performance evaluation of contour based segmentation

Research ArticlePerformance Evaluation of Contour Based SegmentationMethods for Ultrasound Images

R J Hemalatha 1 V Vijaybaskar2 and T R Thamizhvani 3

1Department of Biomedical Engineering Sathyabama Institute of Science and Technology Chennai TamilNadu-600 India2Department of Electronica andTelecommunication Sathyabama Institute of Science andTechnology Chennai TamilNadu-600 India3Department of Biomedical Engineering Vels Institute of Science Technology and Advanced Studies ChennaiTamilNadu-600117 India

Correspondence should be addressed to R J Hemalatha rjhemalathagmailcom

Received 26 May 2018 Revised 9 August 2018 Accepted 30 August 2018 Published 16 September 2018

Academic Editor Huiyu Zhou

Copyright copy 2018 R J Hemalatha et alThis 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

Active contour methods are widely used for medical image segmentation Using level set algorithms the applications of activecontourmethods have become flexible and convenientThis paper describes the evaluation of the performance of the active contourmodels using performancemetrics and statistical analysisWe have implemented five differentmethods for segmenting the synovialregion in arthritis affected ultrasound image A comparative analysis between the methods of segmentation was performed and thebest segmentation method was identified using similarity criteria standard error and F-test For further analysis classification ofthe segmentation techniques using support vector machine (SVM) classifier is performed to determine the absolute method forsynovial region detection With these results localized region based active contour named Lankton method is defined to be thebest segmentation method

1 Introduction

Musculoskeletal disorder (MSD) an epidemic disease is nowa significant health problem in emerging and most develop-ing countries in the world MSD remains the most prevalentdisease in society due to its impact on mobility ability towork and life style Arthritis is one of the prevalent MSDamong all the age groups of people [1] The most affectedportion is the joint region At relatively early stage of thedisease the synovial membrane around the joints used toinflame and leads to degeneration of the joints The mainsystem to visualize the joint state is through ultra sounddiagnosis (USD)TheUSD is less expensive and availablewithall the clinicians [2 3] The USD images represent differenttissue regions with variations of gray shades These imagesare further processed to segment the synovial region and toanalyze the disease condition and further progression of thedisease

One of the essential tasks in image analysis is image seg-mentation Segmentation is to extract objects from the imagesby dividing the image into set of regions with different

properties Segmentation plays an important role in auto-matic object recognition or pattern identification process toidentify pathologies and medical diagnosis [4 5] The mostchallenging task is to extract the contour and boundariesof the desired region for dynamic analysis of anatomicalstructures One of the most robust segmentation methods formedical images is active contour method (ACM)Themethodimplies curve evolution to detect the region of interest in agiven image [6ndash8]The segmentation process is based on edgeand region based approach The geometric active contourmodel proposed by Caselles and Malladi et al is an edgebased approach which is based on evolution of curves andgeometric flows [9] Chan and Vese have proposed edgeless active contour model which is one of the most well-known region based method Bernard et al have proposed aparameterized active contour method More recently Li et aland Lankton proposed a method which utilizes local regioninformation for segmentation [9ndash14]

In this paper we have applied different segmentationmethods to arthritis affected finger joint ultrasound imagesto segment the synovial region The efficiency of these

HindawiAdvances in MultimediaVolume 2018 Article ID 4976372 8 pageshttpsdoiorg10115520184976372

2 Advances in Multimedia

methods is analyzed using performance analysis metrics andstatistical analysis method For performance analysis metricswe have used Dice coefficient and Hausdorff coefficient andfor statistical analysis method standard error F-test wereused At the end classification is used to define the absoluteactive contour method for segmentation of synovial regionby training and analyzing the performance metrics andstatistical values The rest of the paper is organized as followsSection 2 The Methods and Materials Section 3 the resultsand discussion and Section 4 the conclusion

2 Methods and Materials

Inmedical imaging the segmentation of regions with specificparameters is carried out with the help of active contourmodels Because these models develop a contour aroundthe target object and segregate it from the image the seg-mented image possesses only the required information ofthe target object [15] The level set segmentation methodslike Caselles ChanndashVese Bernard Li and Lankton areapplied on arthritis affected finger joint images obtainedfrom the MEDUSA database httpmedusaaeipolslpl[16ndash18] Further using performance analysis metrics like dicecoefficient and Hausdroff distance and statistical analysismetrics like standard error and F-test describes the significantdifference between the techniques used for segmentationClassification using SVM defines the best suited method forsynovial region segmentation

