Þ

This relationship gave the best agreement with manualmeasurement from our training group and was subsequentlyused in this study.

To calculate scan-specific thresholds for NCP and CP, weused algorithms developed in our prior work with non-contrast cardiac CT25 and also considered the scan-specificvariation in contrast distribution through the coronary arter-ies. The uniform normal blood pool region in the aorta is firstfiltered with a median filter, and an image histogram is com-puted. A Gaussian curve is fitted to the smoothed image his-togram by using the iterative Levenberg-Marquardtalgorithm.25,26 From the fitted Gaussian curve, we calcu-lated the peak value corresponding to normal contrast inthe aorta, as well as the SD s. The lower and upper contrastlevels at the level of the aorta are defined as27:

l5p23s

u5p13sð4Þ

where p corresponds to the Gaussian peak value.For each cross-sectional arterial plane, the normal

contrast is calculated by 2D region growing, from theinterpolated luminal centerline point defined by equation 1,above a initial lower contrast threshold (equal to the mini-mum normal blood pool value) and below u. For eachcross-sectional plane in each artery, the contrast distributionfactor (CDF) is defined as:

CDF 5 Ci=Ci21 ð5Þ

where Ci is the average normal arterial contrast computedin cross-sectional plane i, and i–1 corresponds to the previ-ous, more proximal cross-sectional plane. Finally, the

Figure 2 CCTA mixed plaque (A) and attenuation (HU) line profile through mid-plaque cross-sectional plane, indicated with greenarrows (A and B). EF, NCP, and lumen regions in the attenuation profile are marked (B). For this study, the APQ upper NCP thresholdwas at 180 HU, and the EF threshold was at 220 HU, indicated with blue lines (B). The arterial wall HU is shown with green point(B), found as described in ‘‘Methods,’’ the corresponding attenuation value is at 210 HU.

Table 2 Attenuation threshold values computed by APQ

Values

No. of plaques 29EF upper threshold, HU,

mean 6 SD (range)216 6 13 (23 to 241)

NCP upper threshold, HU,mean 6 SD (range)

227 6 40 (145–335)

Mid-lesion normal contrast, HU,mean 6 SD (range)

419 6 78 (333–618)

CP lower threshold, HU,mean 6 SD (range)

511 6 80 (397–689)

APQ, automated computer algorithm; CP, calcified plaque; EF,

epicardial fat; HU, Hounsfield unit; NCP noncalcified plaque.

Dey et al Automated plaque quantification from contrast CT 375

corrected lower contrast threshold (upper threshold forNCP, denoted by nt) is defined by multiplying l by theCDF corresponding to the middle cross-section of the pla-que. In the case of total occlusion, CDF for the most distalplaque cross section is used. The lower threshold for CP isu as defined in equation 4. We tested upper and lower con-trast levels (equation 4), the upper threshold for NCP andthe lower threshold for CP by varying the constant multi-plying the standard deviation as well as by using the full-width tenth maximum; we found that 3s yielded visuallythe best results in our training group, and this assumptionwas subsequently used in this study.

Although coronary plaques are easily visualized byCCTA, it is known that visualization of the coronary arterialwall, with thickness typically ,1 mm, is challengingbecause of partial volume effects.5 To segment the arterialwall, we use a multistep adaptive algorithm. The vesselneighborhood is first refined by 2D region growing fromthe interpolated luminal centerline points, using EFT as thelower threshold; this neighborhood still includes both the ar-terial wall and volume-averaged wall and EF voxels. Next,we searched radially for the maximum attenuation gradientin the 2D cross-sectional planes, in the vessel neighborhood,in connected voxels with attenuation values in the lowerthird of the NCP attenuation range (EFT, nt/3). To maintainconsistency and accuracy, we constrained the search to startfrom an arterial radius equal to 85% of the previous, moreproximal cross-sectional plane, as suggested by data frompreviously reported histologic observations.28 Because weexpected contrast enhancement in the plaque,13,21 the outermaximal radial gradient boundary found in each cross-sec-tional plane defined the arterial wall boundary. The averageattenuation value over the entire plaque corresponding to thearterial wall (aw) was computed for the next step (Fig. 2).

