1 23

International Journal of ComputerAssisted Radiology and SurgeryA journal for interdisciplinary research,development and applications of imageguided diagnosis and therapy ISSN 1861-6410 Int J CARSDOI 10.1007/s11548-012-0770-x

Automatic left ventricle segmentation involumetric SPECT data set by variationallevel set

Mohammad Hosntalab, FarshidBabapour-Mofrad, Nazgol Monshizadeh& Mahasti Amoui

1 23

Your article is protected by copyright and all

rights are held exclusively by CARS. This e-

offprint is for personal use only and shall not

be self-archived in electronic repositories.

If you wish to self-archive your work, please

use the accepted author’s version for posting

to your own website or your institution’s

repository. You may further deposit the

accepted author’s version on a funder’s

repository at a funder’s request, provided it is

not made publicly available until 12 months

after publication.

Int J CARSDOI 10.1007/s11548-012-0770-x

ORIGINAL ARTICLE

Automatic left ventricle segmentation in volumetric SPECT dataset by variational level set

Mohammad Hosntalab · Farshid Babapour-Mofrad ·Nazgol Monshizadeh · Mahasti Amoui

Received: 10 January 2012 / Accepted: 30 May 2012© CARS 2012

AbstractIntroduction Left ventricle (LV) quantification in nuclearmedicine images is a challenging task for myocardial perfu-sion scintigraphy. A hybrid method for left ventricle myo-cardial border extraction in SPECT datasets was developedand tested to automate LV ventriculography.Methods Automatic segmentation of the LV in volumetricSPECT data was implemented using a variational level setalgorithm. The method consists of two steps: (1) initializa-tion and (2) segmentation. Initially, we estimate the initialclosed curves in SPECT images using adaptive thresholdingand morphological operations. Next, we employ the initialclosed curves to estimate the final contour by variational levelset. The performance of the proposed approach was evaluatedby comparing manually obtained boundaries with automatedsegmentation contours in 10 SPECT data sets obtained fromadult patients. Segmented images by proposed methods werevisually compared with manually outlined contours and theperformance was evaluated using ROC analysis.Results The proposed method and a traditional level setmethod were compared by computing the sensitivity andspecificity of ventricular outlines as well as ROC analysis.The results show that the proposed method can effectively

M. Hosntalab (B) · F. Babapour-Mofrad · N. MonshizadehFaculty of Engineering, Science and Research Branch,Islamic Azad University (IAU), Tehran, Irane-mail: mhosntalab@yahoo.com

F. Babapour-Mofrade-mail: farshid.mofrad@yahoo.com

N. Monshizadehe-mail: nazgol_monshizadeh@yahoo.com

M. AmouiNuclear Medicine Department,Shahid Beheshti University of Medical Sciences, Tehran, Irane-mail: amoui@sbmu.ac.ir

segment LV regions with a sensitivity and specificity of 88.9and 96.8 %, respectively. Experimental results demonstratethe effectiveness and reasonable robustness of the automaticmethod.Conclusion A new variational level set technique was ableto automatically trace the LV contour in cardiac SPECT datasets, based on the characteristics of the overall region of LVimages. Smooth and accurate LV contours were extractedusing this new method, reducing the influence of nearby inter-fering structures including a hypertrophied right ventricle,hepatic or intestinal activity, and pulmonary or intramam-mary activity.

Keywords Cardiac SPECT · LV segmentation ·Myocardial perfusion scintigraphy · Variational level set

Introduction

Coronary artery disease is one of the most common causes ofdeath in the world. Modern functional medical imaging tech-niques such as Single Photon Emission Computed Tomog-raphy (SPECT), Magnetic Resonance Imaging (MRI) andEchocardiography (cardiac ECHO or ECHO) can contributesignificantly to diagnose CRD and particularly its quantita-tive assessment.