MEDUSA is a standardized and authorized databasewhich consists of finger joint images of different grades (grade0 grade 1 grade 2 and grade 3) of synovitis Various studiesrelated to arthritis and synovitis are performed using thisdatabase

21 Caselles Caselles is geodesic based active contour meth-ods which largely depend on the level set functions thatdescribe the specific regions in the image for segmentationContours are described based on the geometric flow of curveand detection of objects in the image [11]This type of contourmodelmodifies the curve in the plane bymoving the points ofthe curve perpendicularThemotion of the points is at a speedproportional to the curvature of the region in the imageBy adding an area of minimizing region (balloon force)propagation of contour occurs internally by minimization ofthe interior energy given by

E (C) = int 119892 (119868 (119888 (119875))) 100381710038171003817100381710038171198621015840 (119875)10038171003817100381710038171003817 119889119875 (1)

119892 (119868) = 11 + nabla (119866 lowast 119868)2 (2)

I is image intensity G is Gaussian Filter of unity variance andC is derived parametric curve to regions with high gradientwhere set level function is executed as a signed distancefunction (P= 0(x) )

Contour models use the energy forces for geometric flowcurve description Geometric contours can be obtained basedon regions and edges in the curvature of the image [12]

22 ChanndashVese ChanndashVese is a region based method whichsegments an image into two homogeneous regions Themethod utilizes energy minimization technique defined byweighted values corresponding to the average value of sumof intensity difference from outside and inside the segmentedregion [9 10] Contours are based on either the varianceinside and outside contour or the squared difference betweenaverage intensities inside and outside the contours alongwith the total contour length This contour model helps todetermine different image properties not only edges and italso includes regions based on texture and other geometricalfeatures Energy defines the entire region of interest from theimage

The total energy of the model is given in

E0 = 120583intΩ120575 (120593 (a)) 1003816100381610038161003816nabla120593 (a)1003816100381610038161003816 da + vint

ΩH (120593 (119886))da

+ 1205821 intΩH (120593 (a)) 1003816100381610038161003816f minus C1

1003816100381610038161003816+ 1205822 int

Ω(1 minusH (120593 (a)) 1003816100381610038161003816f minus C21003816100381610038161003816 da

(3)

120583 ] 1205821 1205822 real parametersC1 C2 constants determined for segmentationminus1le 120593(a)le1 level set function in which 120593(a)=0specifies the interfacef original imageH heavy side function in 1 dimension centered at 0and 120575=H1015840

23 Bernard Bernard method utilizes B-spline coefficientsas energy minimization function These utilize parameter-ized active contour method [12] Spline coefficients definethe contour models for the pixels of interest The energybased functions inside and outside is described with thesecoefficients Contour models describe the entire structureswith inflation force that can overpower forces from weakedges amplifying the issue with localization of initial guessTo speed up the process a linear combination of B-spline basisfunctions is used and given in

E0 = intsdotΩF (I (a) 0 (a)) dx (4)

F (I (a) 0 (a)) = (I (a) minus 120592)2H (0 (a))+ (I (119886) minus 120583)2 (I minusH (0 (a))) (5)

Φ(a) is linear combination of B-spline basis functions

24 Chumming Li In order to separate the region intotwo homogenous regions this method utilizes using localneighbourhood statistics for each pixel given in (6) It useslocal region information for segmentation [13] The energyfunction of the region based active contour model is rangeof region based domain kernel function By minimizing the

Advances in Multimedia 3

energy function the region of elements of the target could bedetermined in images with contours

E0 = 1205821∬K120590 (a minus b) 1003816100381610038161003816I (b) minus f1 (a)10038161003816100381610038162H (0 (a)) db da+ 1205822∬K120590 (a minus b) 1003816100381610038161003816I (b) minus f2 (a)10038161003816100381610038162sdot (1 minusH (0 (a))) db da + 120592int120575 (0 (a))sdot nabla0 (a) da + 120583int 12 (nabla0 (a) minus 1)2 da

(6)

