Content uploaded by Angus J.M. Watson
Author content
All content in this area was uploaded by Angus J.M. Watson on Mar 08, 2023
Content may be subject to copyright.
Diagnostics2023,13,1038.https://doi.org/10.3390/diagnostics13061038www.mdpi.com/journal/diagnostics
Review
Clinicians’GuidetoArtificialIntelligenceinColonCapsule
Endoscopy—TechnologyMadeSimple
IanI.Lei
1
,GoharJ.Nia
1
,ElizabethWhite
2
,HagenWenzek
2
,SantiSegui
3
,AngusJ.M.Watson
4
,
AnastasiosKoulaouzidis
5,6,7
andRameshP.Arasaradnam
1,8,9,
*
1
DepartmentofGastroenterology,UniversityHospitalofCoventryandWarwickshire,
CoventryCV22DX,UK;brian.lei@uhcw.nhs.uk(I.I.L.)
2
CorporateHealthInternational,InvernessIV25NA,UK
3
MathematicsandComputerScienceDepartment,TheUniversityofBarcelona,Barcelona58508007,Spain
4
InstituteofAppliedHealthSciences,UniversityofAberdeen,AberdeenAB243FX,UK
5
DepartmentofGastroenterology,OdenseUniversityHospital&SvendborgSygehus,
5700Odense,Denmark
6
DepartmentofClinicalResearch,UniversityofSouthernDenmark(SDU),5000Odense,Denmark
7
DepartmentofSocialMedicineandPublicHealth,PomeranianMedicalUniversity,70‐204Szczecin,Poland
8
WarwickMedicalSchool,UniversityofWarwick,CoventryCV47AL,UK
9
DepartmentofGastroenterology,LeicesterCancerCentre,UniversityofLeicester,Leicester,LE17RH,UK
*Correspondence:r.arasaradnam@warwick.ac.uk
Abstract:Artificialintelligence(AI)applicationshavebecomewidelypopularacrossthehealthcare
ecosystem.Coloncapsuleendoscopy(CCE)wasadoptedintheNHSEnglandpilotprojectfollow‐
ingtherecentCOVIDpandemic’simpact.Itdemonstrateditscapabilitytorelievethenationalback‐
loginendoscopy.Asaresult,AI‐assistedcoloncapsulevideoanalysishasbecomegastroenterol‐
ogy’smostactiveresearcharea.However,withrapidAIadvances,masteringthesecomplexma‐
chinelearningconceptsremainschallengingforhealthcareprofessionals.Thisformsabarrierfor
clinicianstotakeonthisnewtechnologyandembracetheneweraofbigdata.Thispaperaimsto
bridgetheknowledgegapbetweenthecurrentCCEsystemandthefuture,fullyintegratedAIsys‐
tem.Theprimaryfocusisonsimplifyingthetechnicaltermsandconceptsinmachinelearning.This
willhopefullyaddressthegeneral“fearoftheunknowninAI”byhelpinghealthcareprofessionals
understandthebasicprincipleofmachinelearningincapsuleendoscopyandapplythisknowledge
intheirfutureinteractionsandadaptationtoAItechnology.ItalsosummarisestheevidenceofAI
inCCEanditsimpactondiagnosticpathways.Finally,itdiscussestheunintendedconsequences
ofusingAI,ethicalchallenges,potentialflaws,andbiaswithinclinicalsettings.
Keywords:artificialintelligence(AI);machinelearning(ML);deeplearning(DL);convolutional
neuralnetworks(CNN);decision‐makingsystems(DMS);coloncapsuleendoscopy(CCE)
1.Introduction
Therecentdecade’sprofoundtechnologicaladvanceshaveconsiderablytrans‐
formedthemedicallandscape.Artificialintelligence(AI)applicationshavebecome
widelypopularingenomicanalysis,roboticsurgery,predictionandsupportdiagnosis,
andtreatmentdecision‐makingacrossthehealthcareecosystem.Therehasalsobeensig‐
nificantinterestinAIsolutionsingastroenterologyinrecentyears.Withmanystudies
publishedandpotentialopportunitiesavailableinthisfield,gastroenterologyandendos‐
copyhealthcareprofessionalsmustunderstandandevaluateAIstudiesascriticalstake‐
holdersinsuccessfullydevelopingandimplementingAItechnologies.