Cardiac perfusion studies are carried out routinely in hos-pitals around the world. It allows physicians to see how wellblood is reaching the heart muscle through the coronary arter-ies. It provides a reliable assessment of coronary stenosis orobstruct leading to inducible perfusion abnormalities and ofthe prognosis of disease. Therefore, it can be used both fordiagnosis and for triage of patients between initial medicaltherapy and invasive investigation with regard to revascu-larization. Myocardial perfusion scintigraphy (MPS) is an

123

Author's personal copy

Int J CARS

effective and cost-effective technique and SPECT is used asthe most common form of image acquisition [1]. Addition-ally functional parameters obtained with Gated acquisitioncould assist the diagnosis. It includes cardiac end-diastolicand end-systolic volumes and left ventricular ejection frac-tion which are calculated by detecting myocardial borders ofthe left ventricle (LV).

To the best our knowledge, several authors have reporteddifferent techniques for ventricles segmentation, but only afew of them deal with SPECT data. In [2], an automaticmethod based on region-driven active contours was used forleft and the right ventricle cavity segmentation in magneticresonance (MR) images. The algorithm in [3] was developedfor automatic segmentation of the left ventricular endocardialboundary in echocardiographic images. In [4,5], geometricactive contour models have been used in cardiac SPECT seg-mentation with various modifications for MPS. A deformablemodel based on charged particle dynamics and geometriccontour propagation was utilized in [6] to trace LV of the heartin SPECT images. In [7], the LV was segmented by registra-tion using a dynamic anatomic model and prior knowledge.

In the area of segmentation, manual segmentation is notonly a tedious and time consuming process, but also onewith unfeasible and inaccurate calculation results. Manualsegmentation by experts is variable up to 20 % [8]. Also, con-ventional approaches such as global thresholding employedin most existing systems, are not adequate for segmentationof LV in SPECT images. In this regard, computer-based auto-matic segmentation methods have been introduced with highaccuracy and little user interaction, desirable for fully exploit-ing medical data.

The aim of this paper is to introduce an efficient approachthat could potentially be used in segmentation of LV inSPECT volumetric data set. It is proposed to meet demandsof clinicians in common clinical practice and resolve someconstraints of current cardiac software in LV segmentation.In order to validate the proposed method, the resulting auto-matically defined borders are compared to those manuallyidentified by experts.

Subsequently, In section “Materials and methods”, thedata set and proposed techniques are given in details. Exper-imental results are mentioned in section “Experiments”. Insection “Discussion”, we clarify the practical aspects of theproposed algorithms, i.e. the benefits and the limitations. Thepaper in closed with remarkable conclusion and future plansmade in section “Conclusion”.

Materials and methods

In this section, we describe the data set and proposed tech-nique in details. This section consists of the two parts. Ini-tially we introduce data acquisition and processing protocol.Then we present proposed method of LV segmentation.

Acquisition and processing protocol

Ten cases with definite or suspected coronary artery disease,4 female and 6 male patients, were included in the studywith mean age of 60.80 ± 10.90 (ranging from 44 to 79).Stress and rest studies were carried out on 2 separate days.Images were obtained 45–90 min after intravenous injection296–925 MBq (8–15 mCi) of Tc-99m MIBI.

Image system was a single head e-cam SPECT camera(SIEMENS-Germany) equipped with a low-energy high res-olution collimator. In SPECT study, the data were acquiredin 64 × 64 image matrix for 32 projections over 180◦ arc,25 s per projection, from 45◦ right anterior oblique (RAO)to the 45◦ left posterior oblique (LOP). Then, images werereconstructed by filtered back projection method, low passButterworth filter (cut off: 0.4-order: 5), followed by viewingLV myocardium in transaxial, vertical long axis and horizon-tal long axis planes. Thereafter, semi quantitative analysis ofLV myocardium perfusion has been done by Cedars-SinaiSoftware (20 segment model).

Proposed method of LV segmentation

We propose a hybrid approach for LV segmentation in car-diac SPECT data set. Major steps of the proposed techniquesare as follows: (1) initialization (2) segmentation. Figure 1shows the block diagram of the techniques.

