Automatic quantification of epicardial fat volume on non-enhanced cardiac CT scansusing a multi-atlas segmentation approachRahil Shahzad, Daniel Bos, Coert Metz, Alexia Rossi, Hortense Kirili, Aad van der Lugt, Stefan Klein, Jacqueline

Witteman, Pim de Feyter, Wiro Niessen, Lucas van Vliet, and Theo van Walsum

Citation: Medical Physics 40, 091910 (2013); doi: 10.1118/1.4817577 View online: http://dx.doi.org/10.1118/1.4817577 View Table of Contents: http://scitation.aip.org/content/aapm/journal/medphys/40/9?ver=pdfcov Published by the American Association of Physicists in Medicine Articles you may be interested in Assessment of dedicated low-dose cardiac micro-CT reconstruction algorithms using the left ventricular volumeof small rodents as a performance measure Med. Phys. 41, 051908 (2014); 10.1118/1.4870983 Lumen segmentation and stenosis quantification of atherosclerotic carotid arteries in CTA utilizing a centerlineintensity prior Med. Phys. 40, 051721 (2013); 10.1118/1.4802751 Evaluation of a multi-atlas based method for segmentation of cardiac CTA data: a large-scale, multicenter, andmultivendor study Med. Phys. 37, 6279 (2010); 10.1118/1.3512795 Automatic segmentation of intracranial arteries and veins in four-dimensional cerebral CT perfusion scans Med. Phys. 37, 2956 (2010); 10.1118/1.3397813 Geometric feature-based multimodal image registration of contrast-enhanced cardiac CT with gated myocardialperfusion SPECT Med. Phys. 36, 5467 (2009); 10.1118/1.3253301

Automatic quantification of epicardial fat volume on non-enhanced cardiacCT scans using a multi-atlas segmentation approach

Rahil Shahzada)

Quantitative Imaging Group, Faculty of Applied Sciences, Delft University of Technology, 2628 CJ Delft, TheNetherlands and Biomedical Imaging Group Rotterdam, Departments of Radiology and Medical Informatics,Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Daniel BosDepartment of Radiology, Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands and Department ofEpidemiology, Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Coert MetzBiomedical Imaging Group Rotterdam, Departments of Radiology and Medical Informatics, Erasmus MedicalCenter, 3015 GE Rotterdam, The Netherlands

Alexia RossiDepartment of Radiology, Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Hortense KirisliBiomedical Imaging Group Rotterdam, Departments of Radiology and Medical Informatics, Erasmus MedicalCenter, 3015 GE Rotterdam, The Netherlands

Aad van der LugtDepartment of Radiology, Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Stefan KleinBiomedical Imaging Group Rotterdam, Departments of Radiology and Medical Informatics, Erasmus MedicalCenter, 3015 GE Rotterdam, The Netherlands

Jacqueline WittemanDepartment of Epidemiology, Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Pim de FeyterDepartment of Radiology, Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Wiro NiessenQuantitative Imaging Group, Faculty of Applied Sciences, Delft University of Technology, 2628 CJ Delft, TheNetherlands and Biomedical Imaging Group Rotterdam, Departments of Radiology and Medical Informatics,Erasmus Medical Center, 3015 GE Rotterdam, The Netherlands

Lucas van VlietQuantitative Imaging Group, Faculty of Applied Sciences, Delft University of Technology, 2628 CJ Delft, TheNetherlands

Theo van WalsumBiomedical Imaging Group Rotterdam, Departments of Radiology and Medical Informatics, Erasmus MedicalCenter, 3015 GE Rotterdam, The Netherlands

(Received 2 May 2013; revised 2 July 2013; accepted for publication 23 July 2013; published 12August 2013)

Purpose: There is increasing evidence that epicardial fat (i.e., adipose tissue contained within thepericardium) plays an important role in the development of cardiovascular disease. Obtaining the epi-cardial fat volume from routinely performed non-enhanced cardiac CT scans is therefore of clinicalinterest. The purpose of this work is to investigate the feasibility of automatic pericardium segmenta-tion and subsequent quantification of epicardial fat on non-enhanced cardiac CT scans.Methods: Imaging data of 98 randomly selected subjects belonging to a larger cohort of subjects whounderwent a cardiac CT scan at our medical center were retrieved. The data were acquired on twodifferent scanners. Automatic multi-atlas based method for segmenting the pericardium and calculat-ing the epicardial fat volume has been developed. The performance of the method was assessed by(1) comparing the automatically segmented pericardium to a manually annotated reference standard,(2) comparing the automatically obtained epicardial fat volumes to those obtained manually, and(3) comparing the accuracy of the automatic results to the inter-observer variability.

