1College of Electronic Science and Engineering Jilin University Changchun 130012 China2Division of Research and Innovations Carestream Health Inc Rochester NY 14615 USA3Affiliated Hospital of the Changchun University of Chinese Medicine Changchun 130021 China4Collaborative Innovation Center of High-End Manufacturing Equipment Xirsquoan Jiaotong University Xirsquoan 710054 China
Correspondence should be addressed to Bingheng Lu bhluxjtueducn and Jing Wang wjwjggggmailcom
Received 3 June 2016 Revised 23 August 2016 Accepted 15 December 2016 Published 7 February 2017
Academic Editor Dariusz Mrozek
Copyright copy 2017 Xiaohua Qian et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Ventricle segmentation is a challenging technique for the development of detection system of ischemic stroke in computedtomography (CT) as ischemic stroke regions are adjacent to the brain ventricle with similar intensity To address this problem wedeveloped an objective segmentation system of brain ventricle in CTThe intensity distribution of the ventricle was estimated basedon clustering technique connectivity and domain knowledge and the initial ventricle segmentation results were then obtained Toexclude the stroke regions from initial segmentation a combined segmentation strategy was proposed which is composed of threedifferent schemes (1) the largest three-dimensional (3D) connected component was considered as the ventricular region (2) thebig stroke areas were removed by the image difference methods based on searching optimal threshold values (3) the small strokeregionswere excluded by the adaptive template algorithmTheproposedmethodwas evaluated on 50 cases of patients with ischemicstroke The mean Dice sensitivity specificity and root mean squared error were 09447 0969 0998 and 0219mm respectivelyThis system can offer a desirable performanceTherefore the proposed system is expected to bring insights into clinic research andthe development of detection system of ischemic stroke in CT
1 Introduction
Computed tomography (CT) is generally used to assesspatients with acute ischemic stroke in America because of itsfaster speed the better contrast of bones and blood and thelower cost thanmagnetic resonance images (MRI)The ische-mic stroke and cerebrospinal fluid (CSF) regions have asimilar appearance inCT images thus accurate ventricle seg-mentation can significantly facilitate ischemic stroke regionlocalization and is an indispensable step for the develop-ment of computer-aided detection (CAD) for acute ischemicstroke
Several state-of-the-art methods have been proposed tosegment ventricles inMRI [1] including active contour-basedmethods [2ndash4] fuzzy schemes [5 6] and probabilitymethods[7 8] However these methods may be inappropriate to workon CT images since there are lower contrast higher noiselevel and larger slice thickness in brain CT images
Only little literature on the segmentation of brain CTimages has been published For example Wei et al proposeda segmentation scheme based on 2D Otsu thresholdingapproach [9] Lee et al applied the 119896-means and expectationmaximization clustering to segment CT images [10] Anothermethod by Chen et al was based on a Gaussian mixturemodel [11] Gupta et al integrated the adaptive thresholdconnectivity and domain knowledge to classify the cere-brospinal fluid white matter and gray matter on CT images[12] These methods mentioned above were not designedspecifically for ventricle segmentation and were not validatedon the images with severe abnormalities Chen et al devel-oped a ventricular segmentation system by combining low-level segmentation and high-level template matching [13]Similarity Liu et al proposed a model-guided segmentationfor ventricle region [14] The two methods are both basedon the template or model scheme for ventricle extraction inCT Since these templates were yielded from the MRI brain
HindawiBioMed Research InternationalVolume 2017 Article ID 8690892 11 pageshttpsdoiorg10115520178690892
2 BioMed Research International
image and registrationwas linear the templates only provideda rough mask for the ventricle segmentation Therefore itis still challenging for these two methods to exclude strokeregions from segmentation results Qian et al proposed alevel set model to segment CSF but the result includes thestroke regions [15] This study will improve the methods andextensively validate our previous work [16]
The significant difficulty of the accurate ventricle seg-mentation is to deal with CT images of patients withischemic stroke Some of the stroke regions and ventriclesare connected and have similar intensities To address thischallenge we developed an objective segmentation strategyof brain ventricles in unenhanced CT with ischemic strokeWe applied the following three schemes to exclude the strokeregions from segmentation results
(1) We took the largest three-dimensional (3D) con-nected component in a preliminary segmentation asthe ventricular region removing the lesion or otherregions without the 3D connectivity relationship withthe ventricle since the initial segmentation resultcontained not only the ventricle but also some non-ventricular regions such as lesion or CSF
(2) The large stroke regions were removed by the imagedifferencemethodThe large stroke areas tend to closethe brain edge and their intensities were generallylower than that of the main parts of ventricles Thusthe stroke region can be extracted by the differ-ence between segmentation results from two optimalthreshold values
(3) The small stroke regions were removed by the adap-tive template algorithm The adaptive template wasdirectly generated from the corresponding imageitself based on the big intensity difference between themain part of the ventricle and the brain parenchymaThis template did not contain the whole ventricle butdid cover the main part of the ventricular regionThus we applied this template to remove the smalllesions around the main part of the ventricle whichwas not subjected to the registration Another effectwas that the exclusion of these small lesions mightbreak the connectivity relationship between the lesionregions and the ventricular region in 3D space
2 Materials and Methods
As shown in Figure 1 the automated ventricle segmentationmethod is comprised of two phases that is alignment phase(Section 22) and segmentation phase (Section 23) In thealignment phase the light curvessegment of the brain wasdetected to determine the midsagittal line for each slice Wethen aligned the midsagittal line (MSL) with the verticalline of each slice to achieve brain alignment In the segmen-tation phase we first estimated the intensity range of theventricle region based on clustering technique connectivityand domain knowledge An image difference algorithm wasdeveloped to identify and remove the large stroke regions inthe initial segmentation The remaining small stroke regionwas further excluded by an adaptive template of the ventricle
Finally the largest 3D connectivity of the segmented ventriclewas employed to refine the segmentation result
21 Dataset We tested the proposed method on 50 CT scansof patients with ischemic stroke in this study This datasetwas collected from Jilin University Medical Center using CTscanners (Light Speed 16 GE Medical System) with an X-raytube voltage of 120 kVp Each patient has 14 slices with thethickness of 5mm in this study The matrix size of each sliceis 512 times 512 pixels and the pixel size is 0426mm with a 16-bit gray levelThe 50 patients were composed of 29 males and21 females and their average age is 57 years with the rangebetween 41 years and 76 years We established a referencestandard of ventricle for evaluation of segmentation result Amedical physicist (XQ eight years of experience) manuallydelineated the ventricle boundaries for all the slices on anLCD screen as the reference standard to assess the accuracyof segmentation results
22 Alignment of the Brain Image Prior to the alignment ofbrain image the skull was stripped by a threshold methodsince CT number of bone tissues are consistently higher thanbrain tissues Generally the CT number of soft tissue is lessthan 60 Hounsfield units (HU) (such as 1ndash12HU of ventricle25ndash38HU of white matter and 35ndash60HU of gray matter)while average CT intensity is 1000HU for bones Thus weextracted the skull using a fixed threshold of 100HU Theregion inside the skull was considered as brain region and theregion outside the skull served as background
After the extraction of the brain the inclination angle andposition were corrected by aligning MSL with the verticalcenterline of each slice The determination of MSL is a keystep in this alignment Since the falx cerebri (ie narrow lightcurvesegment) presents on about 30 images we applied thefalx cerebri as a reference to identify the MSL Therefore weutilized two steps to achieve alignment of the brain including(1) detection of a light curve in the brain and (2) affinetransformation based on MSL
221 Detection of Light Curves in Brain Figure 2 shows theschematic diagram of light curve detection To acceleratethe detection we defined a rectangle region of interest(ROI) whose size was chosen to include the light curvesto be detected We selected a smallest minimum boundingrectangle of the brain area in the whole scan and then definedthe half width of this rectangle as the width of the ROIThe height of the ROI was taken the default value of 512Figure 2(b) shows the rectangle ROI of the brain
CT brain image has a high level of noise The commonfiltering may blur the weak edge making detection of thelight curve difficult The light curve has a slight angle withthe vertical direction however it is still regarded as verticalThus we designed a one-dimensional (7 times 1) Gaussian filterwith the variance of 2 to smooth the image along the verticaldirection which can preserve the edge information of thelight curve in the horizontal direction as shown in Figure 2(c)
