This may be the author s version of a work that was submitted/accepted for publication in the following source:

(1)

for publication in the following source:

Alonso-Caneiro, David, Szczesna, Dorota, Iskander, D. Robert, Read, Scott, & Collins, Michael

(2013)

Application of texture analysis in tear film surface assessment based on videokeratoscopy.

Journal of Optometry, 6(4), pp. 185-193.

This file was downloaded from: https://eprints.qut.edu.au/220285/

Consult author(s) regarding copyright matters c

This work is covered by copyright. Unless the document is being made available under a Creative Commons Licence, you must assume that re-use is limited to personal use and that permission from the copyright owner must be obtained for all other uses. If the document is available under a Creative Commons License (or other specified license) then refer to the Licence for details of permitted re-use. It is a condition of access that users recog- nise and abide by the legal requirements associated with these rights. If you believe that this work infringes copyright please provide details by email to [email protected]

License: Creative Commons: Attribution-Noncommercial-No Derivative Works 2.5

Notice: Please note that this document may not be the Version of Record (i.e. published version) of the work. Author manuscript versions (as Sub- mitted for peer review or as Accepted for publication after peer review) can be identified by an absence of publisher branding and/or typeset appear- ance. If there is any doubt, please refer to the published source.

https://doi.org/10.1016/j.optom.2013.07.006

(2)

www.journalofoptometry.org

ORIGINAL ARTICLE

Application of texture analysis in tear ﬁlm surface assessment based on videokeratoscopy

David Alonso-Caneiro

^a^,∗

, Dorota H. Szczesna-Iskander

^b

, D. Robert Iskander

^c

, Scott A. Read

^a

, Michael J. Collins

^a

aSchoolofOptometryandVisionScience,QueenslandUniversityofTechnology,Brisbane,Australia

bInstituteofPhysics,WroclawUniversityofTechnology,Wroclaw,Poland

cInstituteofBiomedicalEngineeringandInstrumentation,WroclawUniversityofTechnology,Wroclaw,Poland

Received28March2013;accepted19July2013

KEYWORDS Highspeed videokeratoscopy;

Tearﬁlmkinetics;

Dryeyedetection;

Imageprocessing

Abstract

Purpose: Videokeratoscopyimages canbe usedfor thenon-invasiveassessmentofthe tear ﬁlm.Inthisworktheapplicabilityofanimageprocessingtechnique,textural-analysis,forthe assessmentofthetearﬁlminPlacidodiscimageshasbeeninvestigated.

Methods:Inthepresenceoftearfilmthinning/break-up,thereflectedpatternfromthevideok- eratoscopeisdisturbedintheregionoftearfilmdisruption.Thus,thePlacidopatterncarries informationaboutthestabilityoftheunderlyingtearfilm.Bycharacterizingthepatternregu- larity,thetearfilmqualitycanbeinferred.Inthispaper,atexturalfeaturesapproachisused toprocessthePlacidoimages.Thismethodprovidesasetoftexturefeaturesfromwhichan estimateofthetearfilmqualitycanbeobtained.Themethodistestedforthedetectionofdry eyeinaretrospectivedatasetfrom34subjects(22-normaland12-dryeye),withmeasurements takenundersuppressedblinkingconditions.

Results:Toassessthecapabilityofeachtexture-featuretodiscriminatedryeyefromnormal subjects,thereceiveroperatingcurve(ROC)wascalculatedandtheareaunderthecurve(AUC), speciﬁcityandsensitivityextracted.Forthedifferentfeaturesexamined,theAUCvalueranged from0.77to0.82,whilethesensitivitytypicallyshowedvaluesabove0.9andthespeciﬁcity showedvaluesaround0.6.Overall,theestimatedROCsindicatethattheproposedtechnique providesgooddiscriminationperformance.

Conclusions: Textureanalysisofvideokeratoscopyimagesisapplicabletostudytearﬁlmanoma- lies indry eye subjects.The proposedtechniqueappears tohave demonstrateditsclinical relevanceandutility.

∗Correspondingauthorat:SchoolofOptometryandVisionScience,QueenslandUniversityofTechnology,KelvinGrove,Brisbane,Australia.

E-mailaddress:[email protected](D.Alonso-Caneiro).

http://dx.doi.org/10.1016/j.optom.2013.07.006

(3)

ARTICLE IN PRESS

2 D.Alonso-Caneiroetal.

PALABRASCLAVE Videoqueratoscopio dealtavelocidad;

Cinéticadela películalagrimal;

deteccióndeojo seco;

Procesamientode imágenes

Aplicacióndeunanálisisdetexturaenlaevaluacióndelasuperﬁciedelapelícula lagrimalmediantevideoqueratoscopio

Resumen

Objetivo:Puedenutilizarselasimágenesdelvideoqueratoscopioparalaevaluaciónnoinvasiva delapelículalagrimal.Enestetrabajosehainvestigadola capacidaddeaplicacióndeuna técnicadeprocesamientodeimágenes,análisisdelatextura,paraevaluarlapelículalagrimal enlasimágenesdeldiscodePlacido.