I(a) pixel intensity at xH heavy side functionK120590 Gaussian Kernel

K120590 (120583) = 1(2120587)1205872 120590120587 119890minus120583221205902 (7)

25 Lankton Lankton is a region based active contourmethod which segments non homogeneous objects Thismethod utilizes localizing region based energy which seg-ments the region based on local information It is not suitablefor unsupervised image segmentation as it requires appro-priate curve initialization [14] These models form contourboundaries with energy forces required for the particularregion of interest The energy inside and outside depends onthe local region pixels of the image that describes the requiredregion The energy equation is illustrated in

E0 = intΩ120575 (0 (a)) int

ΩF (I (a) 0 (b)) B (a b) dbda

+ 120582intΩnabla0 (a) 120575 (0 (a)) da (8)

120575 is Dirac functionB is Ball of radius r centered at point x

26 Performance Evaluation Metrics The segmentationmethods are qualitatively and quantitatively assessed andcompared with each other based on three kinds of criteriaBased on this the best suited algorithm is chosen forparticular applications

Visual criteria The segmented region using the activecontour methods is compared with the annotated images byexpert radiologist as reference image The segmented regionobtained from level set function methods is compared withreference image

Computation time The time taken for each algorithm tosegment the region represents the speed of the algorithmThespeed of the respective algorithm is compared

Similarity criteria This criterion measures the similaritybetween the reference and segmented image The quality ofthe segmented image is measured by calculating the Dicecoefficient Hausdorff distance and PSNR

Dice coefficient compares the segmented region with thereference region from the annotated image and provides thedice coefficient values ranging between 0 and 1 If it is 1 thesegmented region is more similar and it is different when it is0 [19]

Dice = 2 (A cap B)A + B

(9)

Hausdroff distance is a metric to measure dissimilaritybetween two point sets Distance transform is used to com-pute the HD in an image This is used to control the progressof level set based algorithms and to evaluate the quality of theclusters [20]

1198631 (119860 119861) = max119909isin119860

(min119910isin119861

(1003817100381710038171003817119909 minus 1199101003817100381710038171003817)) (10)

27 Statistical Analysis The features like mean varianceand standard errors are calculated for the segmentationmethods Among the five segmentation methods the bestsuited method was identified using statistical analysis

Mean The mean value is termed as average value whichis computed by taking sum of all perceived outcomes dividedby overall number of gray levels The following shows math-ematical expression for mean represented as119909

Mean 119909 = 1119899119899sum119894=1

119909 (11)

where n is sample size and x is observed valueVariance is study of deviation of actual value versus

predicted value The deviation from actual and predictedindicates the performance of the methods used

Standard error is defined as the measure of predictionrsquosaccuracy Estimated standard error is related to sum ofsquared deviations of prediction (that is sum of squareserror) described

120590119890119904119905 = radicsum(119884 minus 1198841015840)2

119873 (12)

120590est is the standard error of the estimate Y is an actual rangeY1015840 is a predicted range andN is the number of pairs of scoressum(119884 minus 1198841015840)2 is the sum of squared differences between theactual scores and the predicted scores

28 Classification Classification is a process to describe theeffective type or class based on the features derived fromthe region of interest Support vector machines (SVM) aremachine learning models SVM is the representation ofobservations as points that maps to form separate divisionsand a clear boundary factor defined as decision boundaryMulticlass support vector machine classifies the types basedon the kernel models Multiclass support vector machineis used to illustrate the appropriate type of active contourtechnique for the segmentation of the synovial region


ORIGINAL IMAGE (ANNOTATED)

CASELLES

CHAN-VESE

CHUMMING LI

LANKTON

BERNARD

Figure 1 Different types of active contour segmentation techniques

3 Results

In thismethod different types of active contour segmentationtechniques are used for the detection of synovial region Thesegmentation methods were evaluated using performancemetrics and statistical analysis Ultrasound images fromthe database are used for the identification of synovialregion Fifty images of different grades are considered for

segmentation of the synovial region from the databaseDifferent active contour segmentation techniques are usedto segment the synovial regions Visual changes in thesegmentation process are illustrated through the images inFigure 1

In this figure the annotated image is defined withgreen colour and the synovial region defined by the fivedifferent types of segmentation is displayed in red colourThe