We then design a horizontal Laplacian detection maskthat is [05 0 1 0 05] to detect the light curve since thevertical strip included more edge points of the light curve
BioMed Research International 3
Light curve detection Affine transformation
Alignment
Estimation for intensity range of ventricle
Preliminarysegmentation
Large stroke region removed by image difference method
Small stroke region excluded by the adaptive template matching
Maximum 3D connected region as
the ventricle
Segmentation
Figure 1 Schematic framework for segmentation of the brain ventricle in CT of patients with ischemic stroke
than other places With the Laplacian image (Figure 2(c))we employed an adaptive threshold to yield an edge map asshown in Figure 2(d) We empirically set the threshold as theaverage value with 25 multiple of the standard deviation ofthe Laplacian image
After that we erased the small unconnected noise pointclouds in the edge map based on 3D connectivity The noisepoints in edge map may negatively affect the subsequent3D fitting of the middle sagittal plane However the 3Dconnected volume of these noise points is small thus wecan remove them with a threshold in 3D connected volumeIn our experiment we applied thirty pixels as the thresholdto obtain the clean edge map of light curve (Figure 2(f))Figure 2(g) shows the 3D edge map of light curves
222 Affine Transformation Based on MSL To obtain theprecise MSL we first fitted a middle sagittal plane in 3DEuclidean space through a set of edge segments of lightcurves using least-squares fitting approach Let (119909119894 119910119894 119911119894) bea point of edge segments which has totally 119872 points and
119894 = 1 2 119872 So the optimum fitting plane can be achievedby the following formulation as
(119911119894 minus 119886119909119894 minus 119887119910119894 + 119888)2 (1)
TheMSL of each slice was defined as the intersection linebetween the image and middle sagittal plane Let 119911119894 denotethe 119894th slice of 3D image and we can obtain the MSL of thisslice as
119886119909 minus 119887119910 = 119911119894 minus 119888 (2)
The determined MSL was shown in Figure 3(a) Finallywe aligned the MSL of the brain with the vertical center lineof a slice using the affine transformation defined by
1199091015840 = (119909 minus 1199090) cos 120579 + (119910 minus 1199100) sin 120579 + 11990901199101015840 = (119909 minus 1199090) sin 120579 + (119910 minus 1199100) cos 120579 + 1199100 (3)
where (1199090 1199100) is the center point of the vertical center lineof a slice and 120579 is the inclination angle between the MSL
4 BioMed Research International
ROI extraction
Vertical filtering
3D display of light curve
8040100
150200
250300
350400
450
48
12
Denoising
(a) (b) (c) (d) (e) (f)
Horizontal Laplacian detection
Light curvedetection
(g)
150200
250300
350
y
z
x
Figure 2 Diagram of light curvesegment detection (a) original image without skull (b) the ROI of the light curve (c) the vertical filteredROI (d) the Laplacian image (e) the detected light curve (f) denoising light curve (g) 3D display of light curves 119911-axis represents the slicenumber 119909- and 119910-axes denote the pixel number
(a) (b) (c)
Figure 3 Alignment of the brain image (a) original image with the midsagittal line (MSL dashed line) (b) the vertical center line of a slicewith white color and the MSL (c) aligned brain image
and vertical center line Figures 3(b) and 3(c) show that theinclination angle and position of the brain were corrected
23 Segmentation of the Ventricle In the phase of ventriclesegmentation we focused on excluding the stroke area in theventricle segmentation result The flowchart was shown inFigure 4
231 Parameter Estimation for the Ventricle Prior to thesegmentation of ventricle we estimated parameters of theintensity distribution of the ventricle We first applied the119870-means algorithm (119870 = 2) on the 3D images for stratificationof the brain image and took the largest 3D connectedcomponent of low-intensity category as the ventricle Thenan estimation method based on connectivity and domain
knowledge from the literature [8] was utilized to computethe intensity distribution of different tissues Specifically wetracked the slop of the histogram corresponding to the 3Dlargest connected component in rough intensity range ofventricle to determine a critical intensity which serves asan initial classifier of cerebral spinal fluid and white matterThresholds of cerebral spinal fluid white matter and graymatter are optimally derived to minimize spatial overlaperrors in different tissue types In this study ventricularintensity range of [119881min 119881max] will be adopted to extract theventricular region
232 Preliminary Segmentation for the Ventricle Based onEstimated Parameters 119881max the estimatedmaximum of ven-tricular intensity range was applied as a threshold value for
BioMed Research International 5
(a)
(b)
(c)
(g)
(d)
(e)
(f)
(h)
Ventricle region withoutbig stroke regions
Extraction of the big stroke region by image
difference approach
Determination of the 2ndcritical segmentation
without ldquostroke regionrdquo
Preliminary segmentation
Checking of stroke region
Small stroke regionexcluded by the adaptive
template matching
Maximum 3D connected region as the ventricle
Yes
No
Figure 4 Flowchart of the exclusion of stroke area in the ventricular segmentation result
preliminary segmentation of the ventricle If the intensityrange of the stroke is greater than 119881max the preliminarysegmentation is a good result Whereas if the intensity rangeof the stroke is less than119881max the segmentation result may beunacceptable since it may also contain some stroke regions
Then we utilized the 3D connectivity of the preliminarysegmentation result to obtain the largest volume as the initialsegmentation of the ventricle The stroke regions or noiseareas without the 3D connectivity to the ventricle could beexcluded by this step Figure 4(b) shows that the large strokeregions are connected to the ventricle in the segmentation
233 Detection of the Big Stroke Regions Since big strokeregions are mainly related to the anterior cerebral artery ormiddle cerebral artery these stroke regions are mostly closedto the brain edge Thus we proposed a brain edge checkingalgorithm to determinewhether the big stroke regions exist in
the segmentation result An annular region of the brain edgewas defined to detect the objects Assumed that theminimumside length of the minimum bounding rectangle of the brainwas 119871min the width of the annular region could be calculatedby 015 times 119871min to avoid some parts of the ventricle fallingwithin the annular regionThemask of the brain edge annularregion was shown in Figure 4(c) Thus if the objective areawas greater than the threshold we labeled it as the strokeregion The threshold was empirically selected as 20 pixels toallow the presence of noise
234 Determination of the Big Stroke Regions We proposedan image difference technique based on the heuristic search-ing algorithm to extract the big stroke regions which weresuccessfully detected in the preliminary segmentation bythe edge checking method This image difference techniqueessentially applied the difference between two segmentation
6 BioMed Research International
results by different threshold values for determining thestroke regions We first defined the critical threshold value(ie119879critical) If a threshold was greater than119879critical the strokeregions in the segmentation result of this threshold couldbe detected by the edge analysis method whereas if thethreshold was smaller or equal to 119879critical none stroke regioncould be detected We then obtained the stroke regions by
PA asymp 119891 (119881max) minus 119892 (119891 (119879critical)) (4)
where 119891(lowast) was the threshold method 119892(lowast) represented thesubsequent refine algorithms such as morphology methodand PA represented the stroke regions So we obtained theventricle areas
119891 (119881max) minus 119892 (PA) (5)
The vital step in the image difference method is todetermine the critical threshold value 119879critical We applied thegold searching method and the edge checking method toobtain the 119879critical in range [119881min 119881max]235 Exclusion of the Small Stroke Regions Some smallstroke regions may still present in the segmentation resultfrom the image difference approach To address this problemwe developed an adaptive template matching approachwhich applied the mask of the main part of the ventricleto exclude the remaining small stroke regions The templatewas generated from each image It did not contain the wholeventricle but covers the main part of the ventricle
Figure 5 shows a sectional view of the gray-scale map fora brain image The intensity difference between the ventricleand brain parenchyma was around 20 intensity values whilethe transition area was only 6 to 7 pixels Thus we applied119881min as a threshold for ventricle segmentation and took the3D largest connected region as the ventricle as shown inFigure 6(b) The ventricle segmentation merely containingthe right and left lateral ventricles and without the 3rd and4th ventricle was adaptively selected as the templates Toensure that the template covers the ventricle we conductedsomemorphological analysis including closed operation andexpansion operation The generated template was shown inFigures 6(c) and 6(d)
After these steps we linearly registered the templatewith the corresponding segmentationThe objects within thetemplate served as the ventricle so that the remaining smallstroke areas could be excluded from the segmentation results
236 Refinement of the Ventricular Segmentation Weemployed connected component labeling to the segmentedventricle regionThe largest volume served as the ventricularWe then removed the calcification regions in the results andsmoothed the ventricular edges using the morphologicallyclosed operation