Initialization: We propose a morphology-based operationto automatically estimate the initial closed curves. In thisregard, the images are initially binarised with an adaptivethresholding such as Otsu’s thresholding [9]. The result isgiven in Fig. 1b. Then a process of thinning [10,11] thatiteratively searches for the locus of the points that are equi-distant to the object border is provided. The final set of pointsis shown in Fig. 1c.

Segmentation: We employ the initial closed curves to esti-mate the final contour by variational level set (Fig. 1d). Weutilize the results of proposed method by applying connectedcomponent analysis using 8-connectivity [10] to extract thecontour of each LV. A set of black pixels, P, is an 8-connectedcomponent if for every pair of pixels pi and p j in P, thereexists a sequence of pixels pi , …, p j such that all pixels inthe sequence are in the set P i.e. are black, and every two pix-els that are adjacent in the sequence are 8-neighbors. Resultsare illustrated in Fig. 1e, f.

Since active contour model was first proposed in [12],has been widely studied and used in the field of image anal-ysis. Active contours are curves that deform within digitalimages to recover object shapes. They are classified as eitherparametric active contour models [13] or geometric activecontour models [14] according to their representation and

123

Author's personal copy

Int J CARS

Fig. 1 Block diagram of the proposed framework

implementation. Geometric active contours are representedimplicitly based on the theory of curve evolution imple-mented via level set techniques [15]. Level set method pro-vides an alternative solution to energy minimization-basedimage segmentation problem. Level set method based onedges is driven by the derivatives of image intensities. Math-ematical formulation of level set is explained as fallow:

Let � be a bounded open subset of �2, with ∂� asits boundary. Let U0 : � → � be a given image, andC : [0, 1] → �2 be a parameterized curve. The curveC is represented implicitly via a Lipschitz function φ byC = {(x, y) |φ (x, y) = 0 }, and the evolution of the curveis given by the zero-level curve at time t as the functionφ(x, y, t). Evolving the curve C in normal direction withspeed F leads to differential equation

{∂φ∂t = |∇φ| F,

φ(x, y, 0) = φ0(x, y),(1)

where the set C = {(x, y) |φ0 (x, y) = 0 } defines the ini-tial closed curve. A particular case is the motion by mean

curvature, when F = div( ∇φ

|∇φ|)

is the curvature.

Myocardial perfusion SPECT data is noisy, the level ofdetail is low and LV may be extracted from it only if anexternal model is supplied, hence, such techniques will notbe able to accurately extract the LV contour. To address this

limitation, we utilized a variational level set method [16]. Thevariational level set method utilizes an energy functional con-sisting of surface tension (proportional to length) and bulkenergies (proportional to area). This approach combines thelevel set method with a theoretical variational formulation. Itis very robust to noise, which presents serious challenge tomany traditional techniques. Assume that there are n disjointregions �i (1 ≤ i ≤ n) in the image. The common boundarybetween �i and � j is denoted as �i j . Variational level setfunction can be expressed as

inf E =∑

1≤i≤ j≤n

fi j Length(�i j

)

+∑

1≤i≤n

νi × Area (inside (Ci )). (2)

The level set function is expressed as

⎧⎪⎪⎨⎪⎪⎩

∂φi∂t =|∇φi |

⎛⎝γi div

( ∇φ|∇φ|

)−ei −λ

⎛⎝ n∑

j=1

H(φ j

)−1

⎞⎠

⎞⎠ ,

∂φi∂n = 0 on ∂�,

(3)

where n denotes the exterior to the boundary ∂�, and ∂φ/∂ndenotes normal derivative of φ at the boundary.