NCP and CP components are automatically classifiedwithin the arterial wall by an iterative, recursive 3D region-growing algorithm, which expands from a starting seed

voxel similar to an expanding 3D balloon. The seed voxel isfirst localized from the cross-sectional plane with thegreatest luminal stenosis, using attenuation values rangingfrom (aw, nt) for NCP and using a lower attenuation thresh-old of u for CP. If the seed voxel for either plaque compo-nent is not ‘‘found,’’ the algorithm iteratively searches forthe nearest unclassified seed voxel within the attenuationrange. We then performed 3D region growing recursivelyuntil all connected voxels within the corresponding attenu-ation range were identified, weighted by the constraint,tested with our training group, that the attenuation gradientis required to be %50% in each plaque component. Theclassified voxels are displayed with color-coded overlayand NCP and CP volume, and plaque composition (percent-age of NCP, percentage of CP) was calculated.

Manual plaque analysis

Each plaque was independently manually quantified by 2experienced readers (V.Y.C., R.N.), blinded to each other,using Vitrea workstation (Vital Images, Minnetonka, MN)

376 Journal of Cardiovascular Computed Tomography, Vol 3, No 6, November/December 2009

version 4.1.1. The manual 3D plaque outlining method wassimilar to previously reported studies.16,22 Standard CTAwindow width and level settings (width, 800–900 HU; level,250–300 HU) were initially used, and the reader could mod-ify these settings as needed, particularly when quantifyingCP lesions. After identification of the plaque in standardtransverse images, serial oblique multiplanar reformatted(MPR) images orthogonal to the longitudinal axis of the

Figure 3 Results of APQ quantification for a mid-LAD (left anteriorreferred for CCTA, with a coronary calcium score of 761. The maximumMPR views showing a curved mixed plaque from the mid-LAD coronaAPQ quantification, with yellow overlay corresponding to CP and red ovlay can also be visualized as NCP (red) and CP (yellow) contours (C).volume 75.8 mm3 and 14.4 mm3, respectively, with 88% stenosis. MeaA view is shown from invasive coronary angiography (D), with arrow iquantitative coronary angiography.

involved vessel segment and the plaque were displayed.The 3D step increment (typically between 0.3 and 0.45mm) for each cross-sectional segment along the longitudinalaxis was calculated by marking the starting and ending posi-tions of the plaque, computing the 3D vector from the start tothe end, and dividing the 3D vector by the number of cross-sectional plaque segments. Plaque areas were then manuallytraced in each cross-sectional segment, and the total NCP and

descending coronary artery) plaque for a 51-year-old male smokerheart rate during CCTA was 80 beats/min (image quality good).

ry artery are shown (A). Corresponding color-coded overlay fromerlay corresponding to NCP, is shown (B). Alternately, APQ over-For this lesion, APQ calculated results were NCP volume and CPn (6SD) value within the NCP component was 81.4 6 58.3 HU.ndicating the mid-LAD lesion. Percentage of stenosis was 67% by

Dey et al Automated plaque quantification from contrast CT 377

CP volumes were calculated by multiplying the correspond-ing total plaque areas with the 3D step increment. In total, 798and 849 2D cross-sectional segments were manually tracedby the 2 observers.

Plaque quantification by interactive thresholdadjustment

In addition, each plaque was quantified with plaqueanalysis software (SUREPlaque; Vital Images) by an expertreader (V.Y.C.), by visually adjusting attenuation thresholdsfor each plaque (ITA method). Curved multiplanar refor-matted (CMPR) images were rendered and displayed byusing the software (in 3 cases, additional manual operation,editing of the centerline to obtain appropriate CMPRdisplay, was necessary). The reader marked starting andending positions of the plaque in the CMPR display andadjusted attenuation thresholds for NCP and CP to matchvisual impression of plaque components, to obtain NCP,CP, and total plaque volumes.

Figure 4 Results of APQ quantification for a curved proximal RCA (patient referred for CCTA, with a coronary calcium score of 979. The mexcellent). MPR views showing a curved mixed plaque from the mid-LAlay from APQ quantification, with yellow overlay corresponding to CP aresults with color-coded contours (NCP in red, CP in yellow) are shownCP volume 43 mm3 and 14.4 mm3, respectively, with 31% stenosis. M

Interobserver variability for APQ and ITA

We also evaluated the variability between 2 independentobservers for APQ and ITA. For APQ, each observerindependently marked the proximal and distal limits of theplaque and drew the ‘‘normal blood pool’’ region ofinterest. For ITA, each observer independently marked theproximal and distal limits of the plaque and visuallyadjusted the attenuation thresholds.