091910-1 Med. Phys. 40 (9), September 2013 © 2013 Am. Assoc. Phys. Med. 091910-10094-2405/2013/40(9)/091910/9/$30.00

091910-2 Shahzad et al.: Automatic epicardial fat volume quantification 091910-2

Results: Automatic segmentation of the pericardium was achieved with a Dice similarity index of89.1 ± 2.6% with respect to Observer 1 and 89.2 ± 1.9% with respect to Observer 2. The correlationbetween the automatic method and the manual observers with respect to the epicardial fat volumecomputed as the Pearson’s correlation coefficient (R) was 0.91 (P < 0.001) for both observers. Theinter-observer study resulted in a Dice similarity index of 89.0 ± 2.4% for segmenting the peri-cardium and a Pearson’s correlation coefficient of 0.92 (P < 0.001) for computation of the epicardialfat volume.Conclusions: The authors developed a fully automatic method that is capable of segmenting the peri-cardium and quantifying epicardial fat on non-enhanced cardiac CT scans. The authors demonstratedthe feasibility of using this method to replace manual annotations by showing that the automaticmethod performs as good as manual annotation on a large dataset. © 2013 American Association ofPhysicists in Medicine. [http://dx.doi.org/10.1118/1.4817577]

Key words: registration, segmentation, pericardium delineation, adipose tissue, computed tomogra-phy (CT)

1. INTRODUCTION

Cardiovascular disease (CVD) is one of the leading causesof death worldwide.1 Epicardial fat is the adipose tissuewhich is found between the myocardium and the viscerallayer of the pericardium, and thus directly surrounds the en-tire heart as well as the coronary arteries. Increasing evi-dence implicates epicardial fat in the etiology of CVD.2–4 Itis thought that through local production of inflammatory fac-tors it may directly contribute to the formation of coronaryatherosclerosis.5–7 Few studies have found that epicardial fatis associated with cardiovascular risk factors.8, 9 Other studieshave shown that epicardial fat is a dominant factor in case ofcoronary artery disease.10–12 A few population based studieshave also been performed. Ding et al.13 investigated whetherepicardial fat is an independent predictor of future heart dis-ease events as compared to conventional risk factors on 998individuals from the MESA study. Mahabadi et al.14 quanti-fied epicardial fat volume on 4093 subjects in order to deter-mine if epicardial fat predicts coronary events in the generalpopulation.

Several methods for epicardial fat quantification have re-cently been developed. Most of these methods are completelymanual, which is a tedious procedure to perform. The man-ual methods are also prone to inter and intra-observer vari-ability. The objective of our study is to develop and evaluatea fully automatic method, which can accurately and robustlysegment the pericardium and quantify the amount of adiposetissue contained within. To the best of our knowledge, thisis the first fully automatic method presented in the literature.Hence, it has the potential to be applied to large scale clinicalor population based studies.

Most of the methods described previously require manualdelineation of the pericardium, which is subsequently used toquantify the volume of fat. Taguchi et al.15 traced the epi-cardial, subcutaneous, and visceral fat. Wheeler et al.16 usedlandmark points to initialize the heart segmentation. Rositoet al.8 traced the pericardium to delineate the heart from thesurrounding structures. Ding et al.17 used a few landmarkpoints around the heart to enclose it in an envelope, similarto the method proposed by Wheeler et al.16

More recently, a semi-automatic method was proposed byDey et al.18, 19 Their method needs two interactions. First, theuser needs to select the top and the bottom slice in betweenwhich the heart is contained. Once this is done, the methoduses region growing and anatomical information to segmentthe heart. Second, the user needs to select five to seven con-trol points on the axial slices to pinpoint the location of thepericardium. The degree of interaction of this method is stillsubstantial. Population and clinical studies, as well as clinicalworkflow, would greatly benefit from a precise and fully au-tomatic method for epicardial fat quantification. Ultimately,findings from such studies could further establish the role ofepicardial fat in the development of CVD.

In this paper, we present a method which is able ofautomatically segmenting the pericardium on non-enhancedcardiac CT scans and subsequently quantifying the epicardialfat volume within the pericardium. Our method uses anatlas-based segmentation approach in order to segment thepericardium. The atlas-based segmentation approach is anadaptation of our previous work, where atlas-based segmen-tation was evaluated with respect to segmenting the heart andits chambers in a multicenter, multivendor contrast-enhancedCT (CTA) study.20 Our method was evaluated on 98 CT scanswith respect to (1) the accuracy of pericardium segmentation,(2) the accuracy of epicardial fat quantification, and (3) accu-racy of the results with respect to the inter-observer variabil-ity. The evaluation was conducted by comparing the perfor-mance of our method to two independent manual observers.