Dice Coefficient for Segmentation Techniques

DiceCoefficient

Hausdorff Distance for Segmentation techniques

HausdorffDistance

Caselles Chan-vese ChummingLi

Lankton Bernard

Computation Time for Segmentation Techniques

ComputationTime

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

050

100150200250300350400450500

Com

puta

tion

Tim

e

0

20

40

60

80

100

120

Hau

sdor

ff D

ista

nce

0

02

04

06

08

1D

ice c

oeffi

cien

t

Figure 2 Graphical representation of the performance metrics of five different active contour techniques

Table 1 Performance metrics value for several segmentation meth-ods

Segmentationmethods

DiceCoefficient

HausdorffDistance

ComputationTime

Caselles 0823 16136 40608Chan-Vese 0562 18419 55109Chumming Li 0303 39841 388815Lankton 0879 14814 13416Bernard 0204 88601 402057

segmented synovial region when compared to the annotatedimages visually defines the fact that Caselles and Lanktonpossess similarity To analyze the absolute segmentation tech-nique further performance metrics description statisticalanalysis and classification are carried out The performanceanalysis metrics like dice coefficient Hausdorff distance andcomputation time values of an image is tabulated in Table 1

Comparison between each performancemetric of the fivedifferent active contour techniques is graphically representedin Figure 2 These representations illustrate that the Casellesgeodesic active contour and Lankton localized region basedactive contour methods have slightly similar values From thedice coefficient values it is shown that Caselles and Lanktonare more towards 1 that is the highest range Hausdroffdistance also shows the similarity between the two methodsComputation time in seconds defines the fact that thelocalized region based active contour Lankton is efficientTo further classify which is the best suitable segmentationtechnique we perform statistical analysis

In statistical analysis the features like standard erroraverage mean and average variance values are derived fromthe similarity index parameters (Dice coefficient and Haus-droff distance) for the exact determination of the segmenta-tion technique In this Table 2 shows the statistical analysisof the similarity index parameters in performance metricsdefining the average mean standard error and averagevariance values


Table 2 Statistical values of performance metrics for segmentation techniques

Segmentation methods Average Mean Standard Error Average VarianceDice coefficient Hausdorff distance Dice coefficient Hausdorff distance Dice coefficient Hausdorff distance

Caselles 0809plusmn0060 30plusmn1560 0809plusmn00837 30plusmn1118 0809plusmn00451 30plusmn10332Chan-vese 0579plusmn0150 909plusmn2540 0579plusmn01404 909plusmn5682 0579plusmn00233 909plusmn4610Chumming Li 0306plusmn0050 23517plusmn100 0306plusmn01829 23517plusmn6570 0306plusmn00520 23517plusmn862Lankton 0873plusmn0005 187plusmn0010 0873plusmn00858 187plusmn1331 0873plusmn00403 187plusmn3792Bernard 0158plusmn0160 2844plusmn200 0158plusmn00907 2844plusmn1118 0158plusmn00081 2844plusmn489

300

200

100

0

Hau

sdor

ff D

istan

ce

0

05

1

15 2 25 3 35 41050

Dic

e coe

ffici

ent

BernardCaselles Chan-vese Chumming Li Lankton

Figure 3 Graphical representation of mean average values of similarity index parameters

Mean average value defines significant variations amongall the types of active contour techniques used for the processof segmentation of the synovial region But the values possessslight similarity between the two active contour methodsTheslight deviation in the synovial segmentation between the twomethods is illustrated in Figure 3 which represents the meanvalues of the similarity index parameters of the five activecontour methods

The standard error measurement describes the accuracyof the predicted value The difference in standard errorbetween Caselles and Lankton active contours is 00021which is really insignificant So there is a little more simi-larity among the geodesic and localized region based activecontours

Average variance is derived to find out the significant dif-ference between the methods The difference can be definedappropriately using F-test which is performed over the vari-ance valueThe F-test results shows that the Caselles geodesicactive contour and Lankton localized region based activecontour method are likely significant which shows that theregion segmented using these methods is not similar Theresult of the F-test of the geodesic active contour Caselles andlocalized region based active contour Lankton is shown inTable 3