Métodos: Enpresenciadeladelgazamiento/roturadelapelículalagrimal,elpatrónreﬂejado porelvideoqueratoscopiosealteraenlaregióndelaalteracióndelapelículalagrimal.Por ello,elpatróndePlacidoaportainformaciónacercadelaestabilidaddelapelículalagrimal subyacente.Alcaracterizarlaregularidaddelpatrón,puedeinferirselacalidaddelapelícula lagrimal.Enestetrabajoseutilizalaevaluacióndelascaracterísticasdelatexturaparaproce- sarlasimágenesdeldiscodePlacido.Estemétodoaportaunconjuntodecaracterísticasde textura,apartirdelcualpuedeobtenerseunaestimacióndelacalidaddelapelículalagrimal.

Elmétodoseprobóparadetectarelojosecoenunabasededatosretrospectivaqueincluía a34pacientes(22normales,y12conojoseco),realizándoselasmedicionessuprimiendoel parpadeo.

Resultados: Paraevaluar lacapacidad decadaparámetrodetexturapara discriminarentre el ojo seco y los pacientes normales, se calculó la curva operativa del receptor (ROC), extrayéndoseeláreabajolacurva(AUC),laespecificidadylasensibilidad.Paralasdiferentes características examinadas, el valor del AUC osciló entre 0,77 y 0,82, mientras que la sensibilidad mostró normalmente valores superiores a 0,9, y la especificidad reflejó unos valores cercanosa 0,6.Generalmente,lasROCestimadasindican quela técnicapropuesta aportaunbuenrendimientodiscriminatorio.

Conclusiones:Elanálisisdelatexturamediantelaimagendelvideoqueratoscopioesaplicable alestudiodeanomalíasdelapelículalagrimalenpacientesconojoseco.Latécnicapropuesta parecehaberdemostradosurelevanciayutilidadclínicas.

Introduction

Assessingthe qualityandstability ofthe pre-cornealtear filmisacommonclinicaltaskthatisimportantforassessing suitabilityfor contactlenswear,refractivesurgeryandin the detection and monitoring of dry eye disease.¹ Tradi- tionalclinical techniques toassess the tear film typically relyuponinvasiveand/orsubjectivemethodslikethefluo- rescein tear break-uptime test to assess the integrityof thetear film.This commonly used techniquehas the dis- advantagethattheinstillationofdyemayalterthenatural tearfilmcomposition.² Thus,thedevelopmentof reliable non-invasive,objectivemethodsoftearfilmassessmentis crucialforimproveddiagnosisandclinicalmonitoringoftear filmdisorders.

Oneof the objective,non-invasive techniques for tear filmassessmentisthedynamicvideokeratoscopytechnique, which analyses a continuous recording of images from a Placido disc corneal topographer, and provides dynamic informationregardingchangesinthetearfilm.Earlywork withthistechniqueusedthechangesintheestimatedtopog- raphydataderivedfromtheinstrumenttoassessthetear film.^3,4However,asdemonstratedbyAlonso-Caneiroetal.⁵, the actual topography data may no longer be calculable or accurate under tear film thinning or break-ups. Thus, methods for assessing the tear film surface based onthe rawPlacidodiscimagesratherthantopographyhavebeen proposed.^6,7Thetearfilmlayerisknowntoactlikeamirror andreflectstheprojectedPlacidodiscringpattern.Thus,a

healthyandregulartearfilmsurfaceformsawell-structured reflectedpattern,whileatearfilmthatisthinning,irregular or breaking up forms a non-structured reflected pattern.

Fromnowonandfollowingengineeringterminology,wewill refertothewell-structuredpatternasthesignal,andthe unstructuredpatternastheinterference.Asetofexemplary videokeratoscopy images with different kinds of interfer- enceisshowninFig.1.

Inpreviousstudies,measurementsofpatterncoherence⁸ (ametricdescribingtheconsistencyorregularityofthelocal orientation of theimage) have been usedtoquantify the reflectedpattern,⁶fromwhichthetearfilmbehaviourwas inferred.Thecoherencemeasurementhasshownpromising clinical results for theclassificationof patientsymptoms⁹ and the quantification of tear film surface quality over time.⁶However,amoresensitive indicatorisstillrequired for cases in which the tear film is of poor quality. For thosecases,commonindryeye patientsandcontactlens wearers,¹⁰ thereflectedpatternloses muchofitsorienta- tionandbecomeslessstructured.Thus,thereisaneedto identifytechniquesthatprovideamoredetailedquantifica- tionoftheunderlyingbehaviourofthetearfilmsurface.

Several methods of image processing have been pro- posedfortheanalysisandcharacterizationofunstructured imagepatterns(suchastheonesseenundersuboptimaltear ﬁlm conditions),¹¹ including fractal dimension, run-length encoding, discrete wavelet transform, and co-occurrence matrices. Of those mentioned, texture analysis based on co-occurrencematricesisoneofthemostcommonlyused

(4)

Figure1 Asetof4representativevideokeratoscopyimageswithtearﬁlm-relatedinterference(irregularpatterns-duetotear thinningandbreak-ups).Notethattheareasintheimagecontainingshadowsfromtheeyelashesarenotconsideredinthisstudy.