The remaining of the paper is organized as follows. Sec-tion 2 gives details about the imaging data, overview of ourmethod and the experiments we performed. We present ourresults in Sec. 3, discussion and future work in Sec. 4, andfinally the conclusion is provided in Sec. 5.

2. MATERIAL AND METHODS

2.A. Study population and imaging protocol

For this study, we randomly selected 98 subjects fromthe population-based Rotterdam Study,21 who underwent amulti-detector computed tomography (MDCT) scan of theheart. This study was part of a larger MDCT-project involving

Medical Physics, Vol. 40, No. 9, September 2013

091910-3 Shahzad et al.: Automatic epicardial fat volume quantification 091910-3

TABLE I. Characteristics of the subjects (n = 98). Values are mean ± SD forcontinuous variables and numbers (%) for dichotomous variables.

Variable Value

Women 46 (46.9%)Age (years) 69.4 ± 5.6BMI (kg/m2) 27.4 ± 3.9Systolic blood pressure (mmHg) 147.8 ± 22.3Diastolic blood pressure (mmHg) 82.2 ± 11.1Smoking (ever) 69 (70.4%)Diabetes 7 (7.1%)Total cholesterol (mmol/L) 5.8 ± 0.9

calcium-scoring in multiple vessel beds.22 The participantswere scanned on two different generations of Siemens scan-ners [Sensation 16 (n = 55) or Sensation 64 (n = 43), SiemensMedical Solutions, Forchheim, Germany]. The cardiac scanranged from the apex of the heart to the tracheal bifurcation.The Rotterdam Study was approved by the Institutional Re-view Board with additional specific approval of the CT study.All participants gave written informed consent for the CT ex-amination.

Subject characteristics are provided in Table I. Scan set-tings were as follows. The scan reached from the apex ofthe heart to the tracheal bifurcation, no contrast material wasused. On the 16-slice scanner, consecutive nonoverlapping 3.0mm thick slices were acquired within a single breath hold.The collimation was 12 × 1.5 mm, the tube voltage was 120kV, the effective tube current 30 mAs, and prospective ECGtriggering at 50% of the cardiac cycle was used. For the 64-slice scanner, all parameters except the collimation and tubecurrent changed. Collimation was set to 32 × 0.6 mm and

the tube current was adopted with respect to the body weight(CARE DOSE, Siemens, Forcheim, Germany) with a refer-ence value of 50, 100, and 190 mAs. The images from bothscanners were reconstructed with a 3.0 mm increment and anaverage field of view (FOV) of 180 mm. The images were re-constructed with a matrix of 512 × 512 using a b35f (mediumsharp) kernel, and contained on average 52 slices.

2.B. Method overview

The quantification method consists of two steps: (1) peri-cardium segmentation, and (2) epicardial fat volume quantifi-cation. For pericardium segmentation, we used a multi-atlassegmentation approach, as described in the work of Kirisliet al.20 In this approach, manually segmented CTA scans (at-lases) are registered to the subject’s CT scan. The segmenta-tions of these atlases are mapped onto the subject’s scan to beanalyzed. The mapped segmentations from each of the atlasesare fused to obtain the final pericardium segmentation. Thiswhole procedure is fully automatic. The epicardial fat is sub-sequently quantified by applying a threshold of −200 to −30HU (Ref. 23) to the segmented pericardium, followed by con-nected component analysis. Details with respect to the atlasesused, the registration approach, and the fat quantification arepresented in the sections 2.C, 2.D, and 2.E. Figure 1 shows anoverview of all the steps involved in the automatic method.

2.C. Atlas selection and surface computation

CTA scans were used as atlas images because of theirhigher resolution and the better visibility of the cardiac cham-bers (due to the presence of contrast material). This en-ables the observers to accurately delineate the pericardium

FIG. 1. Overview of the segmentation process. Top left: Eight CTA atlas scans and the corresponding manually segmented 3D surfaces. Top right: CT scan tobe segmented. The atlas scans are registered and the 3D surfaces are transformed to match the subject scan. Bottom right: The 3D surfaces are combined usingmajority voting. Bottom left: Resulting fat voxels, obtained after thresholding and connected component analysis.