Statistical analysis

The results of plaque quantification were analyzed byAnalyse-It software (www.analyse-it.com). All continuousvariables were expressed as mean 6 SD. To compareagreements of any 2 plaque quantification methods,Pearson’s correlation coefficient was calculated andBland-Altman plots were created. Paired t test was usedto compare the difference from manual quantification. AP value of ,0.05 was considered statistically significant.

right coronary artery) plaque for a 66-year-old symptomatic maleaximum heart rate during CCTA was 50 beats/min (image qualityD coronary artery are shown (A). Corresponding color-coded over-nd red overlay corresponding to NCP, is shown (B). Quantification(C). For this lesion, APQ calculated results were NCP volume andean (6SD) value within the NCP component was 96 6 59 HU.

378 Journal of Cardiovascular Computed Tomography, Vol 3, No 6, November/December 2009

Results

Study processing times ranged from 5 to 25 minutes formanual quantification, from 1 to 3 minutes for ITA. Theaverage length of the quantified plaques was 10.2 6 4.5mm (range, 2.7–24.8 mm). For manual quantification, thenumber of cross-sectional segments that needed to betraced ranged from 9 to 61 per plaque. The time forautomated plaque segmentation and quantification was ,2seconds for all plaques in this study on a standard Windows2.5 GHz computer.

Strong correlation was observed between the 2 expertreaders for NCP volumes (r 5 0.99, P , 0.0001) and forCP volumes (r 5 0.85, P , 0.0001). The 95% limits ofagreement range between the 2 observers were 225.9 to23.2 mm3 for NCP and 215.5 to 23.4 mm3 for CP;

Figure 5 APQ algorithm performance compared with expert manua(B) are shown. Correlation coefficient for NCP was 0.94 (P , 0.0001)(represented by the equations y 5 1.36x for NCP and y 5 1.005x for CP)(C) and Bland-Altman comparison of CP with expert readers (D) are shthe 95% limits of agreement were 237.3 to 102.5 mm3. For CP, there wawere 214.9 to 17.0 mm3.

with a small positive bias (3.96 mm3) for CP. Mean abso-lute differences between the 2 readers were 8.4 6 9.6mm3 for NCP volumes and 6.2 6 8.7 mm3 for CPvolumes.

The HU values for EF, upper NCP threshold (nt), normalblood pool at the mid plaque lesions and lower CP thresh-olds are shown in Table 2. Examples of 2 lesions quantifiedby APQ are shown in Figures 3 and 4. For the NCP com-ponent, APQ showed a significantly lower mean absolutedifference from expert readers than for ITA (APQ, 26.76 21.2 mm3; ITA, 42.1 6 34.8 mm3; P 5 0.01). For theCP component, APQ also showed a significantly lowermean absolute difference from expert readers than forITA (APQ, 6.0 6 5.5 mm3; ITA, 30.5 6 22.5 mm3; P ,

0.0001). The correlation of NCP and CP volumes byAPQ to the corresponding manually quantified plaque

l quantification. Correlation for NCP (A) and correlation for CP, and for CP it was 0.88 (P , 0.0001). Best-fit lines for the dataare shown. Bland-Altman comparison of NCP with expert readers

own. For NCP, there was a positive bias of 32.6 mm3 for NCP, ands a small positive bias of 1.1 mm3, and the 95% limits of agreement

Figure 6 Figure shows results of ITA plaque quantification compared with expert manual quantification. Correlation for NCP (A) andcorrelation for CP (B) are shown. Correlation coefficient for NCP was 0.90 (P , 0.0001), and for CP it was 0.89 (P , 0.0001). Best-fit lines for the data (represented by the equations y 5 1.53x for NCP and y 5 1.27x for CP) are shown. Bland-Altman comparison ofNCP with expert readers (C) and Bland-Altman comparison of CP with expert readers (D) are shown. For NCP, there was a positivebias of 64.4 mm3 for NCP, and the 95% limits of agreement were 221.6 to 150.4 mm3. For CP, there was a small positive bias of 5.2mm3, and the 95% limits of agreement were 214.5 to 24.9 mm3.