24 Evaluation of the Segmentation Method We applied fourmeasures including Dice metric (Dice) root mean squarederror (RMSE) reliability (R)28 and correlation coefficient(119877) to assess the performance of the proposed segmentationmethod The four measures are defined as follows
Location of the column in image
Inte
nsity
val
ue
6 pixels
10005
10
1520
25
30
35
40
45
50
150 200 250 300 350 400
6 pixels
Figure 5 A sectional view of the gray-scale map for brain image
(1) Dice Metric Let 119881119904 represent the automatically segmentedvolume and 119881119903 represent the manual segmentation (iereference standard) The Dice is defined as
Dice = 2119881119904 cap 119881119903119881119904 + 119881119903 (6)
The value of Dice is between 0 and 1 Higher Dice indicatesbetter overlap between segmented volumes and the referencestandard
(2) Root Mean Squared Error The RMSE calculates the dis-tance between the corresponding points on the automaticallysegmented and reference boundaries defined by
RMSE = ( 1119873119873sum119894=1
(119909119904119894 minus 119909119903119894)2 + (119910119904119894 minus 119910119903119894)2)12
(7)
where (119909119904119894 119910119904119894) is a point on the segmented boundary and(119909119903119894 119910119903119894) is the closest point to (119909119904119894 119910119904119894) on the referenceboundary The lower RMSE the better performance
(3) Reliability The reliability function is used to assess thereliability of segmentation method defined as
R (119889)= Number of volumes segmented with Dice gt 119889
Total number of volumes (8)
BioMed Research International 7
(a) (b)
(c) (d)
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
image and registrationwas linear the templates only provideda rough mask for the ventricle segmentation Therefore itis still challenging for these two methods to exclude strokeregions from segmentation results Qian et al proposed alevel set model to segment CSF but the result includes thestroke regions [15] This study will improve the methods andextensively validate our previous work [16]
The significant difficulty of the accurate ventricle seg-mentation is to deal with CT images of patients withischemic stroke Some of the stroke regions and ventriclesare connected and have similar intensities To address thischallenge we developed an objective segmentation strategyof brain ventricles in unenhanced CT with ischemic strokeWe applied the following three schemes to exclude the strokeregions from segmentation results
(1) We took the largest three-dimensional (3D) con-nected component in a preliminary segmentation asthe ventricular region removing the lesion or otherregions without the 3D connectivity relationship withthe ventricle since the initial segmentation resultcontained not only the ventricle but also some non-ventricular regions such as lesion or CSF
(2) The large stroke regions were removed by the imagedifferencemethodThe large stroke areas tend to closethe brain edge and their intensities were generallylower than that of the main parts of ventricles Thusthe stroke region can be extracted by the differ-ence between segmentation results from two optimalthreshold values
(3) The small stroke regions were removed by the adap-tive template algorithm The adaptive template wasdirectly generated from the corresponding imageitself based on the big intensity difference between themain part of the ventricle and the brain parenchymaThis template did not contain the whole ventricle butdid cover the main part of the ventricular regionThus we applied this template to remove the smalllesions around the main part of the ventricle whichwas not subjected to the registration Another effectwas that the exclusion of these small lesions mightbreak the connectivity relationship between the lesionregions and the ventricular region in 3D space
2 Materials and Methods
As shown in Figure 1 the automated ventricle segmentationmethod is comprised of two phases that is alignment phase(Section 22) and segmentation phase (Section 23) In thealignment phase the light curvessegment of the brain wasdetected to determine the midsagittal line for each slice Wethen aligned the midsagittal line (MSL) with the verticalline of each slice to achieve brain alignment In the segmen-tation phase we first estimated the intensity range of theventricle region based on clustering technique connectivityand domain knowledge An image difference algorithm wasdeveloped to identify and remove the large stroke regions inthe initial segmentation The remaining small stroke regionwas further excluded by an adaptive template of the ventricle
Finally the largest 3D connectivity of the segmented ventriclewas employed to refine the segmentation result
21 Dataset We tested the proposed method on 50 CT scansof patients with ischemic stroke in this study This datasetwas collected from Jilin University Medical Center using CTscanners (Light Speed 16 GE Medical System) with an X-raytube voltage of 120 kVp Each patient has 14 slices with thethickness of 5mm in this study The matrix size of each sliceis 512 times 512 pixels and the pixel size is 0426mm with a 16-bit gray levelThe 50 patients were composed of 29 males and21 females and their average age is 57 years with the rangebetween 41 years and 76 years We established a referencestandard of ventricle for evaluation of segmentation result Amedical physicist (XQ eight years of experience) manuallydelineated the ventricle boundaries for all the slices on anLCD screen as the reference standard to assess the accuracyof segmentation results
22 Alignment of the Brain Image Prior to the alignment ofbrain image the skull was stripped by a threshold methodsince CT number of bone tissues are consistently higher thanbrain tissues Generally the CT number of soft tissue is lessthan 60 Hounsfield units (HU) (such as 1ndash12HU of ventricle25ndash38HU of white matter and 35ndash60HU of gray matter)while average CT intensity is 1000HU for bones Thus weextracted the skull using a fixed threshold of 100HU Theregion inside the skull was considered as brain region and theregion outside the skull served as background
After the extraction of the brain the inclination angle andposition were corrected by aligning MSL with the verticalcenterline of each slice The determination of MSL is a keystep in this alignment Since the falx cerebri (ie narrow lightcurvesegment) presents on about 30 images we applied thefalx cerebri as a reference to identify the MSL Therefore weutilized two steps to achieve alignment of the brain including(1) detection of a light curve in the brain and (2) affinetransformation based on MSL
221 Detection of Light Curves in Brain Figure 2 shows theschematic diagram of light curve detection To acceleratethe detection we defined a rectangle region of interest(ROI) whose size was chosen to include the light curvesto be detected We selected a smallest minimum boundingrectangle of the brain area in the whole scan and then definedthe half width of this rectangle as the width of the ROIThe height of the ROI was taken the default value of 512Figure 2(b) shows the rectangle ROI of the brain
CT brain image has a high level of noise The commonfiltering may blur the weak edge making detection of thelight curve difficult The light curve has a slight angle withthe vertical direction however it is still regarded as verticalThus we designed a one-dimensional (7 times 1) Gaussian filterwith the variance of 2 to smooth the image along the verticaldirection which can preserve the edge information of thelight curve in the horizontal direction as shown in Figure 2(c)
We then design a horizontal Laplacian detection maskthat is [05 0 1 0 05] to detect the light curve since thevertical strip included more edge points of the light curve
BioMed Research International 3
Light curve detection Affine transformation
Alignment
Estimation for intensity range of ventricle
Preliminarysegmentation
Large stroke region removed by image difference method
Small stroke region excluded by the adaptive template matching
Maximum 3D connected region as
the ventricle
Segmentation
Figure 1 Schematic framework for segmentation of the brain ventricle in CT of patients with ischemic stroke
than other places With the Laplacian image (Figure 2(c))we employed an adaptive threshold to yield an edge map asshown in Figure 2(d) We empirically set the threshold as theaverage value with 25 multiple of the standard deviation ofthe Laplacian image
After that we erased the small unconnected noise pointclouds in the edge map based on 3D connectivity The noisepoints in edge map may negatively affect the subsequent3D fitting of the middle sagittal plane However the 3Dconnected volume of these noise points is small thus wecan remove them with a threshold in 3D connected volumeIn our experiment we applied thirty pixels as the thresholdto obtain the clean edge map of light curve (Figure 2(f))Figure 2(g) shows the 3D edge map of light curves
222 Affine Transformation Based on MSL To obtain theprecise MSL we first fitted a middle sagittal plane in 3DEuclidean space through a set of edge segments of lightcurves using least-squares fitting approach Let (119909119894 119910119894 119911119894) bea point of edge segments which has totally 119872 points and
119894 = 1 2 119872 So the optimum fitting plane can be achievedby the following formulation as
(119911119894 minus 119886119909119894 minus 119887119910119894 + 119888)2 (1)
TheMSL of each slice was defined as the intersection linebetween the image and middle sagittal plane Let 119911119894 denotethe 119894th slice of 3D image and we can obtain the MSL of thisslice as
119886119909 minus 119887119910 = 119911119894 minus 119888 (2)
The determined MSL was shown in Figure 3(a) Finallywe aligned the MSL of the brain with the vertical center lineof a slice using the affine transformation defined by
1199091015840 = (119909 minus 1199090) cos 120579 + (119910 minus 1199100) sin 120579 + 11990901199101015840 = (119909 minus 1199090) sin 120579 + (119910 minus 1199100) cos 120579 + 1199100 (3)
