科技英語閱讀與寫作

龍終建了科技之冬

科技英語閱讀與寫作

題目：CellSegmentationinDigitalHolographic

Images

學院：____________電子工程學院_________

專業(yè)：信號與信息處理

姓名：_________________________________

學號：_____________1602120898___________

Partner:

姓名：_______________張潤東_____________

學號：1601120188

CELLSEGMENTATIONINDIGITALHOLOGRAPHICIMAGES

NohaEl-Zehiry'OliverHayden2andAliKamen1

,MedicalImagingTechnologies,SiemensHealthcare,Princeton,NJ,USA

2In-VitroDXandBioscience,SiemensHealthcare,Erlangen,Germany

ABSTRACT

DigitalHolographicMicroscopy(DHM)isbecomingrecentlyverypopularfor

cellimaging.①(1)Themainadvantageofdigitalholographicmicroscopyover

classicalmicroscopytechniquesisthatitdoesnotonlyprovidetheprojectedimageof

theobjectbutalsoprovidesthreedimensionalinformationoftheobject'soptical

thickness.DHMtechnologycouldbethecoreofalabel-freeimagingfbrhematology

applications.Inanidealframework,abloodsamplecanbeimagedusingDHM,

machinelearningapproachescan(beusedfor)thecellextraction,differentiationand

consequentlycomputingalltherelevantbloodstatisticssuchastheMeanCorpuscular

Volume(MCV),theRedBloodCell(RBC)count,RedBloodCellDistributionWidth

(RDW)._Themostvitalcomponentinsuchaframeworkisaccurateextractionofthe

cells.(2)(2)Thispaperpresentsanovelapproachtocellsegmentationinwhicha

probabilisticboostingtreeclassifieristrainedtodetectthecentersofthecellsusing

Haar-Featu】es?Thedetectedcellcentersareusedtotriggeramarker-controlledpower

watershedsegmentationtocomputethecellboundaries.Additionally,wepresenta

comprehensiveevaluationofsegmentationmethodsfbrcellextractionindigital

holographicimages.

1.INTRODUCTION

DigitalHolographicMicroscopy(DHM)hasreceivedalotofattentionrecently

incellimaging[1,2,3].(3)DHM,ifassociatedwithproperimageprocessing

algorithms,canserveasthefundamentalbuildingblockinalabelfreecelldiagnostics

workflow.Insuchaworkflow,accuratecellextractionisavitalsteptoperformthe

analysis.Therefore,thoroughinvestigationofcellsegmentationinDHMimagesisa

persistentnecessity.CurrentDHMapplicationstudiessuchas[1,2]usesimple

segmentationtoolsorutilizegenericcellsegmentationtools[4](that)arenot

designedforDHMandtheaccuracyoftheirperformanceforDHMwasnotassessed

intheliterature.Onenotableexceptionisthecellsegmentationapproachpresentedby

Yietal.in[3].Thealgorithmin[3]usesasequenceofmorphologicaloperationson

thephaseimagetogeneratemarkersformaker-controlledwatershedsegmentation.

Theapproachisrelativelycomplicatedandthemaindisadvantageisitssensitivityto

theparametersofthemorphologicaloperatorssuchasthesizeofthestructure

elementusedineveryoperation.Wewillovercomethisproblem.Inthispaper,we

presentatwo-stepsegmentationapproach:Celldetectiontolocalizethecentersofthe

cellsandcellsegmentationtodelineatetheboundariesofthecells.Qualitativeand

quantitativeassessmentofoursegmentationmethodispresented.Moreover,we

introduceacomparisontothestate-of-theartcellsegmentationmethods[3]and[4].

⑷Therestofthepaperisorganizedasfollows:Section2presentsthedetailsof

ourproposedsegmentationmethods,section3introducestheexperimentalresultsand

thecomparisontoothersegmentationmethods,thenaconcludingdiscussionis

Fig.I.PipelineofCellExtractionAlgorithm.Left:originalimagesuperimposedbytheprobabilitymapofapixel

beingacellcenter.Middle:Resultofaggregationoftheprobabilitymapresponseandgeneratingtheinternal

markers.Left:Segmentationresults.

2.METHODS

Markercontrolledwatershedhasbeenusedrepeatedlyforcellsegmentation[4,3,

5,6].Robustmarkergenerationisnecessarytoobtainqualitysegmentationresults.