Dey et al Automated plaque quantification from contrast CT 379

components averaged between the 2 experts is shown(Fig. 5A and B). Bland-Altman comparisons of NCP andCP volumes by APQ, with manually quantified results,are shown (Fig. 5C and D). The correlation of NCP andCP volumes by ITA to corresponding manually quantifiedplaque components is shown (Fig. 6A and B). Bland-Altman comparisons of NCP and CP volume by ITA,with manually quantified results, are shown (Fig. 6C andD). For the NCP component, APQ showed a significantlylower bias than ITA (32.6 mm3 versus 64.4 mm3) andtighter 95% limits of agreement (P 5 0.01). Because ofroutine inclusion of vessel wall and EF as NCP, ITAoverestimated NCP in both normal and abnormal plaquecross sections, as previously reported.5 With APQ, NCPoverestimation was primarily found in abnormal cross sec-tions. For CP, APQ also showed a significantly lower biasthan did ITA (1.1 mm3 versus 5.2 mm3) and tighter 95%

limits of agreement (P , 0.0001). Correlation with expertmanual quantification for NCP was higher for APQ thanfor ITA (0.94 versus 0.90 for NCP volume). For plaquecomposition, there was excellent correlation betweenAPQ and expert readers (r 5 0.90, P , 0.0001 for both per-centage of NCP and percentage of CP volume), andbetween ITA and expert readers (r 5 0.95, P , 0.0001).

Visually, APQ was visually assessed to be successful(extending over entire NCP and CP components with nomissing voxels) in all 29 plaques, with some overestimationof NCP because of inclusion of artifactual low HU beam-hardening areas as NCP, and inclusion of EF as NCP,especially for small plaque lesions located at vessel branchpoints. In addition, an overall limitation of the manualplaque quantification technique is that the adding plaquecomponent areas from the serial MPR images is accuratefor plaques with linear configuration; for curved or tortuous

Table 3 Interobserver variability for plaque quantification with APQ and ITA

NCP CP Total plaque

APQAbsolute difference in volume, mm3, mean 6 SD 19.8 6 22.2 1.6 6 3.4* 21.1 6 22.7Correlation between observers (P value) 0.97 (,0.0001)† 0.98 (,0.0001)† 0.96 (,0.0001)†

ITAAbsolute difference in volume, mm3, mean 6 SD 29.6 6 24.9 9.7 6 9.2* 32.0 6 29.5Correlation between observers (P value) 0.91 (,0.0001)† 0.84 (,0.0001)† 0.88 (,0.0001)†

APQ, automated computer algorithm; CP, calcified plaque; HU, Hounsfield unit; ITA, interactive adjustment; NCP noncalcified plaque.

*Pairwise differences were significant (P , 0.0001).†Difference in all 3 correlation coefficients were significant.

380 Journal of Cardiovascular Computed Tomography, Vol 3, No 6, November/December 2009

plaques, this technique approximates true plaque volumes.In this study, 3 of 29 plaques were curved plaques (onecurved proximal RCA plaque is shown in Fig. 4).

Interobserver variability

The interobserver variability results for plaque quantifi-cation with APQ and ITA are summarized in Table 3. Themean absolute difference between 2 observers for NCP, CP,and total plaque volume was lower for APQ than for ITA;this difference was significant for CP (P , 0.0001). Corre-lation between observers was significantly stronger forAPQ than for ITA, for total plaque volume (P 5 0.04),for the NCP component (P 5 0.04), and for the CP compo-nent (P 5 0.0001).

Discussion

Standardized and automated quantification of NCP, CP,and total plaque burden from CCTA, although extremelychallenging, is of great interest for refinement of cardio-vascular risk stratification.5,19 Toward this end, we devel-oped novel automated software for segmentation andquantitative characterization of plaques imaged by CCTA.The novelty of the APQ algorithm is in the combinationof several image-processing components that allow deriva-tion of near-automatic segmentation of NCP and CP. Thesecomponents are (1) derivation of scan-specific attenuationthresholds for lumen, NCP, CP, and EF, from the CCTAscan; (2) knowledge-based segmentation of coronaryarteries and geometrical coronary artery modeling; and(3) connected voxel grouping in 3 dimensions to defineNCP and CP components. This is an initial feasibility studyof this approach, and we have shown that this combinedautomated algorithm agrees very well with expert manualplaque quantification and is highly reproducible, and theresults can be obtained in a fraction of the time (,2%)needed for manual analysis.