123

Author's personal copy

Int J CARS

We apply a novel variational level set technique based onMumford–Shah function [17] for segmentation. This modelis able to detect contours with or without a gradient. Objectswith smooth boundaries or even with discontinuous bound-aries can be successfully detected. Moreover, the model isrobust to the position of the initial closed curve whereastraditional level set method is sensitive to the placementof the initial closed curves. Therefore, this model estimatesthe accurate internal and external border of LV myocardiumregardless of the initial position of the thinned curve whichis included in the original image data. The 2D version of themodel can be expressed as

inf(c1,c2,C) E = μ Length(C)+ ν Area (Inside(C))+ EMV,

(4)

with

EMV = λ1

∫inside(C)

(u0(x, y) − c1)2 dxdy

+ λ2

∫outside(C)

(u0(x, y) − c2)2 dxdy, (5)

where Ci are the averages of u0 inside and outside C , andμ ≥ 0, ν ≥ 0, and λi � 0 are fixed parameters. The level setfunction they obtain is given by⎧⎪⎪⎪⎨⎪⎪⎪⎩

∂φ∂t = δε(φ)

[μ.div

( ∇φ|∇φ|

)− υ − λ1 (u0 − c1)

2

+λ2(u0 − c2)2]

φ(x, y, 0) = φ0(x, y) in �,δε(φ)∂φ|∇φ|∂n = 0 on ∂�,

(6)

where n denotes the exterior to the boundary ∂�, and ∂φ/∂ndenotes normal derivative of φ at the boundary and δε is theDirac delta function. In order to solve this partial differentialequation, we first need to regularize Hε(φ) and δε(φ).

{Hε(φ) = 1

2 + 1π

arctan(

φε

),

δε(φ) = 1π. εε2+φ2 .

(7)

It is easy to see that as ε → 0, Hε(φ) converges to H(φ) andδε(φ) converges to δ(φ)

H(φ) ={

1 if φ ≥ 00 if φ ≺ 0

, δ(φ) = d

dφH(φ). (8)

Experiments

In this section, we evaluate the performance of our proposedapproach by comparing manually obtained boundaries. Theperformance of the proposed method was evaluated using10 different SPECT data sets. All images were obtainedwith e.cam single-head SPECT system manufactured bySIEMENS. Implementations were performed in MATLAB2011a [18] environment using an Intel Core i5, 2.4 GHz pro-cessor with 4 GB of memory and under Microsoft Windows7 operating system (64-bit).

As discussed in the previous section, first the initial closedcurve was performed. Then, it was employed for segmenta-tion of LV by variational level set technique. We used theresults of proposed method by applying connected compo-nent analysis to extract the contour of LV in image. Finalboundaries of LV after applying the variational level set aregiven in Figs. 2 and 3.

In order to validate the proposed method, manual seg-mentation of data sets were performed by three medical pro-fessionals. In the available data sets, segmented images ofexperts with a Dice similarity index [19] of above 80 % wereemployed as gold standard in the rest of experiments.

Evaluation of the algorithm was performed by computingfour parameters: sensitivity, specificity, accuracy, and mean

Fig. 2 Several mid-slice cardiac SPECT images from one patient, from top to bottom: original images, proposed (variational level set) results,connected component analysis results and LV areas

123

Author's personal copy

Int J CARS

Fig. 3 Mid-slice cardiac SPECT images from five patients, from topto bottom: original images, proposed (variational level set) results, con-nected component analysis results and LV areas

Fig. 4 A confusion matrix

error rate [20]. For this purpose, true positive (TP), falsepositive (FP), true negative (TN) and false negative (FN) val-ues were calculated. TP, TN, FP, and FN, are the four differentpossible outcomes of a single prediction for a two-class casewith classes “yes” (“cardiac tissues”) and “no” (“non-cardiactissues”). A false positive is when the outcome is incorrectlyclassified as “yes” (or “positive”), when it is in fact “no” (or“negative”). A false negative is when the outcome is incor-rectly classified as negative when it is in fact positive. Truepositives and true negatives are obviously correct classifica-tions. Keeping track of all these possible outcomes is suchan error-prone activity, that they are usually shown in what iscalled a confusion matrix [20]. Figure 4 shows a confusionmatrix.

Table 1 shows these parameters for the proposed methodin the presence of the available data sets.

Another parameter which is used for evaluation of theproposed algorithm is the receiver operating characteristic(ROC) curve [20], which is more accurate than the tradi-tional level set method. This is demonstrated in Fig. 5. Thevertical axis is associated with true positive rate (sensitiv-ity) and horizontal axis is associated with false positive rate(or 1-specificity). The area under the ROC curve (AUC) is areasonable performance statistic for classifier systems. Thevalue of AUC for proposed method is equal to 0.98.