Medical Physics, Vol. 40, No. 9, September 2013

091910-4 Shahzad et al.: Automatic epicardial fat volume quantification 091910-4

FIG. 2. The eight pericardium atlas surfaces used for atlas-based segmenta-tion illustrating the encountered shape and size variations in the atlas images.

manually. Eight previously acquired CTA scans from differ-ent subjects were included in as the atlas scans. Readers arereferred to Ref. 20 for detailed information on atlas selection.

The manually obtained contours of the atlas scans wereconverted to 3D surfaces. Figure 2 shows the resulting heartsurfaces for the eight atlas scans.

2.D. Multi-atlas based segmentation

Multi-atlas segmentation24 is a process in which multipleatlas images with corresponding manually annotated label im-ages are individually registered to the subject scans. The seg-mentation of the subject scans is then obtained by fusing allthe resulting transformed label images.25 In this work, we usemajority voting to fuse the label images.

Image registration26 is used to spatially align the atlasscans and the subjects scan. In the registration procedure, thetransformation parameters T that minimizes the cost functionC between the fixed image (If) and the a moving image (Im)are determined. Detailed information on registration is pro-vided in Refs. 27 and 28.

The registration steps and the parameters are explained inmore detail in Sec. 2.H, where we also compare the perfor-mance of registration with and without masking certain areasof the image.

The resulting transformations from the registration stepsare used to map each of the eight atlas surfaces onto the sub-

ject’s scan. Once this is done, the 3D surface intensities areconverted to binary masks and majority voting is applied toobtain the final pericardium segmentation (see Fig. 1 for avisual representation).

2.E. Epicardial fat quantification

The automatically obtained segmented pericardium is usedas a region of interest (ROI) to quantify the adipose tissuevoxels. A threshold window of −200 to −30 HU is applied toobtain the adipose tissue. A connected-component analysis29

is subsequently applied to all adipose tissue voxels using an18-neighborhood rule, in order to remove regions smaller than10 voxels (2.8 mm3) in size, which we consider to be noise.

2.F. Reference standard

Two experienced observers (D.B. and A.R.), blinded to thepatient information as well as to the results of each other,manually traced the pericardium in each of the CT scans, asshown in Fig. 3(a). A dedicated tool implemented in MeVis-Lab (Ref. 30) was used by the observers for manual anno-tations. Once the pericardium was delineated, a thresholdwindow of −200 to −30 HU was applied to the segmentedregion.23 Adipose tissue voxels were then automatically ex-tracted using connected-component analysis and the volumeof fat in milliliters (ml) was computed. The reference standardobtained this way contains both pericardium segmentationsand epicardial fat volume quantifications [Fig. 3(b)].

2.G. Statistical analysis

We report the Dice similarity index and the mean surfacedistance error between the pericardial heart segmentation bythe automatic method and each of the manually obtained seg-mentations. To evaluate the performance of the fat quantifi-cation method per patient, Pearson’s correlation coefficient(R) was calculated, linear regression was performed, and

FIG. 3. (a) A random axial slice showing the result of a manually obtained whole heart segmentation (with adjusted windowing level, for better visibility).(b) Corresponding slice showing the voxels containing epicardial fat.

Medical Physics, Vol. 40, No. 9, September 2013

091910-5 Shahzad et al.: Automatic epicardial fat volume quantification 091910-5

FIG. 4. Random axial slice with the overlay representing the fixed maskused for the registration.

Bland-Altman plots were created. Furthermore, the accuracyof the method was compared to the interobserver variability.The analyses were performed using MATLAB version 7.9.0.(The MathWorks, Natick, MA) and IBM SPSS Statisticsversion 20 (IBM Corp, Armonk, NY).

2.H. Experiments

Our method consists of segmenting nonenhanced CT scanswith the help of contrast-enhanced CT atlas scans. The reg-istration problem we face here is that the fixed image (If)and moving image (Im) have different characteristics, both interms of contrast and resolution. The CT scans in which weaim to quantify the epicardial fat has an average in-plane res-olution of 0.35 × 0.35 mm2 and a slice thickness of 3.0 mm,whereas the CTA atlases have an average in-plane resolutionof 0.32 × 0.32 mm2 and a slice thickness of 0.4 mm. In or-der to obtain the optimal parameters to register the CTA at-lases and the CT subjects, we performed pilot experiments ona subset of 35 randomly selected CT datasets. Two registra-tion strategies were investigated and the segmentation resultswere compared to the results of one of the observers. In bothstrategies, the CTA atlas was used as the fixed image (If) andthe subjects CT scan was used as the moving image (Im).

TABLE II. Table representing the results of the registration strategy. Valuesrepresent mean ± SD.