Other investigators have recently developed methods toquantify noncalcified plaque imaged by CCTA.29 Clouseet al29 have described a ‘‘voxel analysis’’ technique that

uses Analyze-Direct software (www.analyzedirect.com;Mayo Clinic, Rochester, MN), whereby expert readersmanually drew 8 perpendicular line profiles through theplaque and measured attenuation values in 8 radial voxelsfor each line profile. By using these manually definedpoints, the attenuation thresholds for lumen, EF, and vesselwall were calculated from interpolation of the line profiles,and the lumen and plaque volume were calculated. Gertz etal30 have used 3D isotropic wavelet analysis to characterizemicro-CT images of excised human coronary arteries.Compared with histology, they found that wavelet analysisallowed identification of coronary plaque components with81% sensitivity and 86% specificity.30 However, it has notyet been applied to clinical CCTA images from standardmultislice scanners. To our knowledge, this is the firstreport describing a complete algorithm for automatedCCTA plaque segmentation and quantification, a key steptoward standardized quantification of NCP and CP.

Recent in vivo studies comparing manual plaque charac-terization from 64-slice CCTA to invasive intravascularstandard have found significant overlap between lipid-richand fibrous NCP12 and high intraobserver variability.9,12 Pet-ranovic et al12 reported mean attenuation values of 275.3 6

77.2 HU for coronary lumen, 117.9 6 94.2 HU for NCP,and 608.2 6 216.9 HU for CP, with the mean attenuationvalues for NCP and CP significantly differing from eachother, suggesting that accurate NCP and CP quantificationcan be performed from CCTA. These data also show thatthe lower threshold value of 130 HU, used extensively forCP quantification from noncontrast CT,17,18 cannot be ap-plied for CP quantification from CCTAwithout including sig-nificant NCP.12 Our approach, therefore, aims to standardizeNCP and CP quantification by deriving scan-specific attenu-ation thresholds for both components.

Study limitations

Our study lacks validation with an invasive standard,such as IVUS scanning, or with histology. This study was aretrospective analysis; our patients did not undergo IVUSscans at our institution and IVUS data were not available.Histology results are very difficult to obtain in human

Dey et al Automated plaque quantification from contrast CT 381

beings for the coronary arteries, and the software isdesigned for detection of plaque in vivo human coronaryartery images. The validation of the software in animalmodels may also not be appropriate because assumptionsare made which are specific to human scans. The otherpossible external validation, invasive coronary angiographydoes not allow estimation of coronary plaque and thereforeis of limited value in this comparison. Nevertheless, theprimary focus of this feasibility study was to compare thenew automated approach with manual techniques forquantification of NCP and CP, currently the only availablestandard for CCTA; further validation is needed withIVUS-correlated CCTA data.

We restricted our sample to selected coronary plaques inthe proximal and mid segments of the coronary arteries(limiting our analysis to arterial diameters . 1.5 mm),which, because of limitations of CCTA spatial resolution,can be considered standard CCTA practice. The keyparameters for APQ were visually tested on our traininggroup of CCTA data before comparison with expert manualquantification; however, some parameters may need to befurther adjusted for new CCTA datasets with changing datacharacteristics. In our experience, the upper and lowerthreshold determinations for NCP and CP as described wascrucial, but constraints of HU gradient may need furtheradjustment for new CCTA datasets. Compared with expertmanual readers, our results show overestimation of NCP byAPQ; further shape modeling of the coronary artery may benecessary to correct this. For one patient, artifactual low-HU area because of metal beam-hardening was included asNCP. This is a limitation of attenuation-based plaquecharacterization; further pattern recognition around high-HU areas may help flag beam-hardening artifacts. Anotherlimitation is that low-density CP nodules whose attenuationvalues overlap with that of lumen cannot be easily detectedautomatically; further pattern recognition may be helpful incharacterizing such lesions.

Conclusions

We developed a novel automated algorithm for quantita-tive 3D CCTA plaque characterization which shows strongagreement and low bias for NCP and CP volumes with expertmanual quantification, is highly reproducible, and requires afraction of the time needed for manual analysis.

References

1. Yusuf S, Reddy S, Ounpuu S, Anand S: Global burden of cardiovascu-

lar diseases: part I: general considerations, the epidemiologic transi-

tion, risk factors, and impact of urbanization. Circulation. 2001;104:

2746–53.

2. Falk E, Fuster V: Angina pectoris and disease progression. Circulation.

1995;92:2033–5.

3. Virmani R, Burke AP, Farb A, Kolodgie FD: Pathology of the vulner-

able plaque. J Am Coll Cardiol. 2006;47(8 Suppl):C13–8.

