surface55556.com

IFIPInternationalFederationforInformationProcessing2015L.
M.
Camarinha-Matosetal.
(Eds.
):DoCEIS2015,IFIPAICT450,pp.
239–247,2015.
DOI:10.
1007/978-3-319-16766-4_26HighResolutionDigitalTissueImageProcessingUsingTextureImageDatabasesGáborKiss1(),OrsolyaEszterCseri1,dámAltsach1,IstvánImreBándi1,LeventeKovács1,andMiklosKozlovszky1,21budaUniversity/BiotechKnowledgeCenter,Budapest,Hungary{kiss.
gabor,cseri.
eszter,altsach.
adam,bandi.
istvan,kovacs.
levente,kozlovszky.
miklos}@nik.
uni-obuda.
hu2MTASZTAKI/LaboratoryofParallelandDistributedComputing,Budapest,HungaryAbstract.
Texturebasedimagedatabasesintegratedwitheffectivesearchingalgorithmsareusefulsolutionsformanyscientificandindustrialpurposes.
Medicalimageprocessingofhighresolutiontissueimagesisoneoftheareas,wherethecell/tissueclassificationcanrelyonsuchsolutions.
Inthispaperwearedescribingthedesign,developmentandusageofaspecializedmedicaltex-tureimagedatabase.
OurprimaryaimwiththistexturedatabaseistoprovideDigitalImagingandCommunicationinMedicine(DICOM)compatibletextureimagedatasetforcell,glandandepitheliumclassificationinhistology.
Ourso-lutionincludesaPictureArchivingandCommunicationSystem(PACS)sub-system,whichismainlyprovideacommunicationinterface(textureimagesearchingandretrieval)andenablesimageprocessingalgorithmstoworkmoreeffectivelyonhighresolutiontissueslideimages.
InthispaperwedescribehowourLocalBinaryPattern(LBP)basedalgorithmbenefitstexturedatabaseusagewhensolvingimageprocessingproblemsinhistologyandhistopathology.
Keywords:PACS·DICOM·LBP·Texturebasedimagedatabase·Medicalimageprocessing·Digitalmicroscopy1IntroductionWiththedevelopmentofhighresolutionscanningandinformatics,thepractitioner'sworkflowhaschanged,andhighresolutiondigitalmicroscopybecomespossible.
Slidesharing,indigitalworldisjusteasyasforwardingimagesviaInternet,alsovisualizationandarchivingtheimagesbecomesimpletasks.
Unfortunately,themi-grationintodigitalworlddidnotsolveallproblems,andslideassessmenttasksstillremaincomplexandtimeconsumingintheworkflow.
Practitioners-independentlyofwhethertheyusetraditionalmethodsornewdigitalslides-,aresharingthesamelevelofdifficultiestodocancergradinganddetermineefficienttreatments.
Inmanycountries(duetothelargeamountofsamples)patientsarewaitinginlongqueuesfortheirdiagnosisresult.
Suchwaitingqueuesneedtobeavoided,becauseduringwait-ingtime,healthproblemsmaygetworse.
240G.
Kissetal.
2StateoftheArtWehavedoneamarketscreeningoftheavailablehighresolutiondigitalmicroscopyimageassessmentsoftwaresolutions.
Ourevaluatedsystemswerethefollowings:theTissuemorphDP[1]imageanalyzerfromVisiopharm,whichisdesignedmainlyforthepopulationofnucleiexamination,HER2-CONNECT[2]targetingtumorcellsdetection,Virtuoso,DeveloperXD[3]andtheMediaCybernetics[4]doingimageoftissueanalysis.
Wehaveexaminedalsosomeimagedatabasesnamely:ASHImagebank[13],Im-ageAfter[14],TinEye[15],RevIMG[16].
TheASHImagebankisahematologydatabasewithkeywordsearch.
ImageAfterisatexturebasedapproachwiththestring-basedsearchmethod.
FinallyTinEyeandRevIMGareContentBasedImageRetrievalsystems(CBIR).
Alltheexaminedsystemssupportonlyfullsizeimagesinsteadofplaintextureimages.
MorphCheckisourmedicalimageanalysisplatformworkingonhighresolutiondigitalmicroscopyimages.