where (1199090 1199100) is the center point of the vertical center lineof a slice and 120579 is the inclination angle between the MSL
4 BioMed Research International
ROI extraction
Vertical filtering
3D display of light curve
8040100
150200
250300
350400
450
48
12
Denoising
(a) (b) (c) (d) (e) (f)
Horizontal Laplacian detection
Light curvedetection
(g)
150200
250300
350
y
z
x
Figure 2 Diagram of light curvesegment detection (a) original image without skull (b) the ROI of the light curve (c) the vertical filteredROI (d) the Laplacian image (e) the detected light curve (f) denoising light curve (g) 3D display of light curves 119911-axis represents the slicenumber 119909- and 119910-axes denote the pixel number
(a) (b) (c)
Figure 3 Alignment of the brain image (a) original image with the midsagittal line (MSL dashed line) (b) the vertical center line of a slicewith white color and the MSL (c) aligned brain image
and vertical center line Figures 3(b) and 3(c) show that theinclination angle and position of the brain were corrected
23 Segmentation of the Ventricle In the phase of ventriclesegmentation we focused on excluding the stroke area in theventricle segmentation result The flowchart was shown inFigure 4
231 Parameter Estimation for the Ventricle Prior to thesegmentation of ventricle we estimated parameters of theintensity distribution of the ventricle We first applied the119870-means algorithm (119870 = 2) on the 3D images for stratificationof the brain image and took the largest 3D connectedcomponent of low-intensity category as the ventricle Thenan estimation method based on connectivity and domain
knowledge from the literature [8] was utilized to computethe intensity distribution of different tissues Specifically wetracked the slop of the histogram corresponding to the 3Dlargest connected component in rough intensity range ofventricle to determine a critical intensity which serves asan initial classifier of cerebral spinal fluid and white matterThresholds of cerebral spinal fluid white matter and graymatter are optimally derived to minimize spatial overlaperrors in different tissue types In this study ventricularintensity range of [119881min 119881max] will be adopted to extract theventricular region
232 Preliminary Segmentation for the Ventricle Based onEstimated Parameters 119881max the estimatedmaximum of ven-tricular intensity range was applied as a threshold value for
BioMed Research International 5
(a)
(b)
(c)
(g)
(d)
(e)
(f)
(h)
Ventricle region withoutbig stroke regions
Extraction of the big stroke region by image
difference approach
Determination of the 2ndcritical segmentation
without ldquostroke regionrdquo
Preliminary segmentation
Checking of stroke region
Small stroke regionexcluded by the adaptive
template matching
Maximum 3D connected region as the ventricle
Yes
No
Figure 4 Flowchart of the exclusion of stroke area in the ventricular segmentation result
preliminary segmentation of the ventricle If the intensityrange of the stroke is greater than 119881max the preliminarysegmentation is a good result Whereas if the intensity rangeof the stroke is less than119881max the segmentation result may beunacceptable since it may also contain some stroke regions
Then we utilized the 3D connectivity of the preliminarysegmentation result to obtain the largest volume as the initialsegmentation of the ventricle The stroke regions or noiseareas without the 3D connectivity to the ventricle could beexcluded by this step Figure 4(b) shows that the large strokeregions are connected to the ventricle in the segmentation
233 Detection of the Big Stroke Regions Since big strokeregions are mainly related to the anterior cerebral artery ormiddle cerebral artery these stroke regions are mostly closedto the brain edge Thus we proposed a brain edge checkingalgorithm to determinewhether the big stroke regions exist in
the segmentation result An annular region of the brain edgewas defined to detect the objects Assumed that theminimumside length of the minimum bounding rectangle of the brainwas 119871min the width of the annular region could be calculatedby 015 times 119871min to avoid some parts of the ventricle fallingwithin the annular regionThemask of the brain edge annularregion was shown in Figure 4(c) Thus if the objective areawas greater than the threshold we labeled it as the strokeregion The threshold was empirically selected as 20 pixels toallow the presence of noise
234 Determination of the Big Stroke Regions We proposedan image difference technique based on the heuristic search-ing algorithm to extract the big stroke regions which weresuccessfully detected in the preliminary segmentation bythe edge checking method This image difference techniqueessentially applied the difference between two segmentation
6 BioMed Research International
results by different threshold values for determining thestroke regions We first defined the critical threshold value(ie119879critical) If a threshold was greater than119879critical the strokeregions in the segmentation result of this threshold couldbe detected by the edge analysis method whereas if thethreshold was smaller or equal to 119879critical none stroke regioncould be detected We then obtained the stroke regions by
PA asymp 119891 (119881max) minus 119892 (119891 (119879critical)) (4)
where 119891(lowast) was the threshold method 119892(lowast) represented thesubsequent refine algorithms such as morphology methodand PA represented the stroke regions So we obtained theventricle areas
119891 (119881max) minus 119892 (PA) (5)
The vital step in the image difference method is todetermine the critical threshold value 119879critical We applied thegold searching method and the edge checking method toobtain the 119879critical in range [119881min 119881max]235 Exclusion of the Small Stroke Regions Some smallstroke regions may still present in the segmentation resultfrom the image difference approach To address this problemwe developed an adaptive template matching approachwhich applied the mask of the main part of the ventricleto exclude the remaining small stroke regions The templatewas generated from each image It did not contain the wholeventricle but covers the main part of the ventricle
Figure 5 shows a sectional view of the gray-scale map fora brain image The intensity difference between the ventricleand brain parenchyma was around 20 intensity values whilethe transition area was only 6 to 7 pixels Thus we applied119881min as a threshold for ventricle segmentation and took the3D largest connected region as the ventricle as shown inFigure 6(b) The ventricle segmentation merely containingthe right and left lateral ventricles and without the 3rd and4th ventricle was adaptively selected as the templates Toensure that the template covers the ventricle we conductedsomemorphological analysis including closed operation andexpansion operation The generated template was shown inFigures 6(c) and 6(d)
After these steps we linearly registered the templatewith the corresponding segmentationThe objects within thetemplate served as the ventricle so that the remaining smallstroke areas could be excluded from the segmentation results
236 Refinement of the Ventricular Segmentation Weemployed connected component labeling to the segmentedventricle regionThe largest volume served as the ventricularWe then removed the calcification regions in the results andsmoothed the ventricular edges using the morphologicallyclosed operation
24 Evaluation of the Segmentation Method We applied fourmeasures including Dice metric (Dice) root mean squarederror (RMSE) reliability (R)28 and correlation coefficient(119877) to assess the performance of the proposed segmentationmethod The four measures are defined as follows
Location of the column in image
Inte
nsity
val
ue
6 pixels
10005
10
1520
25
30
35
40
45
50
150 200 250 300 350 400
6 pixels
Figure 5 A sectional view of the gray-scale map for brain image
(1) Dice Metric Let 119881119904 represent the automatically segmentedvolume and 119881119903 represent the manual segmentation (iereference standard) The Dice is defined as
Dice = 2119881119904 cap 119881119903119881119904 + 119881119903 (6)
The value of Dice is between 0 and 1 Higher Dice indicatesbetter overlap between segmented volumes and the referencestandard
(2) Root Mean Squared Error The RMSE calculates the dis-tance between the corresponding points on the automaticallysegmented and reference boundaries defined by
RMSE = ( 1119873119873sum119894=1
(119909119904119894 minus 119909119903119894)2 + (119910119904119894 minus 119910119903119894)2)12
(7)
where (119909119904119894 119910119904119894) is a point on the segmented boundary and(119909119903119894 119910119903119894) is the closest point to (119909119904119894 119910119904119894) on the referenceboundary The lower RMSE the better performance
(3) Reliability The reliability function is used to assess thereliability of segmentation method defined as
R (119889)= Number of volumes segmented with Dice gt 119889
Total number of volumes (8)
BioMed Research International 7
(a) (b)
(c) (d)
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Large stroke region removed by image difference method
Small stroke region excluded by the adaptive template matching
Maximum 3D connected region as
the ventricle
Segmentation
Figure 1 Schematic framework for segmentation of the brain ventricle in CT of patients with ischemic stroke
than other places With the Laplacian image (Figure 2(c))we employed an adaptive threshold to yield an edge map asshown in Figure 2(d) We empirically set the threshold as theaverage value with 25 multiple of the standard deviation ofthe Laplacian image
After that we erased the small unconnected noise pointclouds in the edge map based on 3D connectivity The noisepoints in edge map may negatively affect the subsequent3D fitting of the middle sagittal plane However the 3Dconnected volume of these noise points is small thus wecan remove them with a threshold in 3D connected volumeIn our experiment we applied thirty pixels as the thresholdto obtain the clean edge map of light curve (Figure 2(f))Figure 2(g) shows the 3D edge map of light curves