As a result of statistical analysis Caselles and Lanktonhave slightly significant difference and prove to possess dis-similarity in the process of segmentation of synovial regionFor further determination of the exact segmentation tech-nique the performance metrics and statistical featuresobtained from the region undergo classification process In

this process support vector machine (SVM) classifier is usedfor training and identification of the significant segmentationtechnique for synovial region detection Localized regionbased active contour Lankton is described asmore significantwith the help of the results of classification These results aredefined with confusion matrix and scatter plot of the trainedfeatures of the synovial region from the ultrasound imagesIn Figure 4 the confusion matrix is defined with accuracyrates of each category of segmentation With these resultsLankton is determined as the significant nature of themethodin detection of synovial region Features extracted from thesynovial region possess significant variations and localizedregion based active contour is described with the accuracyof the confusion matrix based on the true positive and falsenegative rates

From these results Localized region based active contourismore efficientmethod of active contours for synovial regionsegmentation by training and classifying the features likeperformance metrics and statistical values These featuresdefine the appropriate classification of the region which isperformed with this Lankton active contour

4 Conclusion

In recent days arthritis has become a significant healthproblem Early diagnosis and treatment help the patientsto lead normal life A method was presented to evaluatethe active contour segmentation algorithms to segment thesynovial region from arthritis affected finger ultrasoundimage Performance analysis metrics like Dice coefficient and


Table 3 F-Test for Variances of Caselles and Lankton Method

F-Test Two-Sample for Dice coefficient Variances F-Test Two-Sample for Hausdorff distance VariancesMean 0845719 0829609 Mean 1947198 2757708

Variance 004031 0045104 Variance 3791567 1033178Observations 8 8 Observations 8 8

df 7 7 df 7 7F 0893709 F 0366981

P(Flt=f) one-tail 0442978 P(Flt=f) one-tail 0104682F Criticalone-tail 0264058 F Critical one-tail 0264058

8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class

Figure 4 Confusion matrix of the classification of active contoursegmentation techniques

Hausdroff distance and statistical analysis metrics like stan-dard error and F-test shows the significant difference betweenthe two segmentation method (Caselles and Lankton) forsynovial region Further classification is performed for thederived features such as performance metrics and statisticalvalues Higher accuracy is described for Lankton as the resultof classification process Hence the output of the researchwork shows that Lankton method is the best method forsynovial region segmentation from ultrasound images

Data Availability

The ultrasound images used to support the findings ofthis study were supplied by KrystianRadlak under licenseagreement from MEDUSA Project and so cannot be madefreely available Requests for access to these data should bemade to KrystianRadlak krsytianradlakpolslpl

Additional Points

The complete research work concentrates on evaluation ofcontour based segmentation techniques

Conflicts of Interest

There are no conflicts of interest among the authors withregard to the proposed methodology and performance eval-uation for contour based segmentation techniques for ultra-sound images using statistical analysis

References

[1] R Sharma Ed Epidemiology of Musculoskeletal Conditions inIndia International Journal on Computer Science and Engi-neering (IJCSE) New Delhi India 2012

[2] A Bk J Segen K Wereszczyski P Mielnik M Fojcik and MKulbacki ldquoDetection of linear features including bone and skinareas in ultrasound images of jointsrdquo PeerJ vol 2018 no 3 2018

[3] G Schett ldquoSynovitisndashan inflammation of joints destroying thebonerdquo Swiss Medical Weekly vol 142 p w13692 2012

[4] MKrishnaveni ldquoQuantitative evaluation of Segmentation algo-rithms based on level set method for ISL datasetsrdquo InternationalJournal on Computer Science and Engineering (IJCSE) vol 3 pp2361ndash2369 2011

[5] D Barbosa T Dietenbeck J Schaerer J DrsquoHooge D Fribouletand O Bernard ldquoB-spline explicit active surfaces an efficientframework for real-time 3-D region-based segmentationrdquo IEEETransactions on Image Processing vol 21 no 1 pp 241ndash251 2012

[6] D Reska C Boldak andM Kretowski ldquoA Texture-BasedEner-gy forActive Contour Image Segmentationrdquo in Image ProcessingCommunications Challenges Advances in Intelligent Systems andComputing R Choras Ed vol 313 Springer Cham 2015