Itenablestheloadofhighresolutiontissueimages,auto-maticallydetectsthebasicstructuresoftissue(nuclei,glands,surfaceepithelium)andprovidesquantitativeandqualitativeparametermeasurementsRecentlywehaveadd-edtexturebasedimageanalysistoimproveitsobjectrecognitionaccuracy.
Theuserscandotexture-basesearchandobjectrecognitionwithoutmanualselection.
MostoftheimageacquisitionmodalitiesareusingtheDICOMstandardcombinedwithPACS.
WealsohaveenabledthesupportofaDICOM/PACSbasedmedicaldatabasesystem[5]inoursystem.
DICOMandPACStechnologiesguaranteethestandardstorageformedicalimagesandthesoftwaremakestheadditionalextensionofitsabilitieslater.
PACSserverisbeingusedbyvirtualserversinacloud.
WeareusingIaaS(InfrastructureasaService),becausewecanscaletoanyamountofcom-puterresourcesdynamicallybasedonusageandotherparameters.
ThescalingprocessissupportedbytheEC2compatiblecloudmiddleware.
Thesystemmeasuresalargenumberofmorphologicalparametersonthedetectedimageobjectsandstoretheminthedatabase,allhigherlevelanalysisanddecisionisbasedonthearchiveddatabaseparameters.
Imageprocessingandclassificationtasksdemandhighcomputingcapacity,thusthealgorithmsareabletorunonvarioushardwarearchitecturesandsupportGPUstoenablehigherperformance[6],[7].
Ouraimistoincreasetheefficiencyofthesoftwarewithnewtexture-basedapproaching.
3ArchitectureMorphCheckconsistsoftwowellseparableparts.
Oneofthesepartsistheplainalgo-rithmicmethodsworkingonprocessinganddetection.
Theotherpartisusingthetexture-basedapproachdoingtheimageprocessingtasks(showninFig.
1.
)Theusermaydefineworkflowsandfreelyuseacombinationofbothpartsoftheimageprocessingmethods.
Aworkflowcontainstasks(e.
g.
:texturebasedandclassifieralgorithms)andcanbearbitrarycomplex.
Thesystemalsosupportsmanualimageannotation,wherethepathologistmakestheannotationofthetissuesampleandupl-oadsitintothedatabase.
TheannotatedimageswillbestoredinthePACSsystem,andalgorithmscanreuselaterontheirdataduringclassificationprocesses.
HighResolutionDigitalTissueImageProcessingUsingTextureImageDatabases241Fig.
1.
High-levelsystemarchitectureofMorphCheck4TextureBasedSearchAlgorithmOurimplementedLBPalgorithmisbasedontheLocalBinaryPattern[9](showninFig.
2.
andEq.
1).
ThevalueofLBPcodeofapixelisgivenby:1,0;0,(1)Fig.
2.
partAshowsasampleimagewithintensitypixelsandthecenterpixelisbeingmarkedbybluecolor.
Fig.
2.
partBshowsthedifferencesbetweeneachpixelsandthemarkedcenterpixel,andthedifferencesarebeingwrittentotheplaceoftheoriginalpixel.
Forexample:Thecenterpixelis70andthetopleftpixelis47.
Thedifferenceis-23,anditiswrittentothetopleftpixelplaceatthepartBoftheFigure2.
ThelastcalculationisbeingbasedbyEq.
1.
TakeeachpixelonBimageandifthecurrentpixelgreaterthanorequal0,atCimagethesamepixelwillbemarkedas1.
Otherwiseitwillbewritten0.
FinallytheLBPcodecanberedbytheclockwisedirection.
Fig.
2.
LocalBinaryPattern.
A)Sampleintensitypixelswithcenterpixel,B)Differencesfromcenterpixel,C)LBPcodeinbinarynumberanddecimalsystem.
242G.
Kissetal.
Thetaskistoseparateandclassifythenaturaltexturesontheimages.
Theinputofthealgorithmistheareawheretheexaminationhappens,andthoseclassesshouldbelateridentifiedonimages(morepicturesmaybelongtooneclass).
AtfirsttheSimpleLinearIterativeClustering(SLIC)algorithmslicesimagesintopiecesoftexture[10],[11].
IntheSLICalgorithmitisnecessarytosettwoparame-ters:thesizeofthesuperpixelandhowuniformitshouldbe.
Thesecondparametermeanshowhomogenousthecolorsshouldbeinthesuperpixel.
Athighparameterswegetinhomogeneous,compactandnotrandomsizesuperpixels(showninFig.