Coloncapsuleendoscopy(CCE)wasfirsttestedin2006withthefirstmulticentre
studypublishedinIsrael[1].Comparedwiththegold(reference)standard,i.e.,colonos‐
copy,CCEwasfirstmetwithscepticismduetoitsdisadvantages,includingextensive
Citation:Lei,I.I.;Nia,G.J.;White,E.;
Wenzek,H.;Segui,S.;Watson,
A.J.M.;Koulaouzidis,A.;
Arasaradnam,R.P.Clinicians’Guide
toArtificialIntelligenceinColon
CapsuleEndoscopy—Technology
MadeSimple.Diagnostics2023,13,
1038.https://doi.org/10.3390/
diagnostics13061038
AcademicEditors:Antonella
SantoneandHajimeIsomoto
Received:29December2022
Accepted:7February2023
Accepted:21February2023
Published:8March2023
Copyright:©2023bytheauthors.Li‐
censeeMDPI,Basel,Switzerland.
Thisarticleisanopenaccessarticle
distributedunderthetermsandcon‐
ditionsoftheCreativeCommonsAt‐
tribution(CCBY)license(https://cre‐
ativecommons.org/licenses/by/4.0/).
Diagnostics2023,13,10382of19
bowelpreparationtoachieveareasonablepolypdetectionrate(PDR),highcost,andina‐
bilitytoperformbiopsyortherapy(e.g.,polypectomy).EventhoughthePillCamColon2
wasupgradedtoallowpanoramicviewsin2011,theuptakeofCCEcouldhavebeenbet‐
terintheUK.However,followingtheimpactoftherecentCOVIDpandemic,anNHS
Scotlandevaluationdemonstratedthatthetechnologycouldleadtorelievingthebacklog
inendoscopyonanationallevel.Still,CCEgeneratesavideocontainingmorethan50,000
images;thiscouldbetime‐consumingandinefficienttoanalyse[2,3].Asaresult,thead‐
vancesinAIapplicationonimageanalysismakeAI‐assistedCCEvideoanalysisoneof
themostactiveresearchareas.
Nowadays,itisbroadlyacceptedthatthedatagenerationrateisbeyondthehuman
cognitivecapacitytobeeffectivelyanalysedandmanaged.Therefore,AIwilllikelyhave
acomplementaryroleindeliveringhealthcareservicessoon.Nonetheless,duetothecom‐
plexityofmachinelearning(ML),masteringtheconceptofAIbycliniciansremainschal‐
lenging[4–7].
Robustresearchintocomputer‐aideddetection(CAD)inupperandlowergastroin‐
testinal(GI)endoscopyimageshasdemonstratedencouragingresultsinrecentyears[8,9].
Thesuccessalsobecamevisibleinthewirelesscapsuleendoscopy(WCE)field,wherean
earlyCADmodelonWCEshowedasensitivityof88.2%,specificityof90.9%,andaccu‐
racyof90.8%todetecterosionsandulcerations,withevidenceofrelievingthereader’s
overallworkloadandreadingtime[10,11].Thisrevivedinterestisalsobeingtransferred
intotheCCEworld.AIstartedtobeusedforvarioustasksandachievedthefirstremark‐
ableresults:arecentmeta‐analysisshowedthatthesensitivitieswere86.5–95.5%for
bowel‐cleansingassessmentand47.4–98.1%forthedetectionofcolorectalcancer(CRC)
andpolyps[12,13].