4. Virmani R, Kolodgie FD, Burke AP, Farb A, Schwartz SM: Lessons

from sudden coronary death: a comprehensive morphological classifi-

cation scheme for atherosclerotic lesions. Arterioscler Thromb Vasc

Biol. 2000;20:1262–75.

5. Akram K, Rinehart S, Voros S: Coronary arterial atherosclerotic pla-

que imaging by contrast-enhanced computed tomography: Fantasy

or reality? J Nucl Cardiol. 2008;15:818–29.

6. Narula J, Finn AV, Demaria AN: Picking plaques that pop. J Am Coll

Cardiol. 2005;45:1970–3.

7. Achenbach S: Cardiac CT: state of the art for the detection of coronary

arterial stenosis. J Cardiovasc Comput Tomogr. 2007;1:3–20.

8. Berman DS, Shaw LJ, Hachamovitch R, Friedman JD, DM Polk,

Hayes SW, Thomson LE, Germano G, Wong ND, Kang X, Rozanski A:

Comparative use of radionuclide stress testing, coronary artery calcium

scanning, and noninvasive coronary angiography for diagnostic and prog-

nostic cardiac assessment. Semin Nucl Med. 2007;37:2–16.

9. Leber AW, Becker A, Knez A, von Ziegler F, Sirol M, Nikolaou K,

Ohnesorge B, Fayad ZA, Becker CR, Riser M, Steinbeck G,

Boekstegers P: Accuracy of 64-slice computed tomography to classify

and quantify plaque volumes in the proximal coronary system: a com-

parative study using intravascular ultrasound. J Am Coll Cardiol.

2006;47:672–7.

10. Leber AW, Knez A, Becker A, Becker C, von Ziegler F, Nikolaou K,

Rist C, Reiser M, White C, Steinbeck G, Boekstegers P: Accuracy of

multidetector spiral computed tomography in identifying and differen-

tiating the composition of coronary atherosclerotic plaques: a compar-

ative study with intracoronary ultrasound. J Am Coll Cardiol. 2004;43:

1241–7.

11. Leber AW, Knez A, von Ziegler F, Becker A, Nikolaou K, Paul S,

Wintersperger B, Reiser M, Becker CR, Steinbeck G, Boekstegers P:

Quantification of obstructive and nonobstructive coronary lesions by

64-slice computed tomography: a comparative study with quantitative

coronary angiography and intravascular ultrasound. J Am Coll Cardiol.

2005;46:147–54.

12. Petranovic M, Soni A, Bezzera H, Loureiro R, Sarwar A, Raffel C,

Pomerantsev E, Jang IK, Brady TJ, Achenbach S, Cury RC: Assess-

ment of nonstenotic coronary lesions by 64-slice multidetector com-

puted tomography in comparison to intravascular ultrasound:

evaluation of nonculprit coronary lesions. J Cardiovasc Comput

Tomogr. 2009;3:24–31.

13. Cademartiri F, Mollet NR, Runza G, Bruining N, Hamers R, Somers P,

Knaapen M, Verheye S, Midiri M, Krestin GP, de Feyter PJ: Influence

of intracoronary attenuation on coronary plaque measurements using

multislice computed tomography: observations in an ex vivo model

of coronary computed tomography angiography. Eur Radiol. 2005;

15:1426–31.

14. Cademartiri F, La Grutta L, Runza G, Palumbo A, Maffei E,

Mollet NR, Bartolotta TV, Somers P, Knaapen M, Verheye S,

Midiri M, Hamers R: B ruining N: Influence of convolution filtering

on coronary plaque attenuation values: observations in an ex vivo

model of multislice computed tomography coronary angiography.

Eur Radiol. 2007;17:1842–9.

15. Burgstahler C, Reimann A, Beck T, Kuettner A, Baumann D,

Heuschmid M, Brodoefel H, Claussen CD, Kopp AF, Schroeder S: Influ-

ence of a lipid-lowering therapy on calcified and noncalcified coronary

plaques monitored by multislice detector computed tomography: results

of the New Age II Pilot Study. Invest Radiol. 2007;42:189–95.

16. Schmid M, Achenbach S, Ropers D, Komatsu S, Ropers U, Daniel WG,

Pflederer T: Assessment of changes in non-calcified atherosclerotic

plaque volume in the left main and left anterior descending coronary

arteries over time by 64-slice computed tomography. Am J Cardiol.