Mostoftheprevioussegmentationmethods[3,4]useasequenceofmorphological

operationstoidentifythebestpossiblesetofmarkers.Onekeyproblemwithsuch

approachesis(that)itrequiresthetuningofmanyparametersforthestepswithinthe

morphologicaloperationchain.@(5)Itislesslikelytohavedsinglesetof

parametersthatcouldworkefficientlyforalltheimages.Ontheotherhand,tuning

theparametersforeachsingleimageiscompletelyimpracticalandnegatesthehigh

throughputadvantageassociatedwithDHMtechnology.Motivatedbytheprevious

drawbacks,wepresentanewcellsegmentationapproach.Thenoveltyofour

approachistwo-fold:First,robustmarkergeneration(using)cellcenterdetection.0

⑹Second,weusepowerwatershedforthecellsegmentationwhichhasbeenproven

moreefficientthanconventionalwatershedingeneralsegmentationproblems[7].

Figure1showsthepipelineofourcellextractionapproach.

2.1.RobustMarkerGeneration

(7)Weconsiderthemarkergenerationproblemasanobjectdetectionproblem

whereweaimatfindingthepositionsofcellcenters.Forthispurpose,weusea

machinelearningbasedapproachinsteadofmorphologicaloperationstominimizethe

sensitivitytoparameters9choice.Inthiscontext,weuseProbabilisticBoostingTree

(PBT)learningframework[8].Inthelearningphase,thePBTconstructsatree(in

which)eachnodecombinesasetofweakclassifiersintoastrongclassifier.Intesting

phase,theconditionalprobabilityiscomputedateachnodeandtheprobabilityis

propagatedtothesourceofthetreetoprovidetheoverallprobability.Inourtraining,

weusetheHaarfeatures[9]toformtheweakclassifiers.?⑻Intesting,we

computetheprobabilityofeachpixelintheimagebeingacenterofacell,the

probabilitymapisthresholdedtokeeponlythepixelsthataremorelikelytobeacell

center.

2.2.AggregationoftheDetectionResponses

Aftercomputingtheprobabilitymapandapplyingthethreshold,wegetahigh

responseinsideeachcell.However,theresponseisnotnecessarilysmoothand

connectedwhichmayleadtofalseidentificationofonecellasmultiplecells.

Therefore,toaggregatetheseresponse,weapplyaclusteringsteponthethresholded

probabilitymap.Theclusteringservestwopurposes,first,itaggregatedthecell

responsesinasinglecell.Second,itprovidesalargersetofpixelstoformasthe

internalmarkerforthesegmentation.Thepixelsofeachclusteraremergedintoa

singleconnectedcomponentthatservesasaninternalmarkerforthecell

segmentationstep.Themiddle

imageinFigure1showsasampleofthemarkerscomputedafterclustering.Theseare

usedasinternalcellmarkers,externalcellmarkershighlightingthebackgroundare

obtainedbyapplyingwatershedtransformontheinternalmarkers.

2.3.CellSegmentationusingPowerWatershed

Thepowerwatershedsegmentationreviewedinthissectionwaspresentedby

Couprieetal.in[7].Theformulationisperformedonadiscretegraph.Agraph

g={V,E}consistsofasetofverticesvEVandasetofedgeseG￡QVxV.An

edgeincidenttovertices片andvjisdenotedey.Inourformulation,eachpixelis

identifiedwithanode,片.Aweightedgraphisagraphinwhicheveryedgee,is

assignedaweight

w.⑼Theseededsegmentationenergyin-7]wasgivenas

ngnE啕叫一叼/+￡共產(chǎn)?+￡嗚皿一1R

s.t.x(F)=1.x(B)=0.

s,=1ifX,>I，0ifXa<(1)

wherex^and與arethebinarylabelsassociatedwithvertices,andrespectively.F

andBrepresentthesetsofforegroundandbackgroundmarkers.(Itwasshownin[7]

that)whenpooandq>1,thisleadstoamoregeneralwatershed,namely,power

watershedthatyieldedbetterresults.Cellsegmentationcouldbenefitfbrthe

improvementsassociatedwithpowerwatershed.InSection3,wewillpresentthe

comparisonbetweenthesegmentationresultsusingpowerwatershed!andwatershed

thathasbeenusedinthevastliteraturefbrcellsegmentation.

DonorDonorDataDonorOverall

12Set34

TP97.4%97.8%97.3%96.5%97.2%

FP6.8%4.3%3.7%5.7%5%

Table1.Resultsofthedetectionofthecellcenter.TPandFParethetruepositiveandfalse

positiverates,respectively.

3.EXPERIMENTALRESULTS

Inadditiontothenovelsegmentationapproach,weconsidertheevaluationand

comparisonofsegmentationmethodsforDHMasamajorcontributionofthepaper.