Understandably,thepredominantfocusintheliteratureisontheevidencearound
theaccuracyoftheseAImodelsinCCE,astheauthors’goalwasundoubtedlytobuild
trustaroundartificialintelligenceintheclinicalworld.However,toencouragetheadop‐
tionofCCEAItechnologyinaclinicalsetting,understandingthe“how”inadditionto
anydata‐drivenevidenceisessentialtobuildthattrustamongclinicalprofessionals.
Therefore,thispaperusesadifferentapproachtobridgethegapbetweenrecognitionand
trust.WefirstsimplifytechnicaltermsandthenfocusonhowexistingevidenceofAIin
CCEshowsitsimpactondiagnosticpathways.Wealsohighlighttheunintendedconse‐
quencesofusingAI,potentialflaws,andbiaswithinclinicalsettings.
Theultimateaimisaseamlesscollaborationofmedicalprofessionalsandcomputer
scientiststotranslateprototypeAIsolutionsmorequicklyintovaluableclinicaltools.
2.AITerminologyandtheConceptofMachineLearning
2.1.TheDifferencebetweenMachineLearningandArtificialIntelligence
ThepublichasusedAIinterchangeablywithmachinelearning(ML),whichrefersto
usingcomputerstomodelintelligentbehaviourthatcanperformtasks.However,AIis
commonlydefinedasthedevelopmentofmachinecapabilitiestosimulatenaturalintelli‐
gencebyperformingtasksthatusuallyrequirehumaninput.
Ontheotherhand,MLisasubfieldofAIthatusesacombinationofmathematical
andstatisticalmodellingtechniquesthatutiliseavarietyofalgorithmstolearnandauto‐
maticallyimprovethepredictionoftheresult.Itaimstobuildmathematicalmodelsbased
onthegivendatathathavepredictivepowerovernewandunseendata[14].Thediffer‐
enceisthatAIrelatesonlytocreatingintelligentmodelsthatcansimulatehumancogni‐
tion,whereasMListhesubsetofAIthatallowsmodelstolearnfromdata.
Diagnostics2023,13,10383of19
2.2.TerminologyinMachineLearning
TounderstandandapplythecomplextechnicalscienceofMLinCCE,itisessential
tostartbyaddressingtheterminologygapincomputingengineeringformedicalprofes‐
sionals.Thiscouldensurethatalltheconceptsareunderstoodcorrectly.Furthermore,
sharingjargonandexpertisefrombothfieldscannarrowthiscommunicationgap.There‐
fore,weprovidethemostbasicandrelevanttechnicalterminologyinmachinelearning
forallmedicalprofessionals(Table1).
Table1.Relevanttechnicalterminologyinmachinelearning.
Terminology:Definitions:
Artificialintelligence
(AI)
Itisatechnologythatenablesamachinetosimulateahuman’snatural
intelligenceandbehaviour.
Machinelearning(ML)ItisasubfieldofAIthatfocusesonhowacomputersystemdevelopsits
intelligencetopredicttheresultofunseendataaccurately.
Example
Itisasinglepairofinput‐outputdatausedintraininganMLalgorithm.
Itincludespairedfeaturesandlabelsineachexample.Asetofexamples
formadataset.
Features
Thisistheinputdatathatisfedintothemachinelearningsystem.For
example,inCCE,thevisualpropertiesoftheimages(inputdata)arepro‐
cessedintheformofacollectionofnumberstoformfeaturesfortheML
system.
Labels
Thepreciseoutputdatausedtocomparewiththeprediction(thepre‐
dictedoutputgeneratedbytheMLsystem).
InCCE,labelsaretheannotationsofpolypbyanexpertreader.Then,the
AI‐predictedresultsareverifiedagainsttheselabels.
PredictionTheoutputdataproducedfromanunseeninputbytheMLsystemthat
haslearnedfrommanytrainingsamples.
Trainingloop
Thisisarepeatedtrainingprocesstoallowsufficientmachinetrainingto
takeplace.Thisisperformedonnumeroussetsofinput–outputdata(ex‐
amples)inatrainingset.