TheseexampleimagesI---IVareusedtoexamineaspectsoftheperformanceoftheproposedimageprocessingtechnique.

toanalyzeandinterpret images.Texture analysisis atool whichprovidesasetofparameterstomeasurethevariation oftheintensityofasurfaceandtoquantifyimageproperties suchassmoothness,coarseness,andregularity.Inthispaper, thecapabilityofusingthesecond-ordertexturestatistics¹² toclassifyinterferencepatternsinvideokeratoscopyimages wasinvestigated.Inthisway,bysettingtheco-occurrence matrix to maximize the likelihood of separation between the signal and theinterference, it is expected toextract keyfeaturesthatmaymorecomprehensivelycharacterize anunstructuredpattern.Thus,theultimategoalofthepro- posedsetofimageprocessingtechniquesistoquantifythe Placidodiscregularity,inordertoprovideamoredetailed assessmentofthetearﬁlmthanprovidedbycurrentlyavail- abletechniques.

Methodology

Datacollectionandpreliminaryanalysis

Retrospectivedatafromastudyofthetearfilmusinghigh- speedvideokeratoscopy,capturedat25framespersecond, were used.¹³ The study cohort involved 22 subjects with normal tear film and 12 subjects diagnosed with dry eye syndrome.Only the right eyesof the subjects weremea- sured. Dryeye was diagnosedif the subjectexhibited all threeofthefollowingcharacteristicsofdryeye;significant subjectivedry eyesymptoms(McMonniestest score,>14), objectivesigns of tearfilm instability (FTBUT,<10s), and surfacestaining(cornealand/orconjunctivastainscore,>3 (NationalEyeInstitutegradingscale¹⁴)).Furtherdetailson thestudycanbefoundinRef.13.

For the initial phaseof this study,a number of repre- sentativeimagesandvideos weremanuallyselectedfrom thewholedataset(Fig.1),andanalyzedtoobtainthebest parameterselectionfortheproposedmethod.Thus,these dataderivedfromtheimagesinFig.1areusedtotestthe proposedmethod,whilein thesecondphaseofthestudy, thewholedatasetwhichcontainedmeasurementsfrom22 normaland12dryeyesubjectswasusedtoevaluatetheper- formanceofthe proposedtechniqueinassessing tearﬁlm surfacequality.

Thevideokeratoscopicgreyscaleintensitydigitalimage forms a 2D matrix I[n,m], with n=1, 2,...,N and m=1, 2,...,M.Inthevideokeratoscopeusedforthisstudy,Med- montE300(Melbourne,Australia),thematrixisN=648by M=572 pixels. Before the calculation of the Grey Level

Co-occurrenceMatrix(GLCM),anareaofanalysisismanually selected.Thisareacontainsthemaximumrectangulararea oftheimagethatisnotaffectedbytheshadowsfromeye- lashes.Toextractthisregion,whichavoidstheundesirable effectfromeyelashes, amethodology that haspreviously beenpresented indetail⁶ wasused.The keyfocusofthis paperistheseparationbetweentearfilmrelatedinterfer- encesandthesignal. The selectedareaisfurtherdivided intonon-overlappingblocksofequalsize.Theblocksizeis optimizedsoitcontainsasignificantamountofinformation without reducingthe spatial resolution whenlocating the potentialinterference.Inourcase,theimagewasdivided intosquareblocksof 25×25 pixels.Giventhat themaxi- mumwidthofaringisapproximately8.78pixels,¹⁵a25×25 pixelblockissufficienttoencompassasignificant amount ofinformation(atleasttworings)andtokeepsufficiently highspatialresolution.

Avideokeratoscopyimageiscomposedoftwoelements:

thebackground (theanterioreye---iris andpupil)andthe foreground(thereﬂectedpatternofthePlacidodiscrings).

Thebackgrounddoesnotprovideanyessentialinformation fortheanalysisofthepattern.Forthatreasonitisimportant toremoveitfromtheimagetoavoidmisleadinginformation infurtherprocessing.Thiscanbeachievedwithstatistical blocknormalization, in which theintensity informationis normalizedtozeromeanandunitvariance.¹⁵

GLCMparameterselection

TheGreyLevelCo-occurrenceMatrix(GLCM),proposedby Haralicketal.,¹²p(d),foranimageisthejointconditional probabilitydensityfunction(atwodimensionalhistogram), p(i,j|ϕ(d,));whereϕ(d,)isavectoroflengthdandangle

measuredfromanelementintheimage.Thejointcondi- tionalprobabilitydensityfunctionrepresentstheprobability ofﬁndinggreylevels,iandj,atthebase andtermination pointsofthevectorϕ(d,)foranypixelwithintheimage.