222 Affine Transformation Based on MSL To obtain theprecise MSL we first fitted a middle sagittal plane in 3DEuclidean space through a set of edge segments of lightcurves using least-squares fitting approach Let (119909119894 119910119894 119911119894) bea point of edge segments which has totally 119872 points and
119894 = 1 2 119872 So the optimum fitting plane can be achievedby the following formulation as
(119911119894 minus 119886119909119894 minus 119887119910119894 + 119888)2 (1)
TheMSL of each slice was defined as the intersection linebetween the image and middle sagittal plane Let 119911119894 denotethe 119894th slice of 3D image and we can obtain the MSL of thisslice as
119886119909 minus 119887119910 = 119911119894 minus 119888 (2)
The determined MSL was shown in Figure 3(a) Finallywe aligned the MSL of the brain with the vertical center lineof a slice using the affine transformation defined by
1199091015840 = (119909 minus 1199090) cos 120579 + (119910 minus 1199100) sin 120579 + 11990901199101015840 = (119909 minus 1199090) sin 120579 + (119910 minus 1199100) cos 120579 + 1199100 (3)
where (1199090 1199100) is the center point of the vertical center lineof a slice and 120579 is the inclination angle between the MSL
4 BioMed Research International
ROI extraction
Vertical filtering
3D display of light curve
8040100
150200
250300
350400
450
48
12
Denoising
(a) (b) (c) (d) (e) (f)
Horizontal Laplacian detection
Light curvedetection
(g)
150200
250300
350
y
z
x
Figure 2 Diagram of light curvesegment detection (a) original image without skull (b) the ROI of the light curve (c) the vertical filteredROI (d) the Laplacian image (e) the detected light curve (f) denoising light curve (g) 3D display of light curves 119911-axis represents the slicenumber 119909- and 119910-axes denote the pixel number
(a) (b) (c)
Figure 3 Alignment of the brain image (a) original image with the midsagittal line (MSL dashed line) (b) the vertical center line of a slicewith white color and the MSL (c) aligned brain image
and vertical center line Figures 3(b) and 3(c) show that theinclination angle and position of the brain were corrected
23 Segmentation of the Ventricle In the phase of ventriclesegmentation we focused on excluding the stroke area in theventricle segmentation result The flowchart was shown inFigure 4
231 Parameter Estimation for the Ventricle Prior to thesegmentation of ventricle we estimated parameters of theintensity distribution of the ventricle We first applied the119870-means algorithm (119870 = 2) on the 3D images for stratificationof the brain image and took the largest 3D connectedcomponent of low-intensity category as the ventricle Thenan estimation method based on connectivity and domain
knowledge from the literature [8] was utilized to computethe intensity distribution of different tissues Specifically wetracked the slop of the histogram corresponding to the 3Dlargest connected component in rough intensity range ofventricle to determine a critical intensity which serves asan initial classifier of cerebral spinal fluid and white matterThresholds of cerebral spinal fluid white matter and graymatter are optimally derived to minimize spatial overlaperrors in different tissue types In this study ventricularintensity range of [119881min 119881max] will be adopted to extract theventricular region
232 Preliminary Segmentation for the Ventricle Based onEstimated Parameters 119881max the estimatedmaximum of ven-tricular intensity range was applied as a threshold value for
BioMed Research International 5
(a)
(b)
(c)
(g)
(d)
(e)
(f)
(h)
Ventricle region withoutbig stroke regions
Extraction of the big stroke region by image
difference approach
Determination of the 2ndcritical segmentation
without ldquostroke regionrdquo
Preliminary segmentation
Checking of stroke region
Small stroke regionexcluded by the adaptive
template matching
Maximum 3D connected region as the ventricle
Yes
No
Figure 4 Flowchart of the exclusion of stroke area in the ventricular segmentation result
preliminary segmentation of the ventricle If the intensityrange of the stroke is greater than 119881max the preliminarysegmentation is a good result Whereas if the intensity rangeof the stroke is less than119881max the segmentation result may beunacceptable since it may also contain some stroke regions
Then we utilized the 3D connectivity of the preliminarysegmentation result to obtain the largest volume as the initialsegmentation of the ventricle The stroke regions or noiseareas without the 3D connectivity to the ventricle could beexcluded by this step Figure 4(b) shows that the large strokeregions are connected to the ventricle in the segmentation
233 Detection of the Big Stroke Regions Since big strokeregions are mainly related to the anterior cerebral artery ormiddle cerebral artery these stroke regions are mostly closedto the brain edge Thus we proposed a brain edge checkingalgorithm to determinewhether the big stroke regions exist in
the segmentation result An annular region of the brain edgewas defined to detect the objects Assumed that theminimumside length of the minimum bounding rectangle of the brainwas 119871min the width of the annular region could be calculatedby 015 times 119871min to avoid some parts of the ventricle fallingwithin the annular regionThemask of the brain edge annularregion was shown in Figure 4(c) Thus if the objective areawas greater than the threshold we labeled it as the strokeregion The threshold was empirically selected as 20 pixels toallow the presence of noise
234 Determination of the Big Stroke Regions We proposedan image difference technique based on the heuristic search-ing algorithm to extract the big stroke regions which weresuccessfully detected in the preliminary segmentation bythe edge checking method This image difference techniqueessentially applied the difference between two segmentation
6 BioMed Research International
results by different threshold values for determining thestroke regions We first defined the critical threshold value(ie119879critical) If a threshold was greater than119879critical the strokeregions in the segmentation result of this threshold couldbe detected by the edge analysis method whereas if thethreshold was smaller or equal to 119879critical none stroke regioncould be detected We then obtained the stroke regions by
PA asymp 119891 (119881max) minus 119892 (119891 (119879critical)) (4)
where 119891(lowast) was the threshold method 119892(lowast) represented thesubsequent refine algorithms such as morphology methodand PA represented the stroke regions So we obtained theventricle areas
119891 (119881max) minus 119892 (PA) (5)
The vital step in the image difference method is todetermine the critical threshold value 119879critical We applied thegold searching method and the edge checking method toobtain the 119879critical in range [119881min 119881max]235 Exclusion of the Small Stroke Regions Some smallstroke regions may still present in the segmentation resultfrom the image difference approach To address this problemwe developed an adaptive template matching approachwhich applied the mask of the main part of the ventricleto exclude the remaining small stroke regions The templatewas generated from each image It did not contain the wholeventricle but covers the main part of the ventricle
Figure 5 shows a sectional view of the gray-scale map fora brain image The intensity difference between the ventricleand brain parenchyma was around 20 intensity values whilethe transition area was only 6 to 7 pixels Thus we applied119881min as a threshold for ventricle segmentation and took the3D largest connected region as the ventricle as shown inFigure 6(b) The ventricle segmentation merely containingthe right and left lateral ventricles and without the 3rd and4th ventricle was adaptively selected as the templates Toensure that the template covers the ventricle we conductedsomemorphological analysis including closed operation andexpansion operation The generated template was shown inFigures 6(c) and 6(d)
After these steps we linearly registered the templatewith the corresponding segmentationThe objects within thetemplate served as the ventricle so that the remaining smallstroke areas could be excluded from the segmentation results
236 Refinement of the Ventricular Segmentation Weemployed connected component labeling to the segmentedventricle regionThe largest volume served as the ventricularWe then removed the calcification regions in the results andsmoothed the ventricular edges using the morphologicallyclosed operation
24 Evaluation of the Segmentation Method We applied fourmeasures including Dice metric (Dice) root mean squarederror (RMSE) reliability (R)28 and correlation coefficient(119877) to assess the performance of the proposed segmentationmethod The four measures are defined as follows
Location of the column in image
Inte
nsity
val
ue
6 pixels
10005
10
1520
25
30
35
40
45
50
150 200 250 300 350 400
6 pixels
Figure 5 A sectional view of the gray-scale map for brain image
(1) Dice Metric Let 119881119904 represent the automatically segmentedvolume and 119881119903 represent the manual segmentation (iereference standard) The Dice is defined as
Dice = 2119881119904 cap 119881119903119881119904 + 119881119903 (6)
The value of Dice is between 0 and 1 Higher Dice indicatesbetter overlap between segmented volumes and the referencestandard
(2) Root Mean Squared Error The RMSE calculates the dis-tance between the corresponding points on the automaticallysegmented and reference boundaries defined by
RMSE = ( 1119873119873sum119894=1
(119909119904119894 minus 119909119903119894)2 + (119910119904119894 minus 119910119903119894)2)12
(7)