TrainingdatasetThiscomprisesasetofexamplesthatareusedfortheMLsystemtolearn
thefunctionthatconnectsthefeaturestothelabels.
Validationdataset
Thiscomprisesasetofexamplesthatareonlyusedperiodicallytoassess
andtunethehyperparametervaluesthatweretrainedonthetraining
set.
Testdataset
ThiscontainsasetofexamplesthattheMLsystemhasneverbeenex‐
posedto.ItteststheMLsystem’sgeneralisationperformancetounseen
data.
Deeplearning
Atypeofmachinelearningmodelthatisformedbynumerouslayersof
neuralnetworkandallowsthefeaturestobeorganisedintohierarchical
layers.Themajordifferencefromtraditionalmachinelearningisthatthe
featuresandrelationsaredirectlylearntfromtheinputdata(end‐to‐end
learning)toproducetheprediction.
Hyperparameters
Thesearetheparametersusedtocontrolthelearningprocedureand
trainthemodel.Thisispredeterminedbeforethetrainingset.Theselec‐
tionofhyperparametersincludesthesizeofthesamplesetandthenum‐
beroflayersintheneuralnetwork.Thehyperparametertuningprocess
involveschangingthetrainingconfigurationsandthistakesplacewhen
themodelisevaluatedonavalidationdataset.
Convolutionalneural
network(CNN)
Itisatypeofneuralnetworkthatisdesignedforvisualimagery.Ituses
convolutionalfilters(kernels)tobuildashared‐weightarchitecturethat
includeslayersoffullyconnectedneuralnetworks.
ClassificationAformofsupervisedlearningandthegoalofthemodelistomatchthat
inputwithpredefinedcategoriesattheoutput.Forexample,theCCEML
Diagnostics2023,13,10384of19
algorithmclassifiedthelesionsintopolyporcancer,whichareprede‐
finedcategories.
Overfitting
Aphenomenonoccurswhenthemodelstartstolearnallthedetailedfea‐
tures,suchasthebackgroundnoiseand“memorise”thetrainingset
(tightlyfittedtothetrainingset).Thishappenswhentheerrorintheval‐
idationdatasetstartstodeteriorateduetopoorgeneralisationtonew
data(themodelonlyworkswellinthetrainingexamplesasithasmem‐
orisedallthedetails).
Underfitting
Aphenomenonoccurswhenthemodelcannotobtainagoodfitofthe
trendtothedatasetduetoalackoftrainingorthemodel’sdesignistoo
simpletofitacomplextrend.
RegularisationTechniquesusedtoaddressoverfittingbycommandingthemodelto
learnandretainsomegeneralisationduringthetrainingprocedure.
2.3.MachineLearning
MLissimilartocomputerprogramming,asillustratedinFigure1.Theprocessof
transformingtheinputintotheoutputisknownasafunction.Incomputerprogramming,
theprogrammerencodesthestepsbasedonrulesintofunctionstoprovideanautomated
output.Manualeffortsarerequiredtosupportthisrule‐basedtechnique.
Figure1.Thisshowstheprocessoftransformingtheinputintotheoutputbyusingafunctionthat
programmersincomputersciencecreate.Ontheotherhand,themachinelearningsystemcanlearn
anddevelopthefunctionbystudyingtheexistinginput–outputpairstobuildaperfectfunction
withoutrelyingontheprogrammer.Examplesareincludedtodemonstratethebasicconceptof
machinelearningbyusingasimplemathematicalfunction.
Diagnostics2023,13,10385of19
Incontrast,thatfunctioncorrelatingtheinputandoutputremainsunknowninML.