Fig. 2 shows an example of the GLCM (for d=1 and

=0).Therefore,inthisexampletheGLCMcalculateshow oftena pixel withgrey-level (grey-scale intensity) valuei occurshorizontallyadjacenttoapixelwiththevaluej.In Fig.2,twodifferentsectionsfromavideokeratoscopeimage arepresented: anoptimal (top)andsub-optimal (bottom) reﬂectedpattern.Asanexample,theoccurrencevaluesat greylevel(3,4)(i.e.,grey level3at thestart ofthe vec- torand 4 at the end) are shown in the images (with red lines)and its corresponding occurrence/cumulative GLCM

(5)

ARTICLE IN PRESS

Videokeratoscopic image

tear film break-up [left box] stable tear film [right box]

Example image -- stable tear film

Example image -- tear film break-up Co-occurrence matrix -- tear film break-up Co-occurrence matrix -- stable tear film

gray-scale

level Gray-scale level at j

Gray-scale level at j

Gray-scale level at iGray-scale level at i

occurence level level (3,4)

acculumated value = 7 level (3,4) acculumated value = 10

entropy: 4.13 energy: 0.0224

entropy: 2.9 energy: 0.117 30

20

10

0

0 150

100

50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

1 2 3 4 5 6 7 8 9 10111213141516

1 2 3 4 5678 9 10111213141516

16 14 12 10 8 6 4 2

Figure2 Tworepresentativeblocks(centre)fromavideokeratoscopicimage(left)withtheircorrespondingco-occurrencematri- ces(right).Thetopsubplotsrepresentanexamplewithoptimalreflectionconditions(goodtearfilm)whilethebottomsubplots representsuboptimalreflectionconditions(poortearfilm).Theredlinesinthecentralblockimageillustratethepixelsineachof theblockimageswithoccurrencevaluesofgreylevelequalto(3,4)(seetextfordetails).

value.Thehighoccurrencelevelaroundthe(5,5)greylevel corresponds to the background. In general, for the opti- mallyreﬂectedpattern,extrainformationcanbeseeninthe GLCM.FromtheseGLCMsasetoftexturestatisticaldescrip- torscan beextracted whichrepresent globalmeasures of thedata.Tworepresentativefeaturesarepresentedhereas anexample.TheEntropy(measurestherandomnessofthe distribution)andtheEnergy(measurement ofthe texture uniformity,theoppositeofentropy)areprovidedinFig.2 (rightside),andasexpectedbaseduponthedeﬁnitionsof thesemetrics,bothvaluesbehaveinopposingwaysinthese twoexamples.Therestofthesetoftexturefeaturesused inthisstudyaredescribedindetailinTable1.

ArangeofdifferentGLCMscanbegeneratedfromeach imagedependingonthechoiceofvectorlength(d)andori- entation(),whichshouldbeselectedinawaytomaximize thelikelihoodofseparatingtheinterferencepattern from thesignal.Usingvisuallyrepresentative imagesfromeach patterntype,variouschoicesforvectorlengthandorienta- tionwereexamined.Althoughtheexamplesprovidedhere areforareducednumberofimages,theyrepresentatypical setofimagesthatcanbeencounteredinvideokeratoscopy.

The effectof the level of quantization(Ng),which is the specific number of grey-levels toscale the image, is not considered inthis paper.In general,the loweris the val- uesofNg,thelowertheclassificationaccuracy.Theimages fromthevideokeratoscopehavea256grey-scalelevel,however choosing Ng=64 keeps the computational load low, while ensures that the classificationaccuracy will notbe affected.¹⁶

Features

Once the GLCM is generated, a set of texture statistical descriptorscan beextracted.InHaralick’sinitial study14 featureswere described.¹² This set has been extended in

otherworks.^17,18Throughapreliminarytrial,thosefeatures thatappearedtobethemostsensitivefortheclassiﬁcation ofthedifferentvideokeratoscopytextureswereidentiﬁed.

Following thenotationprovidedinHaralicketal.¹² deﬁni- tions of these features along with a brief explanation of the aspectof the imagethat theyquantifyarepresented inTable1.

Statisticalanalysis

Thereceiveroperatingcharacteristic(ROC)curvesarecom- monly used to measure the performance of a two-class classiﬁcation. The ROC curve provides the sensitivity and speciﬁcity of the measurement in separating the interference from thesignal. Eight different ROCcurves were calculated to show the capability of each of the texture features listed in Table 1 to discriminate between interference and signal pattern in the videokeratoscopic images.

Oncetheimageisdividedintoblocks,theGLCMandthe featuresareextractedforeachoftheblocks.Inordertocal- culatetheROC,thedataofthefeaturesfromeachofthe twogroups ofpattern(i.e., signalandinterference) were usedtoestimatetheprobabilitydensityfunction.Thiswas achievedbyusingkerneldensityestimatorwithanEpanech- nikov window.¹⁹ The ROC curves were then numerically evaluatedfromeachsuchpairofnon-parametrickernelden- sity estimators. Fromthe ROC curve, the area under the curve(AUC)wascalculatedusingnumericalintegration.The AUCis normallyusedasameasure oftheperformance of howwelldetectioniscarriedout,valueofAUC=1indicates an idealdetection,while avalueofAUC=0.5corresponds torandomclassiﬁcation.Additionally,thisapproachutiliz- ing ROC curve analysis was also applied to examine the effectiveness of each of the texturefeatures in discrimi- nating between normal patients and those with dry eye.

(6)

Table1 DeﬁnitionofthetexturefeaturesusedinthisstudyalongwiththeequationappliedtotheGreyLevelCo-occurrence Matrix(GLCM).