Wechoseasubsetofsegmentationmethodsthatweconsiderrepresentativeofthe

currentsegmentationmethods.Qualitativeandquantitativecomparisonwillbe

presentedinthissection.

3.1.DataDescriptionandCellDetection

ThedatasetisacquiredusingtheQModHolographicandFluorescence

Microscope[10].Thecellswereilluminatedwithalightsourceofwavelength=

550nmandthemagnificationobjectiveis60x1.Thenumberofdonorsusedinthis

studyis4.Thenumberofimagesforeachdonor(variesfrom)6-9imageswithatotal

of28images.Eachimagecontainsmultiplecellsvaryingfrom10-30withatotalof

615cells,lbevaluatetheaccuracyofthedetectionalgorithm,weusen-foldleaveone

outcrossvalidation.Specifically,wedid4training/testingsets.Foreachsetweleave

outalltheimagesassociatedwithagivendonorandtraintheclassifierontheimages

oftherestofthedonors.Fortesting,wetestonlyontheunseendatafromtheleftout

donor.(10)Weexcludevllthecellsthataretouchingtheboundariesoftheimageto

ensurethattheevaluationisperformedonlyonvalidcellcandidates.

Theprobabilisticboostingtreeoutputsaprobabilitymapofeachpixel(being)a

cellcenter.Wethresholdtheprobabilitymapatp=0:75toconsideronlythepixels

thatarehighlikelytobecells.(11)Thethresholddoesnotprovideasinglecellcenter

butratheraclusterofpointsinthecenterofthecellasdepictedinthefirstimageof

Figure1.Inthisimage,theprobabilitymapissuperimposedontheoriginalimage

withthelostprobabilityinredandthehighestprobabilityinblue.Table1showsthe

(12)resultforthedetectionsystem.

3.2.CellSegmentation

Wechosearepresentativesubsetofcellsegmentationmethodstocompare

againstMl[4]andM2[3].Moreover,todecoupletheeffectofeachcontribution(the

machinelearningdetectioncomponentandthepowerwatershedsegmentation

component),wepresenttwonovelmethodsM3thatbenefitsfromdetection

componentonlyandM4thatbenefitsfromthepowerwatershedsegmentation

componentonly.ThedescriptionsofmethodsMl-M5aresummarizedasfollows:

l.CellProfiler[4](Ml):CellProfiler(isconsideredas)oneofthebenchmarksin

cellsegmentationandhasbeenrepeatedlyusedbyotherresearchgroupsto

obtainthesegmentationandprovideanalysis.

2.MarkerControlledWatershedforDHM(M2)[3]:Dedicatedcellsegmentation

forDHMtechnologywasonlydiscussedin[3].Themethodusesa

complicatedworkflowtogeneratethemarkersforthewatershedsegmentation.

Theworkflowissummarizedasfollows:(1)Imageisnormalized,(2)Otsu's

thresholdisappliedtoobtainIbin,(3)Holesarefilledusingmorphological

constructiontoIbin,(4)GradientimageIgradiscomputedusingSobel

operator.(5)MorphologicalopeningisappliedtoIbintoobtainlopen.(6)

Morphologicalerosion(13)isappliedtolopentoobtainlerode.(7)

Morphologicalreconstructionisappliedwithlopenasthemaskandlerodeas

amarkertoobtain\rec.(8)ComputeIsubasIopen-Irec.(9)Apply

morphologicaldilationonlerodetoobtainIdilate(10)Obtaintheinternal

markersbycombiningIsubandIdilate.(11)Applywatershedtransformonthe

internalmarkerstogeneratetheexternalmarkers.(12)Applymarkercontrolled

watershedsegmentationusingthemarkersgeneratedin(11).Mostofthe

previousstepsrequiretuningofstructureelementparameterswhichmakesthe

workflowerrorprone.Althoughtheworkflowwascarefullycraftedwith

specificparametersprovidedbytheauthorsin[3]toavoidmergingtouching

cellsoreliminatingsmallcomponents,asinglesetofparametersdoesnotwork

wellinallscenarios.Wetestedseveralcombinationsofparameterstotryto

achievethebestresultsforourdataset.Afterseveralexperiments,wefigured

thattheparametersprovidedbytheauthorsin[3]workbest.(14)Hence,for

thecomparison,weusetheparameterselectionin⑶.Onecanarguethatthe

improvementover[3]isduetoperformingthesegmentationusingpower

watershedratherthanthegenerationoftheaccuratemarkers.So,(15)itis

worthclarifyingthattheimprovementisduetoacombinationofbothfactors.