Insteadofrelyingonaprogrammertofindthefunction,theMLsystemcanlearnand
createthefunctionbystudyingtheexistinginput–outputpairsviatraining.Aftertraining
themachinelearningsystemonnumerousinput–outputpairs,itwillbuildafunctionthat
canaccuratelyprocesstheunseeninputdata(features)toanaccuratelypredictedoutput
data(prediction)(Figure1).Therefore,oneoftheadvantagesofusingmachinelearning
isthatitcanlearnanddevelopatremendouslycomplexfunctionbasedonthemultitude
ofinput–outputpairs,whichwouldbeimpracticalorimpossibleforaprogrammerto
achieve[15].
Anequivalentanalogywouldbetheprocessoflearninghowtodriveamanualcar.
First,thelearneristaughtthebasicprinciples,includinghighwaycodes,gearshifting,
driving,parkingetc.(examples).Forexample,whenstartingacarfromastilltoamoving
position,thedrivermustshiftthegearandapplypressureonthegaspaddle(thisisthe
essentialfunctionlearnedfromexamples).Then,thelearner,taughtbytheexpertinstruc‐
tor,repeatedlytrainsinvariouspreplannedweather,roads,androundabout(learning
andimprovingthefunction)duringtheirtraining(trainingloops).Oncethesebasicskills
andprinciplesarediscovered,thenewdrivercandrivedifferenttypesofcarsonanypre‐
viouslyunseenroadsandareas(newunrecognisedinput)withfurtherassistancefrom
thedrivinginstructor(validationsets).Thedrivingskillswillcontinuetoimproveuntil
theyareadequateforthedrivingtest(e.g.,thetestdataset).Whenmoretypesofdifferent
roads,roundabouts,andcountrysideroadsareexploredthroughthedrivingprocess,the
drivingskillsareimprovedcontinuously(exposedtoanextensivedatasettoimprovethe
overallfunctioninretrainingafterthetestdataset).
TheMLalgorithmusedinCCEispredominantlysupervisedlearning,wherethein‐
puthasbeenprelabelled.UsingCCEasanexample,theMLalgorithmcanproduceapre‐
cisemappingfunctiontoaccuratelyidentifypolypsbyusingtheseprelabelledinputs,co‐
loncapsulepolypimages,andthepairedoutputslabelledas“polyp”(Figure1).
2.4.DataTypes:StructuredandUnstructuredData
Datausedandstoredinourhealthcaresystemcomprisesvariousformats,forexam‐
ple,graphics,laboratoryvalues,andfree‐textmedicalsummaries.Thetypeofdataissep‐
aratedintostructureddataandunstructureddata.Structureddataisstoredandorganised
inawell‐definedmanner,ofteninstructuredsqldatabases,spreadsheets,orlistsofnum‐
bersorcategories(e.g.,listofnames,diagnosiscoding,hospitalnumbers,andlaboratory
values)thatcanbeanalysedbyusingstatisticalmethods(e.g.,addressingaregressionor
classificationquestion).
Theunstructureddatatypedoesnothavethatpredefinedstructure.Withoutthe
structuredformat,theyarestoredintheirrawunstructuredform,usuallyinlargetext
filesornonstructureddatasets.Theyarealsocategorisedasqualitativedata,makingit
moredifficulttocollectandanalyse.Thisincludesimages,audio,andvideointheform
oftextsdesignedtocategorisedataintodifferentclassifications.
SupervisedMLalgorithmsusuallyrequirestructureddata(e.g.,videosinwhichall
imagesarecorrectlylabelledwithallrelevantclassifiers,suchaspolyps,diverticula,in‐
flammation,residueetc.).
Ifunstructureddataistobeusedinmachinelearning,specialisedtechniques,such
asdeeplearning,thatrelyonvastamountsofdatawouldberequired.However,thatis
oftenunavailableduetoconfidentialityconcernsortheneedformoreprocedurestogen‐
eratethosedatasets.Onceunstructureddatacanbeused,thenapplicationsinpredictive
analyticscouldbenefitthemost[16].
Therefore,today,wearelookingatusingstructureddatatoconductmachinelearn‐
ingandunstructureddatatoinferfromitbyusingtheAIsystem(e.g.,byprovidingan
80%probabilityofamucosalstructurebeingapolyp).