Featuredescription Equation

Energy:alsoknownastheangularsecondmoment,measuresthe textureuniformity.Highvaluesofenergyareexpectedwhenthe texturepresentsagreyleveldistributionthatiseitherconstantor periodic

Energy=

i

j

{p(i,j)}²

ClusterShadeandClusterProminence:proposedbyConnersetal.¹⁷ toemulatehumanperceptualbehaviour,theygiveameasureof thedegreetowhichtheoutliersinthehistogramfavouroneside oranotherofthestatisticalmean

ClusterShade=

i

j(i+j−x−y)³p(i,j) ClusterProminence=

i

j(i+j−x−y)⁴p(i,j)

Entropy,DifferenceEntropyandSumEntropy:measuresthe randomnessofagreyleveldistribution.TheEntropyisexpected tobehighifthegreylevelsaredistributedrandomlythroughout theimage.Theenergyandtheentropyareinverselycorrelated.

Theadvantageofusingtheenergyisitsnormalizedrange.Where p_x+y(i)andp_x−y(i)are,

p_x+y(k)=

Ng

i=1 Ng

j=1

p(i,j), i+j=k,k=2,3,...,2Ng

p_x−y(k)=

Ng

i=1 Ng

j=1

p(i,j), |i−j|=k,k=0,1,...,Ng

Differenceentropy=−

Ng−1

i=0

px−y(i)log{px−y(i)} Sumentropy=−

2Ng

i=2

p_x+y(i)log{px+y(i)}

Homogeneity:measuresthelocalhomogeneityofapixelpair.The valuesareexpectedtobelargeifthegreylevelsofeachpixelpair aresimilar

Homogeneity=

i

j 1 1+(i−j)²p(i,j)

Maximumprobability:resultsinthepixelpairthatismost

predominantintheimage.Thevalueisexpectedtobehighifthe occurrenceofthemostpredominantpixelpairishigh

Maximumprobability=MAXi,jp(i,j)

Theequationp(i,j)indicatesthe(i,j)thelementoftheGLCM,Ngisthenumberofgreylevelsintherawimageandx,y,x,

y,are,respectively,themeanandstandarddeviationsfortherowsandcolumnsoftheGLCM:¹²

x=

i

j

i.p(i,j),x =

i

j

(i−x)².p(i,j),y =

i

j

j.p(i,j),y

i

j

(j−y)².p(i,j)

Additionally to the AUC,the sensitivity and speciﬁcity of eachtexture-featuretodetectdryeyewasalsocalculated.

Results

Statisticalperformance:signalversusinterference separation

Thissectionexaminestheoptimumselectionofthevector lengthanditsorientationforthetexturefeatureanalysis, insuch awayastomaximizethelikelihoodof separating theinterferencepatternfromthesignal,usingasetofrep- resentativeimages(imagesI---IVinFig.1).

Orientation

Itisimportanttodifferentiatebetweenthepattern’sdirec- tionalityandtheGLCMorientation.Intheexampleshown in Fig. 2 (top), the pattern’s (Placido disc) directionality isvertical whilefor thisexample theGLCM orientationto

create the co-occurrence matrix was chosen horizontally (red lines in the image). The reﬂected videokeratoscopy pattern has directionality associated with it; hence it is important to ensure that the GLCM is not biased by this pattern’sdirection. Thus the vector’s orientation used to createGLCMshouldbechosentoensurethatthestatistics generated from the GLCM are rotationally invariant. To reduce the angular variance of the statistics, the GLCM canbegeneratedusingmultiplevectorsatthesametime, sothat each pixel contributes multiple times tothe ﬁnal GLCM. Fig. 3 shows the orientation dependency of four differentfeaturesforfourdifferentvectororientations.As thenumberoforientationsusedtocreatedGLCMincreases, the separation of statistics (mean) between signal and interference increases, hence the separation between thesetwoclassesstrengthens.Thismeansthattoproperly characterize theGLCM matrix differentorientations need to be considered. This effect is particularly clear when three or four orientations are used simultaneously (see Fig.3).Forourapplication,discreteanglestheta=[0^◦,45^◦,

(7)

ARTICLE IN PRESS

0.5

0.4

0.3

0.2

0.1 1 2 3 4

Number of orientations

Number of orientations Number of orientations

Number of orientations

Energy

Signal Interference

Signal Interference Signal Interference

Homogeneity

0.8

0.7

0.6 1 2 3 4 1 2 3 4

1 2 3 4

0.7

0.6

0.5

0.4 2500

1500

1000 2000

Maximum probabilityCluster prominence

135.° 90.° 45.°

Pixel of 0.°

interest

Figure3 Class statistics, mean±onestandard deviation,offour differenttexture features as afunction of thenumber of orientationsconsidered.1=[0^◦],2=[0^◦,45^◦],3=[0^◦,45^◦,90^◦],4=[0^◦,45^◦,90^◦,135^◦].Thecentralinsetshowsadiagramwiththe relationbetweentheangleandtheadjacentpixels.