where (119909119904119894 119910119904119894) is a point on the segmented boundary and(119909119903119894 119910119903119894) is the closest point to (119909119904119894 119910119904119894) on the referenceboundary The lower RMSE the better performance
(3) Reliability The reliability function is used to assess thereliability of segmentation method defined as
R (119889)= Number of volumes segmented with Dice gt 119889
Total number of volumes (8)
BioMed Research International 7
(a) (b)
(c) (d)
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Figure 2 Diagram of light curvesegment detection (a) original image without skull (b) the ROI of the light curve (c) the vertical filteredROI (d) the Laplacian image (e) the detected light curve (f) denoising light curve (g) 3D display of light curves 119911-axis represents the slicenumber 119909- and 119910-axes denote the pixel number
(a) (b) (c)
Figure 3 Alignment of the brain image (a) original image with the midsagittal line (MSL dashed line) (b) the vertical center line of a slicewith white color and the MSL (c) aligned brain image
and vertical center line Figures 3(b) and 3(c) show that theinclination angle and position of the brain were corrected
23 Segmentation of the Ventricle In the phase of ventriclesegmentation we focused on excluding the stroke area in theventricle segmentation result The flowchart was shown inFigure 4
231 Parameter Estimation for the Ventricle Prior to thesegmentation of ventricle we estimated parameters of theintensity distribution of the ventricle We first applied the119870-means algorithm (119870 = 2) on the 3D images for stratificationof the brain image and took the largest 3D connectedcomponent of low-intensity category as the ventricle Thenan estimation method based on connectivity and domain
knowledge from the literature [8] was utilized to computethe intensity distribution of different tissues Specifically wetracked the slop of the histogram corresponding to the 3Dlargest connected component in rough intensity range ofventricle to determine a critical intensity which serves asan initial classifier of cerebral spinal fluid and white matterThresholds of cerebral spinal fluid white matter and graymatter are optimally derived to minimize spatial overlaperrors in different tissue types In this study ventricularintensity range of [119881min 119881max] will be adopted to extract theventricular region
232 Preliminary Segmentation for the Ventricle Based onEstimated Parameters 119881max the estimatedmaximum of ven-tricular intensity range was applied as a threshold value for
BioMed Research International 5
(a)
(b)
(c)
(g)
(d)
(e)
(f)
(h)
Ventricle region withoutbig stroke regions
Extraction of the big stroke region by image
difference approach
Determination of the 2ndcritical segmentation
without ldquostroke regionrdquo
Preliminary segmentation
Checking of stroke region
Small stroke regionexcluded by the adaptive
template matching
Maximum 3D connected region as the ventricle
Yes
No
Figure 4 Flowchart of the exclusion of stroke area in the ventricular segmentation result
preliminary segmentation of the ventricle If the intensityrange of the stroke is greater than 119881max the preliminarysegmentation is a good result Whereas if the intensity rangeof the stroke is less than119881max the segmentation result may beunacceptable since it may also contain some stroke regions
Then we utilized the 3D connectivity of the preliminarysegmentation result to obtain the largest volume as the initialsegmentation of the ventricle The stroke regions or noiseareas without the 3D connectivity to the ventricle could beexcluded by this step Figure 4(b) shows that the large strokeregions are connected to the ventricle in the segmentation
233 Detection of the Big Stroke Regions Since big strokeregions are mainly related to the anterior cerebral artery ormiddle cerebral artery these stroke regions are mostly closedto the brain edge Thus we proposed a brain edge checkingalgorithm to determinewhether the big stroke regions exist in
the segmentation result An annular region of the brain edgewas defined to detect the objects Assumed that theminimumside length of the minimum bounding rectangle of the brainwas 119871min the width of the annular region could be calculatedby 015 times 119871min to avoid some parts of the ventricle fallingwithin the annular regionThemask of the brain edge annularregion was shown in Figure 4(c) Thus if the objective areawas greater than the threshold we labeled it as the strokeregion The threshold was empirically selected as 20 pixels toallow the presence of noise
234 Determination of the Big Stroke Regions We proposedan image difference technique based on the heuristic search-ing algorithm to extract the big stroke regions which weresuccessfully detected in the preliminary segmentation bythe edge checking method This image difference techniqueessentially applied the difference between two segmentation
6 BioMed Research International
results by different threshold values for determining thestroke regions We first defined the critical threshold value(ie119879critical) If a threshold was greater than119879critical the strokeregions in the segmentation result of this threshold couldbe detected by the edge analysis method whereas if thethreshold was smaller or equal to 119879critical none stroke regioncould be detected We then obtained the stroke regions by
PA asymp 119891 (119881max) minus 119892 (119891 (119879critical)) (4)
where 119891(lowast) was the threshold method 119892(lowast) represented thesubsequent refine algorithms such as morphology methodand PA represented the stroke regions So we obtained theventricle areas
119891 (119881max) minus 119892 (PA) (5)
The vital step in the image difference method is todetermine the critical threshold value 119879critical We applied thegold searching method and the edge checking method toobtain the 119879critical in range [119881min 119881max]235 Exclusion of the Small Stroke Regions Some smallstroke regions may still present in the segmentation resultfrom the image difference approach To address this problemwe developed an adaptive template matching approachwhich applied the mask of the main part of the ventricleto exclude the remaining small stroke regions The templatewas generated from each image It did not contain the wholeventricle but covers the main part of the ventricle
Figure 5 shows a sectional view of the gray-scale map fora brain image The intensity difference between the ventricleand brain parenchyma was around 20 intensity values whilethe transition area was only 6 to 7 pixels Thus we applied119881min as a threshold for ventricle segmentation and took the3D largest connected region as the ventricle as shown inFigure 6(b) The ventricle segmentation merely containingthe right and left lateral ventricles and without the 3rd and4th ventricle was adaptively selected as the templates Toensure that the template covers the ventricle we conductedsomemorphological analysis including closed operation andexpansion operation The generated template was shown inFigures 6(c) and 6(d)
After these steps we linearly registered the templatewith the corresponding segmentationThe objects within thetemplate served as the ventricle so that the remaining smallstroke areas could be excluded from the segmentation results
236 Refinement of the Ventricular Segmentation Weemployed connected component labeling to the segmentedventricle regionThe largest volume served as the ventricularWe then removed the calcification regions in the results andsmoothed the ventricular edges using the morphologicallyclosed operation
24 Evaluation of the Segmentation Method We applied fourmeasures including Dice metric (Dice) root mean squarederror (RMSE) reliability (R)28 and correlation coefficient(119877) to assess the performance of the proposed segmentationmethod The four measures are defined as follows
Location of the column in image
Inte
nsity
val
ue
6 pixels
10005
10
1520
25
30
35
40
45
50
150 200 250 300 350 400
6 pixels
Figure 5 A sectional view of the gray-scale map for brain image
(1) Dice Metric Let 119881119904 represent the automatically segmentedvolume and 119881119903 represent the manual segmentation (iereference standard) The Dice is defined as
Dice = 2119881119904 cap 119881119903119881119904 + 119881119903 (6)
The value of Dice is between 0 and 1 Higher Dice indicatesbetter overlap between segmented volumes and the referencestandard
(2) Root Mean Squared Error The RMSE calculates the dis-tance between the corresponding points on the automaticallysegmented and reference boundaries defined by
RMSE = ( 1119873119873sum119894=1
(119909119904119894 minus 119909119903119894)2 + (119910119904119894 minus 119910119903119894)2)12
(7)
where (119909119904119894 119910119904119894) is a point on the segmented boundary and(119909119903119894 119910119903119894) is the closest point to (119909119904119894 119910119904119894) on the referenceboundary The lower RMSE the better performance
(3) Reliability The reliability function is used to assess thereliability of segmentation method defined as
R (119889)= Number of volumes segmented with Dice gt 119889
Total number of volumes (8)
BioMed Research International 7
(a) (b)
(c) (d)
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Figure 4 Flowchart of the exclusion of stroke area in the ventricular segmentation result
preliminary segmentation of the ventricle If the intensityrange of the stroke is greater than 119881max the preliminarysegmentation is a good result Whereas if the intensity rangeof the stroke is less than119881max the segmentation result may beunacceptable since it may also contain some stroke regions
Then we utilized the 3D connectivity of the preliminarysegmentation result to obtain the largest volume as the initialsegmentation of the ventricle The stroke regions or noiseareas without the 3D connectivity to the ventricle could beexcluded by this step Figure 4(b) shows that the large strokeregions are connected to the ventricle in the segmentation