Diagnostics2023,13,10386of19
2.5.MachineAnalysisoftheImages
Imagesaredigitalisedintoone(blackandwhiteimage)ormanygridsofnumbers
(colourimage)(seeAppendixA,FigureA1foragraphicaldemonstrationoftheidea).
Insteadofonegridofnumbersrepresentingeachpixel,thecolourimagesarerepresented
inthreegrids(red,green,andblue(RGB))stackedtogether.Eachpixel’smagnituderep‐
resentsthecorrespondingcolourbrightnessineachgrid.Inpractice,asingle224×224
pixelcolourimagewouldgenerate150,528numbersorfeaturesforeachimage(SeeAp‐
pendixA,FigureA1,asimplifiedgraphicalrepresentationoftheconcept).Thisdemon‐
stratestheimmensevolumeofdataincorporatedinallthesequencesofimageswithina
coloncapsulevideo,whichtheAIsystemwillhavetoprocesstoproducethedesiredout‐
put10.Toovercomethis,insteadofusingtherawdatafromtheimagesastheinputfea‐
tures,theexpertsadoptasetofhandcraftedfeaturesmanuallyengineeredbytheexperts.
TheselectedhandcraftedfeatureshaveanenormousimpactontheMLanddependon
thetasktobeaddressed.Forexample,simplefeatureslikeedges,corners,orcolourcan
beused[17].
3.TheConceptofNeuralNetworkandAITraining
3.1.NeuralNetwork
Duetothelargequantityofdata,thenumberofparameters,thespatialinformation
betweeneachpixelandthecomplexityofthedatastructure,deeplearning(DL)models
werecreatedtoorganisethesecomplexfeaturesintoarchitecturallayerscalledneuralnet‐
work(NN)buildingblocks.
Thesenetworksaremadeupofnumerousneurons,eachactingasanindividualmin‐
iaturemachinelearningsystem(e.g.,miniatureregressionmodels).Asetofneurons,
whichtakethesameinput,areorganisedintoalayer.Theseneuronsprocessinputsby
usinglinearcombinationmethods,andthelayer’sparametersgenerateanoutputwhich
isthenfedintothenextlayer.Thisprocessisrepeateduntilthefinaloutputisdelivered,
anditissimilartoournervoussystem[15].Inaddition,therearelayersbetweenthefirst
inputlayerandthelastoutputlayer,calledhiddenlayers.Thenumberofhiddenlayers
variesdependingonthecomplexity,function,andassociateddefinedoutputoftheneural
network[18](seeFigure2).
Diagnostics2023,13,10387of19
Figure2.Adenseneuralnetworkdemonstratesthelayers’architecturecomparedwiththenervous
systemmodel.
ThemaindifferencebetweenNNandtraditionalMLmodelsisthatNNworksdi‐
rectlyfromunstructured,rawdatainsteadofhandcraftedfeatures.Whereastraditional
machinelearningalgorithmsrequireanexperttoselecttheproblem’srelevantcharacter‐
istics,NNcanautomaticallyperformthefeatureengineeringtask.
3.2.ConvolutionalNeuralNetwork
Aconvolutionalneuralnetwork(CNN)ismainlydesignedtoprocessimages,and
itsapplicationispopularinmedicalfields,suchasradiologyandendoscopy.Itisde‐
signedtoaddresstwodifficultiesthatastandardneuralnetworkencounterswhenpro‐
cessingimages.First,eventhoughaneuralnetworkcouldincludeandorganisemany
parametersintothesedensehierarchicallayers,eachparameter(neuron)wouldonlybe
allocatedsimultaneouslytooneorasmallnumberofpixellocations.Takingthehighly
variablepositionsoftheobjectinapracticalimage,thenumberofneuronsrequiredis
enormous;thisinevitablyprolongstheprocessingtimeconsiderably.Thesecondissueis
thatthestandardneuralnetworkcannotrecordthespatialinformationintheimageasit
flattenstheimage(theparametersofeachpixelorganisedinspecificspatialorders)intoa
rollofnumbers(vector)(seediagramintheAppendixAformoreinformation)[18].