Strategy 1 Strategy 2

Dice similarity index % 90.0 ± 3.2 89.6 ± 1.9Mean surface distance mm 3.4 ± 1.3 3.5 ± 0.8

Strategy 1: The similarity metric was computed by ran-domly sampling intensity values from the whole image.

Strategy 2: The similarity metric was computed by ran-domly sampling intensity values within the heart region only(by using a fixed heart mask).

For both strategies, a two-stage registration approach wasused. In the first stage, an affine transformation was used.In the second stage, a non-rigid registration using a B-spline transformation31 was employed while using the resultsof the affine transformation to initialize the registration. Mu-tual information32 was used as similarity measure for thecost function. Optimization was performed using adaptivestochastic gradient descent, the number of voxels sampled ineach iteration was set to 2048,33 and the number of iterationswere set to 512 for the affine transformation and 2048 for theB-spline transformation. For further details about parameterselection and optimizations, readers are referred to our previ-ous study.20 All registrations were performed using Elastix,34

a publicly available software package.35

The heart mask was used in both stages of the registra-tion approach. The main purpose of using the fixed mask wasto prevent the registration to be affected by tissues surround-ing the heart, such as the lungs, rib cage, and the vertebra.Figure 4 shows a random axial slice of one of the fixed masksused for the registration optimization. The mask was createdby dilating the original manually annotated pericardium by1 cm. The masks were created once, only for the atlas scansand not the subject scans. Hence, called the fixed mask.Table II shows the results obtained for both the strategies. Itcan be observed that the average accuracy of both strategieswas very similar in terms of the mean, but using the mask re-sulted in a smaller standard deviation. It was also confirmedvisually that the accuracy of finding the pericardium usingStrategy 2 was better than when using Strategy 1. Further ex-periments on the entire dataset were thus performed using theregistration with the mask.

3. RESULTS

3.A. Agreement between the automatic methodand the observers

A visual check showed that 95 out of 98 segmentationswere successful. Three segmentations failed due to registra-tion errors caused by anatomical and FOV variations. Thesescans were excluded from further analysis.

The subjects had an average fat volume of 101 ± 38 mlaccording to Observer 1 and 113 ± 43 ml according to Ob-server 2. The automatic method found the average fat vol-ume to be 102 ± 34 ml. A Dice similarity index of 89.1%and 89.2% was obtained between the automatic segmentationand each of the manual segmentations, respectively. The meansurface distance between the automatically derived cardiacsurface and the observer segmentations was 3.8 ± 1.1 mm and3.5 ± 0.7 mm, respectively. A Pearson correlation R of 0.91(P < 0.001) was obtained for fat quantification results be-tween the automatic segmentation and each of the manualsegmentations. The mean absolute difference between the

Medical Physics, Vol. 40, No. 9, September 2013

091910-6 Shahzad et al.: Automatic epicardial fat volume quantification 091910-6

FIG. 5. Scatter plots of automatic vs manual fat quantification methods and results from linear regression: (a) correlation between the two observers;(b) correlation between Observer 1 and the automatic method; and (c) correlation between Observer 2 and the automatic method.

automatic method and each of the manual segmentations withrespect to the amount of quantified fat was 11.6 ml and 16.6ml, respectively. The linear regression plots are shown inFig. 5. The numbers obtained from the Bland–Altman anal-ysis and the confidence intervals of the linear regression areshown in Table III with a graphical representation in Fig. 6.It can be noted that the bias from the Bland–Altman analysiswith respect to Observer 1 is almost zero and the automaticmethod slightly overestimates the volume of epicardial fat ascompared to Observer 2.

3.B. Interobserver agreement

An average Dice similarity coefficient of 88.9% was foundbetween the segmentations of the observers. The mean sur-face distance between the two observers was 4.3 ± 1.0 mmover all datasets. With respect to the amount of quantified epi-cardial fat, the mean absolute difference between the two ob-servers was 15.6 ml, and the Pearson correlation coefficientwas 0.92 (P < 0.001). A Bland–Altman analysis of the datashowed that the limits of agreement were between −45.3 and

21.3 ml and a bias of 12.1. Figures 5 and 6 show the correla-tion graph and the Bland–Altman analysis.

4. DISCUSSION

In this study, we presented a fully automatic method forepicardial fat quantification. The method is based on au-tomatic pericardium segmentation. A good correlation withmanual quantification was observed, with differences verysimilar to the inter-observer variability.