3.
).
Fig.
3.
AhighresolutiontissueimagesplitintosuperpixelsAfterthesestepsweexamineallsuperpixelintermsofbelongintoaclass.
Thisprocesshappenstothehelpoftwoparameters:Simplehistogramandrotationinva-riantLocalBinaryPatternhistogram.
Withbothparameterswecalculatedistancemeasureforallimages.
ForthecaseofhistogramwecalculatethedifferencewiththeEarthMover'sDistance[4]metric.
InthecaseoftheLBPhistogramwecalculatetheabsolutedifferenceofbothhistograms.
BasedonLBPtoallsuperpixelwecalculateaclassification,afterthatwecompareallsuperpixelsclassificationwithother.
Therobustclassifiedsuperpixelsarewhichsuperpixelsclassifiedsameclass,onthebasisofbothhistograms(LBPandsimple),allothersareindeterminatesuperpixels.
Cur-rently,toavoidthefalsenegativehitsweclassifythemtotheunclassifiedclass,butlaterwiththestableneighborswewillassignrankstothem.
5MeasurementandTestEnvironmentTomeasureimageobjectclassificationefficiency,wehaveadaptedsomemetricsfromtheliterature.
Itisnecessarytointroducesomeconceptstothis:Referenceresultset:Thepixels,whicharemarkedinthereferenceimagebyaskilledexpert.
Testresultset:Thepixels,whicharemarkedbythealgorithms.
Theselectedpixel-basedparametersfromtheliteratureare[12]:Falsenegative(FN):Referenceresultsetcontainsthecurrentpixel,butthetestresultsetdoesnot.
Falsepositive(FP):Thetestresultsetcontainsthecurrentlypixel,buttheexpertisnotmarkeditonthereferenceresultset.
HighResolutionDigitalTissueImageProcessingUsingTextureImageDatabases243Truepositive(TP):Thepixeliscontainedinbothreferenceresultsetandtestresultset.
TrueNegative(TN):Neithersetcontainsthecurrentpixel.
Accuracy(TP+TN)/(TP+TN+FP+FN):Thisisthemeasurebetween0%and100%,where100%meansthereferenceresultsetandthetestresultsetareequal.
Thusthedoctorandalgorithmsmarkeddifferentpixelsontheimages.
Recall(TP/(TP+FN)):Thisisameasurenumberindicateshowmuchitfoundfromthereferenceresultset.
Iffoundallofthem1,elseconvergetozero.
Precision(TP/(TP+FP)):Measurementofthehit.
Ifthealgorithmmarkedallpix-elsmarked,theprecisionvalueis1,elseconvergetozero.
6PerformanceandAccuracyInthissectionweshowtheperformanceofourLBPbasedalgorithm,accordingtonextaspects:runtime,accuracy,recall,precision.
Weuseanonymtissuesamplestodotheefficiencymeasurements.
Tothetestsweuse20digitalsamples,whichconsistof50ROI(RegionofInterest).
Fromamongthese32ROIwerehealthy,18unheal-thy.
Table1containsallthedatawhichhasbeenproducedbythetests.
Fig.
4.
ResultofaccuracyexaminationWeruntheexaminedalgorithmsunderthetests,thenwehavecalculatedonallimagesthedefinedmetrics(TP,TN,FP,FN).
TheaccuracyisbeingshownbytheFig.
4.
andTable1.
Thismetricofourimplementedalgorithmisintherangeof64.
3-87,3%andtheaverageis75,99%.
Theotherimportantderivedresultisthesocalled"recall"(showninFig.
5.
).
Mixedresultsariseinthecourseofthetests.
TheLBPbasedalgorithmtriestoavoidthefalsepositivehits.
Theareawillbemarkedasundeterminedifthereissmallcon-fidence.
Thisisthemainreasonoflowrecallsinthetestresult.
Thethirdexaminedparameterwastheprecision(showninFigure6.
).
Ourimple-mentedalgorithmisabletofindpixelsbetween0,6and0,8precisiononallreferenceimages.
Amongtheexamined50test,thefoundedpixelsin24caseswerecorrect,whichmeansover80%precision(whichisverygoodifusingsuchsamples).
244G.
Kissetal.
Fig.
5.
ResultofrecallexaminationFig.
6.
ResultofprecisionexaminationTable1.