233 Detection of the Big Stroke Regions Since big strokeregions are mainly related to the anterior cerebral artery ormiddle cerebral artery these stroke regions are mostly closedto the brain edge Thus we proposed a brain edge checkingalgorithm to determinewhether the big stroke regions exist in
the segmentation result An annular region of the brain edgewas defined to detect the objects Assumed that theminimumside length of the minimum bounding rectangle of the brainwas 119871min the width of the annular region could be calculatedby 015 times 119871min to avoid some parts of the ventricle fallingwithin the annular regionThemask of the brain edge annularregion was shown in Figure 4(c) Thus if the objective areawas greater than the threshold we labeled it as the strokeregion The threshold was empirically selected as 20 pixels toallow the presence of noise
234 Determination of the Big Stroke Regions We proposedan image difference technique based on the heuristic search-ing algorithm to extract the big stroke regions which weresuccessfully detected in the preliminary segmentation bythe edge checking method This image difference techniqueessentially applied the difference between two segmentation
6 BioMed Research International
results by different threshold values for determining thestroke regions We first defined the critical threshold value(ie119879critical) If a threshold was greater than119879critical the strokeregions in the segmentation result of this threshold couldbe detected by the edge analysis method whereas if thethreshold was smaller or equal to 119879critical none stroke regioncould be detected We then obtained the stroke regions by
PA asymp 119891 (119881max) minus 119892 (119891 (119879critical)) (4)
where 119891(lowast) was the threshold method 119892(lowast) represented thesubsequent refine algorithms such as morphology methodand PA represented the stroke regions So we obtained theventricle areas
119891 (119881max) minus 119892 (PA) (5)
The vital step in the image difference method is todetermine the critical threshold value 119879critical We applied thegold searching method and the edge checking method toobtain the 119879critical in range [119881min 119881max]235 Exclusion of the Small Stroke Regions Some smallstroke regions may still present in the segmentation resultfrom the image difference approach To address this problemwe developed an adaptive template matching approachwhich applied the mask of the main part of the ventricleto exclude the remaining small stroke regions The templatewas generated from each image It did not contain the wholeventricle but covers the main part of the ventricle
Figure 5 shows a sectional view of the gray-scale map fora brain image The intensity difference between the ventricleand brain parenchyma was around 20 intensity values whilethe transition area was only 6 to 7 pixels Thus we applied119881min as a threshold for ventricle segmentation and took the3D largest connected region as the ventricle as shown inFigure 6(b) The ventricle segmentation merely containingthe right and left lateral ventricles and without the 3rd and4th ventricle was adaptively selected as the templates Toensure that the template covers the ventricle we conductedsomemorphological analysis including closed operation andexpansion operation The generated template was shown inFigures 6(c) and 6(d)
After these steps we linearly registered the templatewith the corresponding segmentationThe objects within thetemplate served as the ventricle so that the remaining smallstroke areas could be excluded from the segmentation results
236 Refinement of the Ventricular Segmentation Weemployed connected component labeling to the segmentedventricle regionThe largest volume served as the ventricularWe then removed the calcification regions in the results andsmoothed the ventricular edges using the morphologicallyclosed operation
24 Evaluation of the Segmentation Method We applied fourmeasures including Dice metric (Dice) root mean squarederror (RMSE) reliability (R)28 and correlation coefficient(119877) to assess the performance of the proposed segmentationmethod The four measures are defined as follows
Location of the column in image
Inte
nsity
val
ue
6 pixels
10005
10
1520
25
30
35
40
45
50
150 200 250 300 350 400
6 pixels
Figure 5 A sectional view of the gray-scale map for brain image
(1) Dice Metric Let 119881119904 represent the automatically segmentedvolume and 119881119903 represent the manual segmentation (iereference standard) The Dice is defined as
Dice = 2119881119904 cap 119881119903119881119904 + 119881119903 (6)
The value of Dice is between 0 and 1 Higher Dice indicatesbetter overlap between segmented volumes and the referencestandard
(2) Root Mean Squared Error The RMSE calculates the dis-tance between the corresponding points on the automaticallysegmented and reference boundaries defined by
RMSE = ( 1119873119873sum119894=1
(119909119904119894 minus 119909119903119894)2 + (119910119904119894 minus 119910119903119894)2)12
(7)
where (119909119904119894 119910119904119894) is a point on the segmented boundary and(119909119903119894 119910119903119894) is the closest point to (119909119904119894 119910119904119894) on the referenceboundary The lower RMSE the better performance
(3) Reliability The reliability function is used to assess thereliability of segmentation method defined as
R (119889)= Number of volumes segmented with Dice gt 119889
Total number of volumes (8)
BioMed Research International 7
(a) (b)
(c) (d)
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
results by different threshold values for determining thestroke regions We first defined the critical threshold value(ie119879critical) If a threshold was greater than119879critical the strokeregions in the segmentation result of this threshold couldbe detected by the edge analysis method whereas if thethreshold was smaller or equal to 119879critical none stroke regioncould be detected We then obtained the stroke regions by
PA asymp 119891 (119881max) minus 119892 (119891 (119879critical)) (4)
where 119891(lowast) was the threshold method 119892(lowast) represented thesubsequent refine algorithms such as morphology methodand PA represented the stroke regions So we obtained theventricle areas
119891 (119881max) minus 119892 (PA) (5)
The vital step in the image difference method is todetermine the critical threshold value 119879critical We applied thegold searching method and the edge checking method toobtain the 119879critical in range [119881min 119881max]235 Exclusion of the Small Stroke Regions Some smallstroke regions may still present in the segmentation resultfrom the image difference approach To address this problemwe developed an adaptive template matching approachwhich applied the mask of the main part of the ventricleto exclude the remaining small stroke regions The templatewas generated from each image It did not contain the wholeventricle but covers the main part of the ventricle
Figure 5 shows a sectional view of the gray-scale map fora brain image The intensity difference between the ventricleand brain parenchyma was around 20 intensity values whilethe transition area was only 6 to 7 pixels Thus we applied119881min as a threshold for ventricle segmentation and took the3D largest connected region as the ventricle as shown inFigure 6(b) The ventricle segmentation merely containingthe right and left lateral ventricles and without the 3rd and4th ventricle was adaptively selected as the templates Toensure that the template covers the ventricle we conductedsomemorphological analysis including closed operation andexpansion operation The generated template was shown inFigures 6(c) and 6(d)
After these steps we linearly registered the templatewith the corresponding segmentationThe objects within thetemplate served as the ventricle so that the remaining smallstroke areas could be excluded from the segmentation results
236 Refinement of the Ventricular Segmentation Weemployed connected component labeling to the segmentedventricle regionThe largest volume served as the ventricularWe then removed the calcification regions in the results andsmoothed the ventricular edges using the morphologicallyclosed operation
24 Evaluation of the Segmentation Method We applied fourmeasures including Dice metric (Dice) root mean squarederror (RMSE) reliability (R)28 and correlation coefficient(119877) to assess the performance of the proposed segmentationmethod The four measures are defined as follows
Location of the column in image
Inte
nsity
val
ue
6 pixels
10005
10
1520
25
30
35
40
45
50
150 200 250 300 350 400
6 pixels
Figure 5 A sectional view of the gray-scale map for brain image
(1) Dice Metric Let 119881119904 represent the automatically segmentedvolume and 119881119903 represent the manual segmentation (iereference standard) The Dice is defined as
Dice = 2119881119904 cap 119881119903119881119904 + 119881119903 (6)
The value of Dice is between 0 and 1 Higher Dice indicatesbetter overlap between segmented volumes and the referencestandard
(2) Root Mean Squared Error The RMSE calculates the dis-tance between the corresponding points on the automaticallysegmented and reference boundaries defined by
RMSE = ( 1119873119873sum119894=1
(119909119904119894 minus 119909119903119894)2 + (119910119904119894 minus 119910119903119894)2)12
(7)
where (119909119904119894 119910119904119894) is a point on the segmented boundary and(119909119903119894 119910119903119894) is the closest point to (119909119904119894 119910119904119894) on the referenceboundary The lower RMSE the better performance
(3) Reliability The reliability function is used to assess thereliability of segmentation method defined as
R (119889)= Number of volumes segmented with Dice gt 119889
Total number of volumes (8)
BioMed Research International 7
(a) (b)
(c) (d)
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Figure 6 Generation of the template for ventricle (a) original image (b) initial segmentation result (c) the generated template (d) thecorresponding brain area in the template