The Dice similarity index (overlap area) between the auto-matic pericardium segmentation and each of the manual an-notations was slightly better than the inter-observer Dice sim-ilarity coefficient. The mean surface distance error betweenthe automatic and manual segmentations corresponds to1.5 voxels in the slice direction, which can be consideredsmall. When the actual amount of fat volume quantified us-ing our method was compared to each of the observers, itresulted in a mean absolute difference of 11.6 ml and 16.6ml, respectively. This difference in volume is very close tothe inter-observer agreement, which was 15.6 ml. The sameconclusion can be drawn from the correlation coefficient R

Medical Physics, Vol. 40, No. 9, September 2013

091910-7 Shahzad et al.: Automatic epicardial fat volume quantification 091910-7

TABLE III. Performance of the whole heart segmentation: comparing the automatic method to each of the observers and the observers to each other.

Automatic vs Automatic vs Observer 1 vsObserver 1 Observer 2 Observer 2

Segmentation measuresDice similarity index (mean ± SD)% 89.1 ± 2.6 89.2 ± 1.9 88.9 ± 2.5Mean surface distance (mean ± SD) mm 3.8 ± 1.1 3.5 ± 0.7 4.3 ± 1.0Quantification measuresCorrelation R 0.91 0.91 0.92Linear regression (CI for β) 0.75–0.90 0.65–0.79 0.96–1.15Bland-Altman bias (95% CI) −0.8 (−31 to 29) 11.3 (−25 to 47) 12.1 (−21 to 45)

obtained with respect to the quantified fat volume of the auto-matic method and the manual observers.

Compared to the existing quantification methods, ourmethod is the first that is fully automatic. The methods pro-posed in Refs. 8 and 13–17 either use manual tracings ofthe different tissue types, or a manual approach to delineatethe pericardium, before quantifying the adipose tissue voxels.This is a tedious and time-consuming task to perform. The

semi-automatic method proposed in Refs. 18 and 19 needstwo interactions, which could limit the use of the method inprocessing a large number of datasets in an epidemiologicsetting.

The three subjects that were excluded from the analy-sis had the following issues: one subject underwent pneu-monectomy (removal of a lung) causing a very unusual po-sition of the heart [see Fig. 7(a)], the other had a heart shape

FIG. 6. Bland–Altman analysis: (a) between observers; (b) between automatic method and Observer 1; and (c) between automatic method and Observer 2.

Medical Physics, Vol. 40, No. 9, September 2013

091910-8 Shahzad et al.: Automatic epicardial fat volume quantification 091910-8

FIG. 7. Excluded subjects: (a) subject with lung removed and (b) segmentation leaking into the ribcage due to rare anatomical variation in heart shape.

anatomically quite different from the others [see Fig. 7(b)],and the last one had a different field of view compared tothe atlas scans used. The large difference between the atlasscan and the subject scan caused the registration to fail, whichresulted in erroneous segmentation of the pericardium.

There has been some confusion in the literature betweenthe nomenclatures of the adipose tissue contained withinthe pericardium;36 some studies call it epicardial fat tissue,whereas others call it pericardial fat tissue. Based on the def-inition provided here,37 we decided to denote the adiposetissue contained within the pericardium as epicardial fat. Inshort, in this definition epicardial fat is the adipose tissuebetween the myocardium and the visceral layer of the peri-cardium.

We did not investigate to what extent the method can beused on multiple scanner types; in this study, we only demon-strated the feasibility of using the method on two genera-tions of Siemens scanners. However, as our method is basedon multi-atlas segmentation, we are confident that the sameapproach would work on other scanner types, as long as thesubject scans and the atlas scans have a similar field of view.It has been demonstrated in our previous study,20 that atlas-based segmentation of the pericardium was performed with asimilar accuracy with respect to multivendor/multicenter CTAdatasets. If required, the method could utilize atlas scans fromthe same scanner.

In the current setup, visual inspection was still required tocheck the accuracy of the pericardium segmentation, whichresulted in discarding the three scans on which the segmen-tation failed. Instead of discarding these scans, or in case ofsmall failures, manual correction before fat quantification isan option. We did not integrate this in our protocol, as theresults were sufficiently accurate without adaptation.

5. CONCLUSION

We developed and evaluated an automatic method for peri-cardium segmentation and subsequent epicardial fat quan-tification. We demonstrated that our automatic approach

achieved good correlation to manual quantifications. The au-tomatic method described in this paper could potentially beused on large clinical or population studies in order to investi-gate the relationship between epicardial fat volume and CVD.