Table 1 Performance measure of proposed method

Measure Definition Proposed method (%)

Sensitivity TP/(TP+FN) 88.94

Specificity TN/(TN+FP) 96.81

Precision TP/(TP+FP) 94.79

Accuracy (TP+TN)/(TP+TN+FP+FN) 93.55

Mean error rate (FP+FN)/(TP+TN+FP+FN) 5.44

Fig. 5 The ROC curves of the proposed method and traditional levelset method

Discussion

As stated earlier, the goal of this research was to performa consistent and feasible contour for LV in cardiac SPECTimages. We propose a simple automatic initialization methodspecifically designed for our cardiac SPECT application.

We employed an adaptive thresholding technique andmorphological operation to estimate the initial LV boundary.Since cardiac SPECT images are not in the same intensityrange, applying conventional techniques such as threshold-ing was not effective to produce acceptable results. More-over, selecting several threshold values was tedious. In turn,the adaptive thresholding technique which is based on Otsu’salgorithm, operates extremely faster and accurate than con-ventional thresholding technique. In terms of providing aninitial closed curve for LV, the proposed techniques presentedacceptable results in the presence of our local data sets.

The variational level set, that is an energy optimizationtechnique, is a region-based active contour model based onglobal image information. Also, it have several advantagesover edge-based active contour models as well as traditionallevel set. Firstly, it does not utilize the image gradient andtherefore has better performance for the images which sufferfrom weak boundaries. Secondly, they are significantly lesssensitive to the location of initial closed curves. Thirdly, it is

123

Author's personal copy

Int J CARS

Fig. 6 A typical mid-slice cardiac SPECT image from one patient,from left to right: original image, proposed (variational level set) result,traditional (level set) result

Fig. 7 Intestinal activity is diagnosed in right lower quadrant of LVmyocardium and markedly distorted LV myocardium is visualizedwhich is due to hypertrophied RV myocardium are mistakenly includedin ROI region of LV by semiquantitation traditional level set method(2nd row). They are accurately recognized and excluded during myo-cardial segmentation by proposed method (3rd row)

very robust to noise. In this regard, the variational level setwas successful to produce consistent and continuous LV con-tours. Final boundaries of a typical LV image after applyingthe level set and variational level set are illustrated in Fig. 6.

In cardiac perfusion study, myocardial border could bedistorted by nearby interfering activities. This phenomenoncould be due to hypertrophied right ventricle myocardium.Retention of tracer in sub diaphragmatic area such as liv-ers, intestine or stomach which could interfere specially withinferior wall of LV myocardium. Retained activity in thelung(s) or breast are the other pathologic issues above thediaphragm which could affect the LV segmentation. They arerecognized well and separated from LV myocardium by pro-posed method. Final boundaries of LV image after applyingtraditional level set method and the level set and variationallevel set are shown in Fig. 7.

As inferred from the experimental results, the proposedmethod demonstrates reasonable robustness against twomajor difficulties in SPECT image processing, i.e., noise andlow level details in an automatic manner.

Conclusion

Heart diseases are amongst major death factors worldwide.A cardiologist needs to know about LV volume and ejection

fraction in the MPS. Extracting the LV borders is a crucialand challenging task in nuclear medicine imaging analysiswhich is required in MPS procedure.

In this research, an automatic method for LV segmentationin cardiac SPECT images by variational level set techniqueis proposed. In this case, firs of all we introduced a hybridinitialization method to define the initial closed curve for car-diac SPECT application, then it was applied to estimate thefinal contour by variational level set.

In our validation methodology, we compared the resultsfrom the implemented methods with the manual estimation ofLV contours which was performed by the experts. Our auto-matic segmentation method correctly segmented the LV withall of the present data sets to which we have access. Also, thismethod is robust for noisy and low-resolution data speciallyto overcome the problematic cases with nearby interferingactivities including hypertrophied right ventricle, hepatic orintestinal activity and pulmonary or intramammary activity.The results show that the variational level set algorithm per-forms better than the traditional level set model for automatedsegmentation of LV in SPECT images.