where 119889 isin [0 1] R(119889) represents the reliability in yieldingDice 119889(4) Correlation Coefficient 119877 between 119881119904 and 119881119903 is used toassess the quality of a least-squares fitting given by
119877= 119899sum119899119894=1 119881119904119894119881119903119894 minus sum119899119894=1 119881119904119894sum119899119894=1 119881119903119894(119899sum119899119894=1 1198812119904119894 minus (sum119899119894=1 119881119904119894)2)12 (119899sum119899119894=1 1198812119903119894 minus (sum119899119894=1 119881119903119894)2)12
(9)
The value of 119877 ranges from 0 no match between the twovolumes to 1 a perfect match
3 Results
31 Qualitative Evaluation Figure 7 displays the alignmentof three representative brain images The original imageswere shown in (a) (b) to (d) were the segmented lightcurvesegment determined midsagittal line and the final
aligned result respectively Only a short light curve segmentwas detected in the brain image of the first row howeverour algorithm still accurately determined themidsagittal linewhich was attributable to 3D fitting of the middle sagittalplane based on segmented light curvesegments We can findthat our alignment algorithm yielded good performance
Figure 8 shows the results of ventricle segmentation Theoriginal brain image ventricle segmentation result and refer-ence standardwere shown in (a) to (c) respectively Althoughsome stroke regions were attached to the ventricle in originalimages they were all excluded in the segmentation resultsThis result means that our proposed segmentation methodcan obtain satisfactory results on images with ischemicstroke
32 Quantitative Evaluation Results We quantitatively asse-ssed the ventricle segmentation results using Dice RMSEthe reliability (R) and correlation coefficient (119877) The meanDice sensitivity specificity and RMSE were 09447 09690998 and 0219 respectively as shown inTable 1The analysis
8 BioMed Research International
(a) (b) (c) (d)
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Figure 7 Alignment performance original image without skull (a) detected light curvesegment (b) determined midsagittal line with redcolor for each slice based on 3D fitting of light curves (c) aligned brain image where the white line shows the midline of the image (d)
Table 1 Quantitative performance evaluations (Dice sensitivity specificity and RMSE) on 50 cases of patients with ischemic stroke regions
results of these metrics confirm the desirable performance ofour proposed method
The proposedmethod produced a reliability ofR(085) =0987 for ventricle segmentation which means all these caseshave a good agreement (Dice gt 085) Figure 9(a) plots R
as a function of 119889 (119889 ge 078) for the ventricle segmentationIt further shows the acceptable performance of the proposedmethod
The correlation coefficients between automatic segmen-tation result and reference standard are 0994 The linear
BioMed Research International 9
(a) (b) (c)
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Figure 8 Performance of ventricle segmentation original brain images (a) ventricle segmentation result outlined with red contours (b)contours of the reference ventricle (c)
regression plotted in Figure 9(b) which indicates a closecorrelation between the results of the proposed method andthe reference standard
4 Discussion
The stroke regions on CT are often adjacent or connected tothe ventricle and their intensities are similar which makesit highly difficult for accurate segmentation of the ventricleTo achieve this goal we developed a combined segmentationstrategy composed of connectivity image difference methodand adaptive template method that is developed to excludestroke regions from the ventricular segmentation resultwhich constitutes the major strength of our segmentationscheme
Image difference method was used to extract the largelesion regions In this approach the most critical step was tosearch the critical threshold for obtaining the ventricular seg-mentation result without stroke regions This result served asldquobenchmark ventricular maskrdquo and acted as the subtrahendin the image difference method However the edge checkingmethod only worked well for the large stroke regions sothis method was not able to efficiently detect the smallstroke areas when they presented in the segmentation resultfrom the critical threshold If the benchmark ventricularmask contains small stroke areas these small stroke regionswould be left in the final segmentation results Therefore theadaptive template method was developed to remove thesesmall stroke regions which would further break up theconnectivity relationship between the lesion regions and the
10 BioMed Research International
Dice metric
Relia
bilit
y
0987
080
02
04
06
08
1
085 09 095 1
(a)
Manual volume
Auto
mat
ic v
olum
e
0
Sample
002
02
04
04
06
06
08
08
1
1
Y = X
Y = 10098X + 00153
R2 = 09943
(b)
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Figure 9 (a) The reliability of our methodR(085) = 0987 (b) segmentation volumes of our method versus manual volumes 119877 = 0997
ventricular region in 3D space Finally we took the largest 3connected component in the segmentation as the ventricularregion to refine the results
The limitation of this segmentation system is that somesmall stroke region may still exist in the segmentation resultdue to the local property of the adaptive template whichcovers the main part of the ventricle Differentiation of theventricle and stroke region is a challenging task In the futurewe will combine the prior template of the ventricle andadaptive template to exclude the stroke region in the initialsegmentation result Besides we will collect more data tovalidate our proposed segmentation system
5 Conclusion
The accurate ventricle segmentation is a critical step in thedevelopment of CAD for acute ischemic stroke Since ische-mic stroke regions are generally adjacent to the brain ventriclewith similar intensity it is a challenging task to segment ven-tricle In this study we developed an objective segmentationsystem of brain ventricle in CT We proposed three differentschemes to exclude the stroke regions from initial segmen-tation which are the main contributions in this work Theexperiments illustrate the proposed segmentation methodthat can obtain a good performance for segmentation ofventricle in brainCT scanswith ischemic strokewhichwouldsignificantly facilitate ischemic stroke region localization
Competing Interests
The authors have no relevant competing interests to disclose
References
[1] M A Balafar A R Ramli M I Saripan and S MashohorldquoReview of brain MRI image segmentation methodsrdquo ArtificialIntelligence Review vol 33 no 3 pp 261ndash274 2010
[2] C M Li R Huang Z H Ding J C Gatenby D N Metaxasand J C Gore ldquoA level set method for image segmentation inthe presence of intensity inhomogeneities with application toMRIrdquo IEEE Transactions on Image Processing vol 20 no 7 pp2007ndash2016 2011
[3] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009
[4] L Wang F Shi W Lin J H Gilmore and D Shen ldquoAutomaticsegmentation of neonatal images using convex optimizationand coupled level setsrdquo NeuroImage vol 58 no 3 pp 805ndash8172011
[5] H Wang and B Fei ldquoA modified fuzzy C-means classificationmethod using a multiscale diffusion filtering schemerdquo MedicalImage Analysis vol 13 no 2 pp 193ndash202 2009
[6] X Yang and B Fei ldquoAmultiscale andmultiblock fuzzy C-meansclassification method for brain MR imagesrdquo Medical Physicsvol 38 no 6 pp 2879ndash2891 2011
[7] S Kumazawa T Yoshiura H Honda F Toyofuku and Y Higa-shida ldquoPartial volume estimation and segmentation of braintissue based on diffusion tensor MRIrdquo Medical Physics vol 37no 4 pp 1482ndash1490 2010
[8] X Li L Li H Lu and Z Liang ldquoPartial volume segmentationof brainmagnetic resonance images based onmaximum a post-eriori probabilityrdquoMedical Physics vol 32 no 7 pp 2337ndash23452005
[9] K Wei B He T Zhang and X Shen ldquoA novel method forsegmentation of CT head imagesrdquo in Proceedings of the 1st Inter-national Conference on Bioinformatics and Biomedical Engineer-ing (ICBBE rsquo07) pp 717ndash720 Wuhan China July 2007
[10] T H Lee M F A Fauzi and R Komiya ldquoSegmentation of CTbrain images using K-means and EM clusteringrdquo in Proceed-ings of the 5th International Conference on Computer GraphicsImaging and Visualisation Modern Techniques and Applications(CGIV rsquo08) August 2008
[11] W Chen and K Najarian ldquoSegmentation of ventricles inbrain CT images using gaussian mixture model methodrdquo in
BioMed Research International 11
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
Proceedings of the ICME International Conference on ComplexMedical Engineering (CME rsquo09) April 2009
[12] V Gupta W Ambrosius G Y Qian et al ldquoAutomatic segmen-tation of cerebrospinal fluid white and gray matter in unen-hanced computed tomography imagesrdquo Academic Radiologyvol 17 no 11 pp 1350ndash1358 2010
[13] W A Chen R Smith S-Y Ji K R Ward and K NajarianldquoAutomated ventricular systems segmentation in brain CTimages by combining low-level segmentation and high-leveltemplate matchingrdquo BMC Medical Informatics and DecisionMaking vol 9 supplement 1 article S4 2009
[14] J Liu S Huang V Ihar A Wojciech L C Lee and WL Nowinski ldquoAutomatic model-guided segmentation of thehuman brain ventricular system from CT imagesrdquo AcademicRadiology vol 17 no 6 pp 718ndash726 2010
[15] X Qian J Wang S Guo and Q Li ldquoAn active contour modelfor medical image segmentation with application to brain CTimagerdquoMedical Physics vol 40 no 2 Article ID 021911 2013
[16] X Qian J Wang and Q Li ldquoAutomated segmentation of brainventricles in unenhanced CT of patients with ischemic strokerdquoin Proceedings of the Medical Imaging 2013 Computer-AidedDiagnosis LakeBuenaVista (OrlandoArea)FlaUSA February2013