[7] MAirouche L Bentabet andM Zelmat ldquoImage SegmentationUsing Active Contour Model and Level Set Method Appliedto Detect Oil Spillsrdquo in Proceedings of the In Proceedings of theWorld Congress on Engineering vol I pp 846ndash850 LondonUK 2009

[8] R Goldenberg R Kimmel E Rivlin and M Rudzsky ldquoFastgeodesic active contoursrdquo IEEE Transactions on Image Process-ing vol 10 no 10 pp 1467ndash1475 2001

[9] T Chan and L Vese ldquoAn active contour model without edgesrdquoin Scale-Space Theories in Computer Vision vol 1682 of LectureNotes in Computer Science pp 141ndash151 Springer Berlin Ger-many 1999

[10] P Getreuer ldquoChan-Vese Segmentationrdquo Image Processing OnLine vol 2 pp 214ndash224 2012

[11] V Caselles F Catte T Coll and F Dibos ldquoA geometric modelfor active contours in image processingrdquo Numerische Mathe-matik vol 66 no 1 pp 1ndash31 1993


[12] O Bernard D Friboulet P Thevenaz and M Unser ldquoVari-ational B-spline level-set a linear filtering approach for fastdeformable model evolutionrdquo IEEE Transactions on ImageProcessing vol 18 no 6 pp 1179ndash1191 2009

[13] C Li C Y Kao J C Gore and Z Ding ldquoMinimization ofregion-scalable fitting energy for image segmentationrdquo IEEETransactions on Image Processing vol 17 no 10 pp 1940ndash19492008

[14] S Lankton and A Tannenbaum ldquoLocalizing region-basedactive contoursrdquo IEEE Transactions on Image Processing vol 17no 11 pp 2029ndash2039 2008

[15] A Khadidos V Sanchez and C-T Li ldquoActive contours basedon weighted gradient vector flow and balloon forces for medicalimage segmentationrdquo pp 902ndash906

[16] MEDUSA ldquoAutomated assessment of joint synovitis activ-ity from medical ultrasound and power Doppler examina-tions using image processing and machine learning methodsrdquohttpeeagrantsorgproject-portalprojectPL12-0015

[17] K Radlak N Radlak and B Smolka ldquoAutomatic detection ofbones based on the confidence map for Rheumatoid Arthritisanalysisrdquo in Proceedings of the 5th EccomasThematic Conferenceon Computational Vision andMedical Image Processing VipIM-AGE 2015 pp 215ndash220 Spain October 2015

[18] P Mielnik M Fojcik J Segen and M Kulbacki ldquoA NovelMethodof Synovitis Stratification inUltrasoundUsingMachineLearning Algorithms Results From Clinical Validation of theMEDUSA Projectrdquo Ultrasound in Medicine amp Biology vol 44no 2 pp 489ndash494 2018

[19] T Dietenbeck M Alessandrini D Friboulet and O BernardldquoCreaseg A free software for the evaluation of image segmenta-tion algorithms based on level-setrdquo in Proceedings of the 201017th IEEE International Conference on Image Processing ICIP2010 pp 665ndash668 Hong Kong September 2010

[20] A A Taha and A Hanbury ldquoAn efficient algorithm for calculat-ing the exact hausdorff distancerdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 37 no 11 pp 2153ndash21632015

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


methods is analyzed using performance analysis metrics andstatistical analysis method For performance analysis metricswe have used Dice coefficient and Hausdorff coefficient andfor statistical analysis method standard error F-test wereused At the end classification is used to define the absoluteactive contour method for segmentation of synovial regionby training and analyzing the performance metrics andstatistical values The rest of the paper is organized as followsSection 2 The Methods and Materials Section 3 the resultsand discussion and Section 4 the conclusion

2 Methods and Materials

Inmedical imaging the segmentation of regions with specificparameters is carried out with the help of active contourmodels Because these models develop a contour aroundthe target object and segregate it from the image the seg-mented image possesses only the required information ofthe target object [15] The level set segmentation methodslike Caselles ChanndashVese Bernard Li and Lankton areapplied on arthritis affected finger joint images obtainedfrom the MEDUSA database httpmedusaaeipolslpl[16ndash18] Further using performance analysis metrics like dicecoefficient and Hausdroff distance and statistical analysismetrics like standard error and F-test describes the significantdifference between the techniques used for segmentationClassification using SVM defines the best suited method forsynovial region segmentation