90^◦, 135^◦] were selected, thesevalues correspond tothe adjacentpixelsasshownintheinsetofFig.3.Asshownin theexample,theadvantageofusingthisapproachisthat itavoidstheangular-dependencyofGLCM.

Vectorlength

Similar to the orientation parameter, the choice of the vector’slength has a substantial impact on the feature’s

statistics.Fig.4 shows theinﬂuence of thelength onthe class statistics. A vector length of 1 pixel is the optimal optioninthiscaseandshowsgoodseparationbetweenthe classes as well asbeingthe shortest possible length.The shorterthelength,themoreentriesgointotheGLCM.

In orderto evaluate the performance of the proposed technique the algorithm was tested ona number of rep- resentative videokeratoscopy images,shown in Fig. 1, as

0.4

0.3

0.2

0.1

00 2 4 6 8 10 12 0 2 4 6 8 10 12

0 2 4 6 8 10 12

Homogeneity

Vector length (pixel) Vector length (pixel)

Signal Interference

Signal Interference Signal Interference

0.8

0.7

0.6

0.5

0.4

Energy

2000

1500

1000

500

Cluster prominenceMaximum probability

0

0.7 0.6 0.5 0.4 0.3 0.2

Figure4 Classstatistics,mean±onestandarddeviation,offourdifferenttexturefeaturesasafunctionofthevector’slength.

(8)

Table2 AreaunderthecurvefortheeightconsideredtexturefeaturesforthefourvideokeratoscopicimagesfromFig.1and thevideosequence.Theboldnumbersindicatethebestfeaturewithineachcase(highestAUC).

Areaunderthecurve Feature Energy Cluster

Shade

Cluster Prominence

Entropy Diff.

Entropy

Sum.

Entropy

Homogeneity Max.

probability

CaseI 0.80 0.77 0.79 0.88 0.84 0.85 0.80 0.79

CaseII 0.96 0.93 0.92 0.97 0.92 0.97 0.87 0.95

CaseIII 0.83 0.77 0.86 0.86 0.83 0.83 0.81 0.82

CaseIV 0.99 0.95 0.98 0.99 0.98 0.99 0.94 0.98

Video 0.94 0.94 0.93 0.95 0.88 0.92 0.93 0.93

Average 0.90 0.87 0.90 0.93 0.89 0.91 0.87 0.89

wellasonavideo sequence.The video, composedof330 frames,correspondstoafullinterblinkintervalofadryeye patient.The detailedresults fromthe ROCcurve analysis for the performance of each of the features are givenin Table2.Theareaunderthecurve(AUC)isprovidedforthe eightdifferentconsideredfeatures.Itisworthnotingthat the entropyfeature resultsin the largest AUC for all the cases,henceseparating moreeffectivelytheinterference fromthe signal. Overall, none of theconsidered features performs belowAUC=0.77. Ofspecialinterest is thehigh AUCvaluesobtainedonthevideosequence.

Clinicalperformance

Inordertoevaluatetheperformanceoftheproposedtech- niquetodiscriminatenormalsubjectsfromthosewithdry eye, data collected in aprevious clinical study¹³ from 34 subjects(22normaland12dry eyes)wereused.The data wereacquiredundersuppressedblinkingconditions(SBC).

IntheSBCs,threemeasurementsweretaken.Subjectswere askedtoblinkseveraltimesbeforethebeginningofthemea- surementandthenfocusontheinstrument’sﬁxationtarget andkeeptheireyesopen aslongastheycould.Themaxi- mumtimeoftherecordingsequencewas30s anda3-min break wasgiven between measurements. More details on themeasurementprotocolaregiveninRef.13.

The subject’s data corresponding to each blink was croppedtoaspecificlength.Thefirst5sofeachrecording was used, excluding the first second after the blink to avoid the effect of tear film formation/spreading on the estimatesofTFSQ.The5sintervalensuresthat100%ofthe subjectswereincludedineach groupintheanalysis(i.e., allsubjectshadaminimumof5sor moreofdatapriorto ablinkintherecording).Thethreerecordingspersubject

were averaged into one single measurement. Receiver operating characteristic (ROC) curves were calculated to showthe capability ofthe considered texturefeatures to discriminate dry eye from normal tear ﬁlm subjects. To arriveat the bestpossible detectionperformance for the non-invasivetest,agroupofROCcurveswereevaluatedat differentsamplingintervalsbetweenthetimeintervalfrom the1stsecondtothe6thsecond.Thismayprovideanopti- maltimethresholdintermsofdiscriminationperformance.

A detailed result of each of the features performance in terms of area under the curve (AUC), speciﬁcity and sensitivityaswellasitsoptimaltimethresholdisgivenin Table3.Theentropyandhomogeneityfeaturesresultedin thelargestAUC,inotherwordsthebestdryeyedetection.

For all the considered texture features, the AUC value rangefrom0.77to0.82,whilethesensitivityshowsvalues above0.9andthespeciﬁcityshowsvaluesaround0.6.