Runtime,accuracy,recallandprecisionvaluesofeachimageImageRuntimeAccuracyRecallPrecision(0-1)1.
48032ms80,5561%34,989983%0,9873527812.
50503ms87,3210%74,7283407%0,6270219843.
45815ms78,4133%27,5994314%0,7078459674.
53523ms86,8289%86,2483716%0,691096685.
48907ms82,3787%89,4566917%0,6521583236.
45962ms74,5288%60,1550501%0,986063247.
49185ms71,5572%54,2035412%0,9977440718.
51919ms75,7369%68,6987809%0,9840419329.
47646ms83,8932%61,719451%0,96828344210.
46471ms84,8415%72,857387%0,62668067211.
46991ms70,9533%16,0547004%0,73916341612.
50978ms71,2926%18,0046956%0,62133238813.
56035ms71,4357%47,9457611%0,57156761414.
53736ms72,6450%89,063646%0,93483871315.
45531ms82,1188%50,8651431%0,88327472616.
51051ms85,9209%67,1163571%0,59118307517.
45935ms79,3216%83,922605%0,57438970218.
57948ms65,2986%66,5004848%0,95679583119.
50487ms71,7855%61,324101%0,64601299720.
46289ms64,3528%87,9682645%0,95743394621.
55556ms82,6466%59,9626072%0,63280693522.
42149ms78,7552%31,8393793%0,66667657223.
42207ms68,0102%49,1480765%0,89602826424.
53199ms78,2759%50,4916674%0,59448594625.
48327ms79,9055%23,9523506%0,986024626HighResolutionDigitalTissueImageProcessingUsingTextureImageDatabases245Table1.
(Continued)26.
54416ms74,5399%70,6376023%0,94875978627.
53046ms80,5891%78,4338479%0,72289876428.
46335ms64,9316%52,0194195%0,77549403729.
58761ms64,8349%36,1403908%0,70559269230.
52995ms64,8009%82,1711576%0,59397349331.
44834ms74,0871%26,2567997%0,82426739932.
42018ms78,9241%25,836577%0,81903848733.
49489ms81,2256%10,4580788%0,71033442134.
48039ms81,3128%72,8684392%0,60929077235.
48893ms78,9584%30,04997%0,68920798736.
51592ms83,4439%80,8248287%0,94403878937.
43584ms65,4951%87,9658039%0,91366402438.
53240ms81,6129%18,0776139%0,87644223839.
49503ms70,2107%84,9993829%0,99204812440.
52690ms71,1560%68,9798564%0,85498464341.
43347ms70,2632%23,2500211%0,84380438942.
56431ms84,0216%60,8157887%0,62328569543.
54683ms64,8362%75,8693771%0,74980723844.
56764ms82,4279%613156813%0,94100970345.
58503ms81,9611%48,3010939%0,68407994446.
44213ms77,9225%74,623875%0,87460400347.
51478ms66,5134%76,3992459%0,85658205548.
47767ms78,8671%39,228208%0,70942241449.
46275ms76,8262%33,2365511%0,95231607250.
43277ms71,0800%84,0721253%0,651843577Fig.
7showssometestimages.
Duringvisualizationweareusingsomefalsecolorsbyouralgorithm:Yellowisacorrecthit,blackisundeterminedarea,blueisaback-groundandgreenisanotexaminedtissuetype.
Onthesampleimagestherearesometissuestructure:purplecircleareglands,pinkareaistheepithelium.
Inthistestwearesearchingforglands.
IntheBandDimagestheavoidanceofthefalsepositivehitsshowedbyblackcolor.
Finallytherearelotsofgoodhitsinagland,epitheliumandthebackground,amongtheblackundeterminedareas.
Fig.
7.
A)Testimage1,B)Resultsoftestimage1,C)Testimage2,D)Resultsoftestimage2246G.
Kissetal.
Thekeyofthealgorithm'saccuracyisthechoseddistancemeasuretechnique.
InthefuturewecanprobablyenhancetheaccuracybychangingthecurrenthistogrambasedapproachtotheGLCM(Gray-LevelCo-OccuranceMatrix)basedapproach.
7SummaryWehavesuccessfullyimplementedintotheMorphChecksystemanLBP-basedtex-tureclassifieralgorithm,whichsignificantlyprovidesmoreefficientobjectclassifica-tionandsegmentationthentheprevioussolutions.