ACKNOWLEDGMENTS

Rahil Shahzad and Hortense Kirisli are supported bya grant from the Dutch Ministry of Economic Affairs(AgentschapNL) under the title “Het Hart in Drie Dimensies”(PID06003). Coert Metz and Theo van Walsum are supportedby a grant from the Information Technology for European Ad-vancement (ITEA), under the title “Patient Friendly MedicalIntervention” (project 09039, Mediate). Stefan Klein is sup-ported by a grant from the Netherlands Science Organisation(NWO), division of Exact Sciences (project 639.021.919).

a)Electronic mail: R.Shahzad@tudelft.nl1World Health Organization, Cardiovascular Diseases, Fact Sheet 317(World Health Organization, Switzerland, 2013).

2K. Cheng, C. Chu, K. Lee, T. Lin, C. Hsieh, C. Chiu, W. Voon, S. Sheu, andW. Lai, “Adipocytokines and proinflammatory mediators from abdominaland epicardial adipose tissue in patients with coronary artery disease,” Int.J. Obes. 32, 268–274 (2007).

3G. Iacobellis, D. Pistilli, M. Gucciardo, F. Leonetti, F. Miraldi, G.Brancaccio, P. Gallo, and C. di Gioia, “Adiponectin expression in humanepicardial adipose tissue in vivo is lower in patients with coronary arterydisease,” Cytokine 29, 251–255 (2005).

4M. Kortelainen, “Myocardial infarction and coronary pathology in severelyobese people examined at autopsy,” Int. J. Obes. 26, 73–79 (2002).

5A. A. Mahabadi, N. Reinsch, N. Lehmann, J. Altenbernd, H. Kälsch,R. M. Seibel, R. Erbel, and S. Möhlenkamp, “Association of pericoronaryfat volume with atherosclerotic plaque burden in the underlying coronaryartery: A segment analysis,” Atherosclerosis 211, 195–199 (2010).

6N. Bettencourt, A. Toschke, D. Leite, J. Rocha, M. Carvalho, F. Sampaio,S. Xara, A. Leite-Moreira, E. Nagel, and V. Gama, “Epicardial adiposetissue is an independent predictor of coronary atherosclerotic burden,” Int.J. Cardiol. 158, 26–32 (2012).

7N. Alexopoulos, D. S. McLean, M. Janik, C. D. Arepalli, A. E. Stillman,and P. Raggi, “Epicardial adipose tissue and coronary artery plaque char-acteristics,” Atherosclerosis 210, 150–154 (2010).

8G. Rosito, J. Massaro, U. Hoffmann, F. Ruberg, A. Mahabadi, R. Vasan,C. ODonnell, and C. Fox, “Pericardial fat, visceral abdominal fat,

Medical Physics, Vol. 40, No. 9, September 2013

091910-9 Shahzad et al.: Automatic epicardial fat volume quantification 091910-9

cardiovascular disease risk factors, and vascular calcification in acommunity-based sample: The Framingham Heart Study,” Circulation 117,605–613 (2008).

9A. M. de Vos et al., “Peri-coronary epicardial adipose tissue is relatedto cardiovascular risk factors and coronary artery calcification in post-menopausal women,” Eur. Heart J. 29, 777–783 (2008).

10S. Sarin, C. Wenger, A. Marwaha, A. Qureshi, B. D. Go, C. A. Woomert,K. Clark, L. A. Nassef, and J. Shirani, “Clinical significance of epicar-dial fat measured using cardiac multislice computed tomography,” Am. J.Cardiol. 102, 767–771 (2008).

11A. A. Mahabadi, J. M. Massaro, G. A. Rosito, D. Levy, J. M. Murabito,P. A. Wolf, C. J. O’Donnell, C. S. Fox, and U. Hoffmann, “Associationof pericardial fat, intrathoracic fat, and visceral abdominal fat with cardio-vascular disease burden: The Framingham Heart Study,” Eur. Heart J. 30,850–856 (2009).

12K. Harada,“Cardiac 64-multislice computed tomography reveals increasedepicardial fat volume in patients with acute coronary syndrome,” Am. J.Cardiol. 108, 1119–1123 (2011).

13J. Ding et al., “The association of pericardial fat with incident coronaryheart disease: The Multi-Ethnic Study of Atherosclerosis (MESA),” Am. J.Clin. Nutr. 90, 499–504 (2009).

14A. A. Mahabadi et al., “Association of epicardial fat with cardiovascularrisk factors and incident myocardial infarction in the general population:The Heinz Nixdorf Recall Study,” J. Am. Coll. Cardiol. 61, 1388–1395(2013).