Discussion

We have investigated the applicability of an image processing technique, textural feature analysis, for the classiﬁcation and detection of interference in videoker- atoscopicimages.Inthemethod greylevelco-occurrence matrix(GLCM)wasusedtoextractasetoftexturalfeatures, in which theGLCM wasoptimally calculatedto maximize thelikelihoodtoseparatethesignalandinterferenceinthe videokeratoscopyimages.TheresultsfromtheROCanaly- sis,withAUCvaluesrangingfrom0.77to0.82,demonstrate the capability of these features to discriminate between signal andinterference in static images.Additionally, the goodperformance,particularlyofspeciﬁctexturalfeatures, such as entropy that exhibited an AUC of 0.82 (sensitivity=0.98,sensitivity=0.64)forthediscriminationbetween

Table3 Theareaunderthecurve(AUC),sensitivityandspeciﬁcityforthedifferenttexturefeatures,showingtheperformance ofthetechniquetodiagnosedryeye.Theoptimaltimethreshold(OTT)indicatesthetimeofoptimaldiscrimination.

Energy ClusterShade Cluster Prominence

Entropy Diff.Entropy Sum.Entropy Homogeneity Max.

probability

AUC 0.77 0.71 0.78 0.82 0.72 0.74 0.82 0.81

Speciﬁcity 0.56 0.51 0.59 0.64 0.50 0.54 0.65 0.62

Sensitivity 0.94 0.90 0.95 0.98 0.90 0.92 0.98 0.97

OTT(s) 3.4 3.2 1.0 3.4 4.6 1.3 5.0 3.4

(9)

ARTICLE IN PRESS

normalanddryeyepatients,indicatesthepotentialofthis technique to assess the tear film quality. These findings suggestthatthevaluesderivedfromthevideokeratoscopy images using the texture feature analysis could provide relevant clinical informationregarding the quality of the tearfilmsurface.Wehavepreviouslyexaminedtheability ofdynamicvideokeratoscopyimagesforthediscrimination ofnormalanddryeyepatients,¹³andfoundthatanalysisof theimagesusing a patterncoherence approach exhibited good discriminative ability for the detection of dry eye (AUCof 0.72).Theslightly greaterAUCfoundfor someof the textural features examined in this manuscript (e.g., entropy)suggeststhatimageanalysisusingtexturalfeatures mayprovidebetterdiscriminationcomparedtootherimage analysis approaches for the analysis of videokeratoscopy images from dry eye patients. Thus with the proposed softwarealgorithm,theclinicalcornealtopographercanbe usedtoprovidefurtherinformationregardingtheintegrity of the tear film, potentially contributing to improved detection of dry eye. However, we recognize that the detectionandmonitoringofdryeyesyndromeisacomplex clinical issue, which cannot rely solely onmeasures from a single instrument. Therefore we anticipate that this methodcouldbeoneof abattery oftests thatis usedto facilitatetheclinicaldiagnosisandmonitoringofdryeye.

Other videokeratoscopy analysis techniques have been previouslyproposedin the literature.However,theyhave typicallybeenbasedupontheassessmentoftearﬁlmbreak- uptimeratherthanthespeciﬁc analysisof thequalityof thePlacidodiscpatternprovidedinourproposedmethod.

Taking into consideration the difference between studies and the principle of the techniques, the methods in the currentpaperappeartobecomparabletopreviouslydevel- opedmethods, suchasMengheretal.²⁰ (specificity=0.83, sensitivity=0.85)and Gotoetal.²¹ (specificity=0.98,sensitivity=0.63), which also showed a good discrimination power.We areyet toextractthemeasurementsofbreak- uptimefromdynamic-areaHSV,sincetheaimofthisstudy wastousethesurfacequalityindicatorasameasurementto discriminatebetweendryeyeandnormaltearfilmsubjects.

In general, from the set of texture statistical descrip- torsthathavebeenconsidered,itappearsthattheentropy descriptor achieves the best AUCvalues for the majority oftheexamples.Thisresultwasexpectedbecauseentropy representsameasureoftherandomnessofthedistribution oftheimage,whichchangesasthetearfilmbecomesthin- nerandthereflectedpatternbecomesmoreunstructured, aspreviously shown in Fig.2. Additionally, the technique couldalsobeusedtoextractinformationaboutthetearfilm kinetics,expanding theflexibilityoftheproposedmethod toextractrelevantclinicalinformationsuchasestimatesof thebreak-uptime.Fig.5showsanexampleoftheestimated entropyvaluesfroma representativevideo sequence,the subjectpresentedabuild-uptimeof0.6sandthetearfilm startedto deteriorate5.7s after theblink. In the plot,a clearexampleofthethree-phasemodeloftearfilmkinet- ics can be seen.⁹ In the first phase following the blink a shortbuild-uptime(tearspreadsovertheocularsurface)is observed,thenastableperiodofabout 5s andfinallythe tearfilm thinning and deterioration phases areobserved.

An example of the image region analyzed in the kinetics calculationisgivenwithintheplot.

Entropy [A.U.] (i) Tear film build-up phase

(ii) Tear film thinning

(ii) Tear film stability

Time [seconds]

2 4 6 8 10 12 14

4.5 5 5.5 6

Entropy value 3 phase fitting

Figure5 Athree-phaseexampleoftearﬁlmsurfacekinetics followingablink.