Todecoupletheeffectofeachcomponent.Wealsocompareagainsttwoother

workflows(M3andM4).

3.MachineLearningMarkerGenerationwithWatershed(M3):Inthismethod,

weusethemarkersgeneratedbythemachinelearningbasedcellcenter

detectionto

triggerwatershedsegmentation.

4.MorphologicalMarkerGenerationwithPowerWatershed(M4):Inthismethod,

wegeneratethemarkersusingthemorphologicalworkflow(M2)andapply

powerwatershedforthesegmentationinsteadofwatershed.

5.MachineLearningMarkerGenerationwithPowerwatershed(M5):Thisisthe

proposedworkflowwhichtakesadvantageofrobustmarkergenerationaswell

asmoreaccuratesegmentationusingpowerwatershed.

Figure2showsasampleofourresults.Thegallerydepictstheresultsofthefive

algorithmsforthreedifferentimagesfromdifferentdonors.Thethirdcolumnshows

thatallthealgorithmsworkcomparablywellifthecellsaresparselydistributedinthe

image.This,however,isnotverypracticalasitdoesnotefficientlyutilizethefull

fieldofview.Inpractice,thecellsmaybeveryclosetoeachother.Insuchscenarios,

mostofthemorphologicalbasedapproacheswouldfailinaccuratelyextractingthe

cells.Intheyellowboxes,weshowanexamplewheretwocellsaretouching.The

algorithmin[4]mergesthetwocellsinoneentity.Thecarefullycrafted

morphologicalworkflowin[3]managedtoseparatethetwocellsasshowninthe

secondrowand(16)sodidallthealgorithmswedevelopedasshowninthethird,

fourthandfifthrows.Whenthecellsslightlyoverlaps,itbecomesmorechallenging

andmorphologicalbasedmethodsfailtoaccuratelyextractthecellsasshowninthe

examplesinthegreenboxes.Ontheotherhand,M4andM5cansuccessfullyextract

thecellsduetotherobustnessofthecellcenterdetectioncomponent.Itisworth

notingthathighthroughputsystemsrequireafullutilizationofthefieldofview

whichresultsinmultipleoccurrencesofsuchtouchingorslightlyoverlappingcells.A

limitationofthecurrentapproachisthatwhentheoverlapbetweenthecellsislarge,

theboundariescannotbedelineatedaccuratelyasdepictedintheblueboxes.However,

eveninsuchscenario,ouralgorithmismoreaccuratethan[4]and[3]becauseitstill

identifiedtwocellswhichisimportantforan

accuratecellcount.

至

,€s￡

K

c

-5*>方

u6

d胃

_M

M至

-

主

?

|國

SS

CJ

M

Oq

co

ld

—

1

0

生

一

里

6i國

U9s

36b

-*

I0M

豈￥

o

d

.

Fig.2.Sampleofthesegmentationresultsobtainedusingthefivedifferentalgorithms.

Forquantitativeassessment,wechosethemostcommonmetricsfor

segmentationevaluation,namely,thesensitivity,specificity,DiceandJaccard

similarityindexdefinedas:

TP

Sensitivity

TP+F.V?

TN

Specificity⑵

TN+FP'

TTP

DiceSimilarityIndex

JaccardSimilarityIndex

TP+FP+N

whereTP,FP,TNandFNrefertothetruepositives,falsepositives,truenegativesand

falsenegatives,respectively.

Table2showstheresultsforthemethodsM1-M5.Thequantitativeassessment

showsthateachcomponentweintroduced(contributedto)improvingthe

segmentationresults.However,(17)itisevidentthattherobustlocalizationofthecells

usingthecelldetectionframeworkplayedamorecrucialrole.Whilethesensitivity

improvedby10%whenusingourmachinelearningmarkergeneration,itonly

improved2%whenusingpowerwatershedinsteadofwatershed.

M1M2M3M4M5

Sensitivity88.7083.3792.9484.5894.83

Specificity98.1898.5797.4498.9298.44

Jaccard80.9577.5581.5380.0686.51

Dice89.3288.8489.7488.8492.66

Table2.QuantitativeComparisonbetweenthedifferentSegmentationMethods.