The plans for future work which suggested by this paperincluded investigation of proposed method performance onmore complex data and different modalities. Future workcan also be done for combining the information of anatomi-cal knowledge with cardiac image to reach a higher perfor-mance.

Acknowledgments The authors would like to thank division ofnuclear medicine imaging at Shohada-e-Tajrish hospital for their helpin providing the data set and their valuable comments as well, especiallyDr. Elahe Pirayesh.

References

1. Reyes E, Underwood SR (2006) Myocardial perfusion scintigra-phy. An important step between clinical assessment and coronaryangiography in patients with stable chest pain. Eur Heart J 27:3–4

2. Grosgeorge D, Petitjean C, Caudron J, Fares J, Nicolas DacherJ (2011) Automatic cardiac ventricle segmentation in MR images:a validation study. Int J Comput Assist Radiol Surg 6(5):573–581

3. Marsousi M, Eftekhari A, Kocharian A, Alirezaie J (2010) Endo-cardial boundary extraction in left ventricular echocardiographicimages using fast and adaptive B-spline snake algorithm. Int J Com-put Assist Radiol Surg 5(5):501–513

4. Debreuve E, Marlaud M, Aubert G, Darcourt J (2001) Space timesegmentation using level set active contours applied to myocardialgated SPECT. IEEE Med Imaging 20(7):643–659

5. Paragios N, Mellina-Gottardo O, Ramesh V (2004) Gradient vec-tor flow fast geodesic active contours. IEEE PAMI 26(3):402–407

6. Yanga R, Mirmehdia M, Hallb D (2006) A charged contour modelfor cardiac SPECT segmentation. Med Image Underst Anal 1:1–5

7. Dornheim L, Tönnies KD, Dixon Kat (2005) Automatic segmen-tation of the left ventricle in 3D SPECT data by registration witha dynamic anatomic model. Med Image Comput Comput AssistInterv(MICCAI) 3749:335–342

8. Warfield S, Dengler J, Zaers J, Guttmann C, Gil W, Ettinger G,Hiller J, Kikinis R (1995) Automatic identification of grey matter

123

Author's personal copy

Int J CARS

structures from MRI to improve the segmentation of white matterlesions. J Image Guid Surg 1(6):326–338

9. Otsu N (1979) A threshold selection method from gray-level his-tograms. IEEE Trans Syst Man Cyben 9(1):62–66

10. Gonzales RC, Woods RE (2002) Digital image processing, 2ndedn. Prentice Hall, New Jersey

11. Gonzales RC, Woods RE, Eddins SL (2003) Digital image pro-cessing using MATLAB, 1st edn. Prentice Hall, New Jersey

12. Kass M, Witkin A, Terozopolous D (1988) Snakes: active contourmodels. Int J Comput Vis 1(4):321–331

13. Delingette H, Montagnat J (2001) Shape and topology constraintson parametric active contours. Comput Vis Image Underst 83:140–171

14. Yezzi A, Kichenassamy S, Kumar A, Olver P, Tannenbaum A(1997) A geometric snake model for segmentation of medicalimagery. IEEE Trans Med Imag 16:199–209

15. Osher S, Sethian JA (1988) Fronts propagating with curvaturedependent speed: algorithms based on Hamilton-Jacobi formula-tions. J Comput Phys 79:12–49

16. Zhao HK, Chan TF, Merriman B, Osher S (1996) A varia-tional level set approach to multiphase motion. J Comput Phys127(1):79–195

17. Vese L, Chan T (2002) A multiphase level set framework for imagesegmentation using the Mumford and Shah model. Int J ComputVision 50(3):271–293

18. http://www.mathworks.com19. Dice LR (1945) Measures of the amount of ecologic association

between species. Ecology 26(3):297–30220. Fawcett T (2006) An introduction to ROC analysis. Pattern Rec-

ognt Lett 27(8):861–874

123

Author's personal copy