MEDUSA is a standardized and authorized databasewhich consists of finger joint images of different grades (grade0 grade 1 grade 2 and grade 3) of synovitis Various studiesrelated to arthritis and synovitis are performed using thisdatabase

21 Caselles Caselles is geodesic based active contour meth-ods which largely depend on the level set functions thatdescribe the specific regions in the image for segmentationContours are described based on the geometric flow of curveand detection of objects in the image [11]This type of contourmodelmodifies the curve in the plane bymoving the points ofthe curve perpendicularThemotion of the points is at a speedproportional to the curvature of the region in the imageBy adding an area of minimizing region (balloon force)propagation of contour occurs internally by minimization ofthe interior energy given by

E (C) = int 119892 (119868 (119888 (119875))) 100381710038171003817100381710038171198621015840 (119875)10038171003817100381710038171003817 119889119875 (1)

119892 (119868) = 11 + nabla (119866 lowast 119868)2 (2)

I is image intensity G is Gaussian Filter of unity variance andC is derived parametric curve to regions with high gradientwhere set level function is executed as a signed distancefunction (P= 0(x) )

Contour models use the energy forces for geometric flowcurve description Geometric contours can be obtained basedon regions and edges in the curvature of the image [12]

22 ChanndashVese ChanndashVese is a region based method whichsegments an image into two homogeneous regions Themethod utilizes energy minimization technique defined byweighted values corresponding to the average value of sumof intensity difference from outside and inside the segmentedregion [9 10] Contours are based on either the varianceinside and outside contour or the squared difference betweenaverage intensities inside and outside the contours alongwith the total contour length This contour model helps todetermine different image properties not only edges and italso includes regions based on texture and other geometricalfeatures Energy defines the entire region of interest from theimage

The total energy of the model is given in

E0 = 120583intΩ120575 (120593 (a)) 1003816100381610038161003816nabla120593 (a)1003816100381610038161003816 da + vint

ΩH (120593 (119886))da

+ 1205821 intΩH (120593 (a)) 1003816100381610038161003816f minus C1

1003816100381610038161003816+ 1205822 int

Ω(1 minusH (120593 (a)) 1003816100381610038161003816f minus C21003816100381610038161003816 da

(3)

120583 ] 1205821 1205822 real parametersC1 C2 constants determined for segmentationminus1le 120593(a)le1 level set function in which 120593(a)=0specifies the interfacef original imageH heavy side function in 1 dimension centered at 0and 120575=H1015840

23 Bernard Bernard method utilizes B-spline coefficientsas energy minimization function These utilize parameter-ized active contour method [12] Spline coefficients definethe contour models for the pixels of interest The energybased functions inside and outside is described with thesecoefficients Contour models describe the entire structureswith inflation force that can overpower forces from weakedges amplifying the issue with localization of initial guessTo speed up the process a linear combination of B-spline basisfunctions is used and given in

E0 = intsdotΩF (I (a) 0 (a)) dx (4)

F (I (a) 0 (a)) = (I (a) minus 120592)2H (0 (a))+ (I (119886) minus 120583)2 (I minusH (0 (a))) (5)

Φ(a) is linear combination of B-spline basis functions

24 Chumming Li In order to separate the region intotwo homogenous regions this method utilizes using localneighbourhood statistics for each pixel given in (6) It useslocal region information for segmentation [13] The energyfunction of the region based active contour model is rangeof region based domain kernel function By minimizing the




(6)


K120590 (120583) = 1(2120587)1205872 120590120587 119890minus120583221205902 (7)


E0 = intΩ120575 (0 (a)) int


+ 120582intΩnabla0 (a) 120575 (0 (a)) da (8)








(9)


1198631 (119860 119861) = max119909isin119860

(min119910isin119861

(1003817100381710038171003817119909 minus 1199101003817100381710038171003817)) (10)



Mean 119909 = 1119899119899sum119894=1

119909 (11)




120590119890119904119905 = radicsum(119884 minus 1198841015840)2

119873 (12)