Detecting interference patterns in videokeratoscopy imagesandidentifyingkeytexturalfeaturestocharacter- izethisinterference canbeuseful inquantifying thetear film surface behaviour, especially in dry eye patients and contactlenswearers,inwhichthestabilityofthepatterns orientation isknowntobepoor. The proposedtechniques provide us with tools that can be used to extract useful informationfromavideokeratoscopyimagefortheassess- mentofthetearfilmsurfacequality;however,furtherwork isneededtounderstandthereliabilityandrepeatabilityof thetechnique.Additionally,thetechniquecouldbeapplied to extract tear film kinetics information and for dry eye detection, which demonstrates the flexibility of the proposed techniquetoquantify videokeratoscopyimagesand obtainrelevantclinicalinformation.

Conﬂicts of interest

Theauthorshavenoconﬂictsofinteresttodeclare.

Acknowledgments

DS-I was supported by the Young Cadre 2015 Plus, 40/PD/2012---theprojectwasco-ﬁnancedbytheEuropean Union and the Polish state budget through the European SocialFund.

References

1.KhanalS,TomlinsonA,McFadyenA,DiaperC,RamaeshK.Dry eyediagnosis.InvestOphthalmolVisSci.2008;49:1407---1414.

2.Mengher LS, Pandher K, Bron A. Non-invasive tear ﬁlm break-uptime: sensitivity and speciﬁcity. Acta Ophthalmol.

2009;64:441---444.

3.NémethJ, ErdélyiB, CsákányB, et al. High-speedvideoto- pographic measurement of tear ﬁlm build-up time. Invest OphthalmolVisSci.2002;43:1783---1790.

4.Montés-Micó R, Alió JL, Mu˜noz G, Charman WN. Temporal changesinopticalqualityofair---tearﬁlminterfaceatanterior corneaafterblink.InvestOphthVisSci.2004;45:1752---1757.

5.Alonso-CaneiroD,IskanderDR,CollinsMJ.Estimatingcorneal surface topography in videokeratoscopy in the presence of strong signal interference. IEEE Trans Biomed Eng.

2008;55:2381---2387.

(10)

6.Alonso-CaneiroD,IskanderDR,CollinsMJ.Assessmentoftear ﬁlmsurfacequalityusingdynamic-areahigh-speedvideokeratoscopy.IEEETransBiomedEng.2009;56:1473---1481.

7.IskanderDR,CollinsMJ,DavisB.Evaluatingtearﬁlmstability inthehumaneyewithhigh-speedvideokeratoscopy.IEEETrans BiomedEng.2005;52:1939---1949.

8.KassM,WitkinA.Analyzingorientedpatterns.ComputVision Graph.1987;37:362---385.

9.Szczesna DH, Iskander DR. Lateral shearing interferometry for analysis of tear ﬁlm surface kinetics. Optom Vis Sci.

2010;87:513---517.

10.Szczesna-IskanderDH,IskanderDR,ReadSA,Alonso-CaneiroD.

Noninvasiveinvivoassessmentofsoftcontactlenstypeontear ﬁlmsurfacequality.InvestOphthVisSci.2012;53:525---531.

11.RaoAR,Chairman-JainR.Ataxonomyfortexturedescription andidentiﬁcation.NewYork:Springer-Verlag;1989.

12.HaralickRM,ShanmugamK,DinsteinIH.Texturalfeaturesfor imageclassiﬁcation.IEEETrans SystManCybern.1973;SMC- 3:610---621.

13.SzczesnaDH,Alonso-CaneiroD,IskanderDR,ReadSA,Collins MJ. Predicting dry eye using noninvasive techniques of tear ﬁlm surface assessment. Invest Ophthalmol Vis Sci.

2011;52:751---756.

14.LempA.ReportoftheNationalEyeInstitute/Industryworkshop onclinicaltrialsindryeyes.CLAOJ.1995;21:221---232.

15.Alonso-Caneiro D, Iskander DR,Collins MJ.Computationally efﬁcientinterferencedetection invideokeratoscopyimages.

In:IEEE10thWorkshoponMultimediaSignalProcessing,vol.1.

2008:486---490.

16.ClausiDA.Ananalysisofco-occurrencetexturestatisticsas a function of grey level quantization. Can J Remote Sens.

2002;28:45---62.

17.ConnersRW,TrivediMM,HarlowCA.Segmentationofahigh- resolutionurbansceneusingtextureoperators.ComputVision Graph.1984;25:273---310.

18.HaralickRM.Statisticalandstructuralapproachestotexture.

ProcIEEE.1979;67:786---804.

19.SilvermanBW.DensityEstimationforStatisticsandDataAnal- ysis.vol.26.London:Chapman&Hall/CRC;1986.

20.Mengher LS, Bron AJ, Tonge SR, GilbertDJ. A non-invasive instrumentforclinicalassessmentofthepre-cornealtearﬁlm stability.CurrEyeRes.1985;4:1---7.

21.Goto T, Zheng X, Okamoto S, Ohashi Y. Tear ﬁlm stability analysissystem:introducinganewapplicationfor videoker- atography.Cornea.2004;23:S65---S70.