Aftertheintegrationwehavedoneefficiencyexaminationofouralgorithm.
Wehaveusedpre-definedmetricsforthealgorithmassessment.
ForthetexturebasedalgorithmwehavecombinedMorpCheckwithaDICOM/PACSbasedsystemandannotatedalargenumberoftissueimages.
Weused1039textureinPACSdatabaseandtheLPBalgorithmwasworkingontheseannotatedtextures.
ThenewtexturebasedalgorithmandthecombinedPACShasbeenvalidatedandfacilitatesdiagnosticworkssignificantly.
Acknowledgements.
ThisworkmakesusesomeofthesoftwareresultsproducedbytheHun-garianNationalTechnologyProgramme,A1,Lifesciences,the"Developmentofintegratedvirtualmicroscopytechnologiesandreagentsfordiagnosing,therapeuticalpredictionandpre-ventivescreeningofcoloncancer"HungarianNationalTechnologyProgramme,A1,Lifesciences,(3dhist08)projectandtheE-RH1104/2-2011project.
AuthorswouldliketothankSemmelweisUniversityandMajor&Co.
toprovideusannotatedtissuesamplesforprocessingandclassification.
References1.
Visiopharmfordigitalpathology.
Visiopharm(March2013).
http://www.
visiopharm.
com/pdf/visiomorphdp-factsheet.
pdf2.
Her2-connect.
Visiopharm(March2013).
http://ww1.
prweb.
com/prfiles/2010/11/18/4253634/Imageforpressrelease.
bmp3.
Developerxd.
Definiers(March2013).
http://www.
definiens.
com/tissue-diagnostics.
html4.
Bickel,E.
L.
P.
:Theearthmover'sdistanceisthemallowsdistance:Someinsightsfromstatistics.
In:ICCV2001,IEEE12thInternationalConferenceonComputerVision,pp.
251–256(2001)5.
Digitalimagingandcommunicationsinmedicine,NationalElectricalManufacturersAs-sociation(2004).
http://dicom.
nema.
org6.
Reményi,A.
,Szénási,S.
,Bándi,I.
,Vámossy,Z.
,Valcz,G.
,Bogdanov,P.
,Kozlovszky,M.
,Sergyán,S.
:ParallelbiomedicalimageprocessingwithGPGPUsincancerresearch.
In:3rdIEEEInternationalSymposiumonLogisticsandIndustrialInformatics,Budapest,Hungary,August2011,pp.
225–248(2011)7.
Kozlovszky,M.
,Szénási,S.
,Vámossy,Z.
:GPGPU-baseddataparallelregiongrowingal-gorithmforcellnucleidetection.
In:CINTI2011,12thIEEEInternationalSymposiumonComputationalIntelligenceandInformatics,pp.
493–499(2011)8.
Fugal,D.
L.
:Conceptualwaveletsindigitalsignalprocessing.
Space&SignalsTechnical(2009)HighResolutionDigitalTissueImageProcessingUsingTextureImageDatabases2479.
XiaoyuWang,S.
Y.
,Han,T.
X.
:Anhog-lbphumandetectorwithpartialocclusionhandling.
In:ICCV2009,IEEE12thInternationalConferenceonComputerVision,pp.
1550–5499(2009)10.
Achanta,R.
,Shaji,A.
,Smith,K.
,Lucchi,A.
,Fua,P.
,Süsstrunk,S.
:Slicsuperpixels,Insti-tuteofElectricalandElectronicsEngineers,Tech.
Rep.
(2010)11.
Achanta,R.
,Shaji,A.
,Smith,K.
,Lucchi,A.
,Fua,P.
,Süsstrunk,S.
:Slicsuperpixelscom-paredtostate-of-the-artsuperpixelmethods.
IEEETransactionsonPatternAnalysisandMachineIntelligence34,2274–2282(2012)12.
Yasnoff,J.
W.
A.
,Mui,J.
K.
:Errormeasuresforscenesegmentation.
PatternRecognition9,217–231(1977)13.
ASHImageBank,hematologyimagebank.
(Marc2013).
http://imagebank.
hematology.
org/14.
ImageAftertexturedatabase(March2013).
http://www.
imageafter.
com/index.
php15.
TinEyeReverseImageSearch(March2013).
https://www.
tineye.
com/16.
RevIMGreversevisualsearch(March2013).
http://www.
revimg.
net/