15R. Taguchi, J. Takasu, Y. Itani, R. Yamamoto, K. Yokoyama, S. Watanabe,and Y. Masuda, “Pericardial fat accumulation in men as a risk factor forcoronary artery disease,” Atherosclerosis 157, 203–209 (2001).

16G. Wheeler et al., “Pericardial and visceral adipose tissues measured volu-metrically with computed tomography are highly associated in type 2 dia-betic families,” Invest. Radiol. 40, 97–101 (2005).

17J. Ding, S. Kritchevsky, T. Harris, G. Burke, R. Detrano, M. Szklo, andJ. Carr, “The association of pericardial fat with calcified coronary plaque,”Obesity 16, 1914–1919 (2008).

18D. Dey, Y. Suzuki, S. Suzuki, M. Ohba, P. Slomka, D. Polk, L. Shaw, andD. Berman, “Automated quantitation of pericardiac fat from noncontrastCT,” Invest. Radiol. 43, 145–153 (2008).

19D. Dey et al., “Computer-aided non-contrast CT-based quantification ofpericardial and thoracic fat and their associations with coronary calciumand metabolic syndrome,” Atherosclerosis 209, 136–141 (2010).

20H. Kirisli et al., “Evaluation of a multi-atlas based method for segmentationof cardiac CTA data: A large-scale, multicenter, and multivendor study,”Med. Phys. 37, 6279–6291 (2010).

21A. Hofman et al., “The Rotterdam Study: 2012 objectives and design up-date,” Eur. J. Epidemiol. 26, 657–686 (2011).

22S. Elias-Smale, R. Wieberdink, A. Odink, A. Hofman, M. Hunink, P. Koud-staal, G. Krestin, M. Breteler, A. van der Lugt, and J. Witteman, “Burdenof atherosclerosis improves the prediction of coronary heart disease butnot cerebrovascular events: The Rotterdam Study,” Eur. Heart J. 32, 2050–2058 (2011).

23T. Yoshizumi et al., “Abdominal fat: Standardized technique for measure-ment at CT,” Radiology 211, 283–286 (1999).

24T. Rohlfing, R. Brandt, R. Menzel, D. B. Russakoff, and C. R. Maurer, Jr.,“Quo vadis, atlas-based segmentation?,” in Handbook of Biomedical ImageAnalysis (Springer U.S., 2005), pp. 435–486.

25T. Rohlfing, R. Brandt, R. Menzel, and C. Maurer, “Evaluation of atlasselection strategies for atlas-based image segmentation with application toconfocal microscopy images of bee brains,” Neuroimage 21, 1428–1442(2004).

26J. Maintz and M. Viergever, “A survey of medical image registration,”Med. Image Anal. 2, 1–36 (1998).

27I. Bankman, Handbook of Medical Imaging: Processing and Analysis, Reg-istration (Academic Press, CA, 2000).

28J. Suri, D. Wilson, and S. Laxminarayan, Handbook of Biomedical ImageAnalysis: Registration Models (Plenum, New York, 2005), Vol. III.

29H. Samet and M. Tamminen, “Efficient component labeling of images ofarbitrary dimension represented by linear bintrees,” IEEE Trans. PatternAnal. Mach. Intell. 10, 579–586 (1988).

30http://www.mevislab.de.31D. Rueckert, L. Sonoda, C. Hayes, D. Hill, M. Leach, and D. Hawkes,

“Nonrigid registration using free-form deformations: application to breastMR images,” IEEE Trans. Med. Imaging 18, 712–721 (1999).

32D. Mattes, D. Haynor, H. Vesselle, T. Lewellen, and W. Eubank, “PET-CTimage registration in the chest using free-form deformations,” IEEE Trans.Med. Imaging 22, 120–128 (2003).

33S. Klein, J. Pluim, M. Staring, and M. Viergever, “Adaptive stochastic gra-dient descent optimisation for image registration,” Int. J. Comput. Vis. 81,227–239 (2009).

34S. Klein, M. Staring, K. Murphy, M. Viergever, and J. Pluim, “Elastix: Atoolbox for intensity-based medical image registration,” IEEE Trans. Med.Imaging 29, 196–205 (2010).

35http://elastix.isi.uu.nl.36S. Y. Thomas, R. Harmancey, and H. Taegtmeyer, “Fat around the heart,”

J. Am. Coll. Cardiol.: Cardiovasc. Imaging 3, 786–787 (2010).37G. Iacobellis, “Epicardial and pericardial fat: Close, but very different,”

Obesity 17, 625–625 (2009).

Medical Physics, Vol. 40, No. 9, September 2013