CASELLES

CHAN-VESE

CHUMMING LI

LANKTON

BERNARD


3 Results






DiceCoefficient


HausdorffDistance


Lankton Bernard


ComputationTime

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

050

100150200250300350400450500

Com

puta

tion

Tim

e

0

20

40

60

80

100

120

Hau

sdor

ff D

ista

nce

0

02

04

06

08

1D

ice c

oeffi

cien

t



Segmentationmethods

DiceCoefficient

HausdorffDistance

ComputationTime









300

200

100

0

Hau

sdor

ff D

istan

ce

0

05

1

15 2 25 3 35 41050

Dic

e coe

ffici

ent









4 Conclusion






df 7 7 df 7 7F 0893709 F 0366981


8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class



Data Availability


Additional Points




References
























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





(6)


K120590 (120583) = 1(2120587)1205872 120590120587 119890minus120583221205902 (7)


E0 = intΩ120575 (0 (a)) int


+ 120582intΩnabla0 (a) 120575 (0 (a)) da (8)








(9)


1198631 (119860 119861) = max119909isin119860

(min119910isin119861

(1003817100381710038171003817119909 minus 1199101003817100381710038171003817)) (10)



Mean 119909 = 1119899119899sum119894=1

119909 (11)




120590119890119904119905 = radicsum(119884 minus 1198841015840)2

119873 (12)





CASELLES

CHAN-VESE

CHUMMING LI

LANKTON

BERNARD


3 Results






DiceCoefficient


HausdorffDistance


Lankton Bernard


ComputationTime

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

050

100150200250300350400450500

Com

puta

tion

Tim

e

0

20

40

60

80

100

120

Hau

sdor

ff D

ista

nce

0

02

04

06

08

1D

ice c

oeffi

cien

t



Segmentationmethods

DiceCoefficient

HausdorffDistance

ComputationTime









300

200

100

0

Hau

sdor

ff D

istan

ce

0

05

1

15 2 25 3 35 41050

Dic

e coe

ffici

ent









4 Conclusion






df 7 7 df 7 7F 0893709 F 0366981


8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class



Data Availability


Additional Points




References
























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




CASELLES

CHAN-VESE

CHUMMING LI

LANKTON

BERNARD


3 Results






DiceCoefficient


HausdorffDistance


Lankton Bernard


ComputationTime

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

050

100150200250300350400450500

Com

puta

tion

Tim

e

0

20

40

60

80

100

120

Hau

sdor

ff D

ista

nce

0

02

04

06

08

1D

ice c

oeffi

cien

t



Segmentationmethods

DiceCoefficient

HausdorffDistance

ComputationTime









300

200

100

0

Hau

sdor

ff D

istan

ce

0

05

1

15 2 25 3 35 41050

Dic

e coe

ffici

ent









4 Conclusion






df 7 7 df 7 7F 0893709 F 0366981


8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class



Data Availability


Additional Points




References
























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




DiceCoefficient


HausdorffDistance


Lankton Bernard


ComputationTime

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

Bern

ard

Case

lles

Chan

-ves

e

Chum

min

g Li

Lank

ton

050

100150200250300350400450500

Com

puta

tion

Tim

e

0

20

40

60

80

100

120

Hau

sdor

ff D

ista

nce

0

02

04

06

08

1D

ice c

oeffi

cien

t



Segmentationmethods

DiceCoefficient

HausdorffDistance

ComputationTime









300

200

100

0

Hau

sdor

ff D

istan

ce

0

05

1

15 2 25 3 35 41050

Dic

e coe

ffici

ent









4 Conclusion






df 7 7 df 7 7F 0893709 F 0366981


8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class



Data Availability


Additional Points




References
























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia






300

200

100

0

Hau

sdor

ff D

istan

ce

0

05

1

15 2 25 3 35 41050

Dic

e coe

ffici

ent









4 Conclusion






df 7 7 df 7 7F 0893709 F 0366981


8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class



Data Availability


Additional Points




References
























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia






df 7 7 df 7 7F 0893709 F 0366981


8080

8282

9090

8080

100100

1010

20

2020

20

1818

Bernard

Bernard

Caselles

Caselles

Chanampvese

Chanampvese

Chumming Li

Chumm

ing Li

Lankton

Lankton

True

clas

s

Predicted class



Data Availability


Additional Points




References
























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia














RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


performance evaluation of contour based segmentation

Documents