2008;101:579–84.

17. Agatston AS, Janowitz WR, Hildner FJ, Zusmer NR, Viamonte M Jr.,

Detrano R: Quantification of coronary artery calcium using ultrafast

computed tomography. J Am Coll Cardiol. 1990;15:827–32.

18. Callister T, Cooil B, Raya S, Lippolis N, Russo D, Raggi P: Cor-

onary artery disease: improved reproducibility of calcium scoring

382 Journal of Cardiovascular Computed Tomography, Vol 3, No 6, November/December 2009

with an electron-beam CT volumetric method. Radiology. 1998;

208:807–14.

19. Schuijf JD, Bax JJ: How do you quantify noncalcified plaque? J Car-

diovasc Comput Tomogr. 2008;2:360–5.

20. Dey D, Lee CJ, Ohba M, Gutstein A, Slomka PJ, Cheng V, Suzuki Y,

Suzuki S, Wolak A, Le Meunier L, Thomson LEJ, Cohen I,

Friedman JF, Germano G, Berman DS: Image quality and artifacts

in coronary CT angiography with dual-source CT: initial clinical expe-

rience. J Cardiovasc Comput Tomogr. 2008;2:105–14.

21. Halliburton SS, Schoenhagen P, Nair A, Stillman A, Lieber M, Murat

Tuzcu E, Geoffrey Vince D, White RD: Contrast enhancement of

coronary atherosclerotic plaque: a high-resolution, multidetector-row

computed tomography study of pressure-perfused, human ex-vivo

coronary arteries. Coron Artery Dis. 2006;17:553–60.

22. Cheng V, Nakazato R, Dey D, Gurudevan S, Tabak J, Budoff MJ, Karls-

berg RP, Min J, Berman DS: Reproducibility of coronary artery plaque

volume and composition quantification by 64-detector row coronary

computed tomographic angiography: an intraobserver, interobserver,

and interscan variability study. J Cardiovasc Comput Tomogr. Published

ahead of print on July 31, 2009, as DOI: 10.1016/j.jcct.2009.07.001.

23. Catmull E, Rom R: A class of local interpolating splines. Proc Int

Conf Comput Aided Geom Des. 1974;74:317–26.

24. Achenbach S: Computed tomography coronary angiography. J Am

Coll Cardiol. 2006;48:1919–28.

25. Dey D, Callister TQ, Slomka PJ, Aboul-Enein F, Nishina H, Kang X,

Gransar H, Wong ND, Miranda-Peats R, Hayes S, Friedman JD,

Berman DS: Computer-aided detection and evaluation of lipid-rich pla-

que in non-contrast cardiac computed tomography. AJR Am J Roent-

genol. 2006;186(6 Suppl. 2):S407–13.

26. Numerical recipes in C the art of scientific computing [computer pro-

gram]. Version 2nd, Acrobat. Cambridge, NY: Cambridge University

Press; 2000.

27. Raggi P, Callister TQ, Cooil B: Calcium scoring of the coronary artery

by electron beam CT: how to apply an individual attenuation thresh-

old. AJR Am J Roentgenol. 2002;178:497–502.

28. Taylor AJ, Burke AP, Farb A, Yousefi P, Malcom GT, Smialek J,

Virmani R: Arterial remodeling in the left coronary system: the role

of high-density lipoprotein cholesterol. J Am Coll Cardiol. 1999;34:

760–7.

29. Clouse ME, Sabir A, Yam C-S, Yoshimura N, Lin S, Welty F,

Martinez-Clark P, Raptopoulos V: Measuring noncalcified coronary

atherosclerotic plaque using voxel analysis with MDCT angiogra-

phy: a pilot clinical study. AJR Am J Roentgenol. 2008;190:

1553–60.

30. Gertz SD, Bodmann BG, Vela D, Papadakis M, Aboshady I,

Cherukuri P, Alexander S, Kouri DJ, Baid S, Gittens AA,

Gladish GW, Conyers JL, Cody DD, Gavish L, Mazraeshahi RM,

Wilner WT, Frazier L, Madjid M, Zarrabi A, Lukovenkov S,

Ahmed A, Willerson JT, Casscells SW: Three-dimensional isotropic

wavelets for postacquisitional extraction of latent images of athero-

sclerotic plaque components from micro-computed tomography of

human coronary arteries. Acad Radiol. 2007;14:1509–19.