4.CONCLUSIONANDFUTUREWORK

Thepaperpresentedanovelapproachforcellsegmentationindigital

holographicmicroscopicimages.Therobustmarkergenerationpresentedinthepaper

outperformsthemorphologicalworkflowformarkergenerationthatiscommonly

usedtoinitiatewatershedsegmentation.Weintroducedthepowerwatershedtothe

cellsegmentationproblemandpresentedquantitativeevidencethatitworksbetter

thanwatershed.Thepaperpresentedaquantitativecomparisonoftheproposed

approachtothecommonmethodsusedforcellsegmentation.Onelimitationofour

approachandallthemethods(M1-M5)isthatitisnotcapableofproperlysegmenting

overlappingcells.Inthefuture,weplantoaddpostprocessingthatidentifiesthe

overlappingcellsandformulatesalayeredsegmentationalgorithmtoseparate

properlytheoverlappingcells.

一、語法分析

(1)Themainadvantageofdigitalholographicmicroscopyoverclassicalmicroscopy

techniquesisthatitdoesnotonlyprovidetheprojectedimageoftheobjectbut

alsoprovidesthreedimensionalinformationoftheobject'sopticalthickness.

本句主語為Themainadvantage,ofdigitalholographicmicroscopy作為后置定

語修飾主語，advantage...over為固定搭配表示與…相比的優(yōu)勢含義，表語由

that引導的表語從句充當，其中notonly…butalso…表示不僅而且的意思，并

且of+n具有形容詞的含義。

(2)Thispaperpresentsanovelapproachtocellsegmentationinwhichaprobabilistic

boostingtreeclassifieristrainedtodetectthecentersofthecellsusing

Haar-Features.

本句主句為Thispaperpresentsanapproach.present:提出，approachto…：固定

搭配，…的方法，inwhich引導定語從句修飾approach,這時可與where互換，

classifieristrainedtodetect…不定式表目的，thecentersofthecellsusing

Haar-Features.using前省略的介詞by,表方式。

(3)DHM,ifassociatedwithproperimageprocessingalgorithms,canserveasthe

fundamentalbuildingblockinalabelfreecelldiagnosticsworkflow.

ifassociatedwithproperimageprocessingalgorithms在句中作條件狀語，可將

其放在句首：Ifassociatedwithproperimageprocessingalgorithms,DHMcan

serveasthefundamentalbuildingblockinalabelfreecelldiagnosticsworkflow.

連詞+分詞做狀語，省略了主語(If也可省略)，主語是主句主語，serveas為

固定短語表示充當…，inalabelfreecelldiagnosticsworkflow介詞短語做狀語。

(4)Therestofthepaperisorganizedasfollows:Section2presentsthedetailsofour

proposedsegmentationmethods,section3introducestheexperimentalresultsand

thecomparisontoothersegmentationmethods,thenaconcludingdiscussionis

providedinsection4.

該段主要是對文章結構進行簡要說明，thepaperisorganizedasfollows采用了

organize的被動態(tài)以及短語asfollows,為避免重復，對每一部分的介紹使用了

不同動詞，分另II是present,introduce,provide；thecomparisontoother

segmentationmethods.comparisonto為compareto…的變形;引入最后一部分

使用了連詞then而不是finally.

(5)Itislesslikelytohaveasinglesetofparametersthatcouldworkefficientlyforall

theimages.

Itislikelytodo…表示做什么事很有可能,less含義與more相反,Itislesslikely

todo…表示做…不太可能，it為形式主語，真正主語為不定式。asetof:一套，

一副，that引導一個定語從句修飾parameters,work表示起作用，工作的意

思。

(6)Second,weusepowerwatershedforthecellsegmentationwhichhasbeenproven

moreefficientthanconventionalwatershedingeneralsegmentationproblems[7].

Second:其次；weusepowerwatershedforthecellsegmentation,use…for…:將...

用在…上；which引導定語從句修飾powerwatershed,句中使用現(xiàn)在完成時

的被動態(tài)hasbeenproven表示此事現(xiàn)在已經(jīng)被證明了；moreefficientthan…比

較級的運用，表示比…更有效。

(7)Weconsiderthemarkergenerationproblemasanobjectdetectionproblemwhere

weaimatfindingthepositionsofcellcenters.

consider…as…：將…視為…;where引導定語從句可將其替換成inwhich(如

(2));aimat*-：旨在…;of+n.做后置定語。

(8)Intesting,wecomputetheprobabilityofeachpixelintheimagebeingacenterof

acell,theprobabilitymapisthresholdedtokeeponlythepixelsthataremore

likelytobeacellcenter.

Intesting:在測試中，做狀語；ofeachpixel…做后置定語修飾probability,each

pixelbeingacenterofacell是分詞的獨立結構，分詞帶有自己的主語；the

pixelsthataremorelikelytobeacellcenter