Consequently,CNNusestheconvolutionallayerstoresolvetheseissuesbyusing
convolutionalfilters(kernels)(Figure3).Thesefilterscomprisesmallgroupsof
Diagnostics2023,13,10388of19
parametersthataremultipliedandsummedinpatches.Theoutputofeachpatchisplaced
relativetothepositionoftheinputpatchinanewsmallergrid.Forexample,anareaof
interest(e.g.,apolyponacoloncapsuleimage)couldrepresentahigh‐valuenumberon
thesmallergrid.
Figure3.ThissimplifieddiagramshowshowCNNprocessestheparametersfromanimageby
usingfilters(kernels)tocondensetheparameterstoasmalleroutputtopreservethespatialinfor‐
mationandimprovethehandlingspeed,astheparametersareanalysedinpatchesratherthanin‐
dividuals.
Theoutputofeachconvolutionallayercanbefedintothenextlayerasan“image
input”.Inthissequence,eachpixelinthenextconvolutionallayerrepresentsapatchof
pixelsinputtedfromthepreviouslayer.Aftergoingthroughvariouslayersofrepeated
processing,theCNNwillbeabletoseetheoveralllargerpatchesoftheimagesandulti‐
matelyproduceoutputprobabilitiesoftheimagecategory[19].
Forexample,thepixelsinthefirstlayerofCNNwillformbasicfeaturessuchassmall
points,lines,andridgesfromtherawpixelontheinputimage.Thesepixelsarethencom‐
binedagaininthesuccessivefewlayers,byusingkernels,intosimpleshapessuchascir‐
cles,squares,andlargedots.Thisprocessrepeatsuntiltheinputdatagoesdeeperintothe
layers.Supposeaspecificcombinationofshapesorfeaturesrepresentingalesionispre‐
sentinthedeeperlayer.Inthatcase,theneuronsinthatlayerwilleventuallyfirethe
processedfeaturestothefinallayerthatpredictstheclassoftheobject(e.g.,polypinthe
imagewithaprobabilityscore(Figure4)).
Diagnostics2023,13,10389of19
Figure4.ThesimplifiedoverallCNNlayersinidentifyingpolypsintheCCEvideoandthe
flowchartdemonstratetheCNNusedinthecoloncapsulevideotopredictpolyps,forexample,
accurately.
3.3.TheAITraining,OptimisationandValidationProcess
Theconvolutedneuralnetworkmodelsorapproximatesanaccuratemappingfunc‐
tionbetweentheinputsandoutputs.Thisrequiresaslowprocessoftraining.
First,theCNNisgivenatrainingdataset,asetofdataexamplesforthemodelto
learnandmapthefunctionthatcorrelatestheinputstotheoutputs.Inthetrainingset,
thedifferenceinerrorbetweentheCNN’spredictionandthetrainingset’slabelwillbe
computedas“loss”.LossisanumericalvaluethatdetermineshowclosetheCNNpre‐
dictedoutputsaretothetrueoutputs.Aftereachrunofthesametrainingset,theCNN
Diagnostics2023,13,103810of19
willupdateitsparameterstoreducetheloss,calledtheoptimisationstep.TheCNNwill
thenbeevaluatedonavalidationdatasettoassessitsperformanceperiodically.Itisim‐
portanttonotethatthevalidationdatasetwasnotexposedtotheCNNduringtraining
andinsteadonlyusedforvalidationwithoutmodifyingthevaluesoftheCNN,i.e.,itwas
notbeingtrained.
Hyperparametersarestudy‐specificoptionalcomponentsorparametersinthetrain‐
ingprogrammethattrainsthemodel.Theyaredefinedmanuallybytheuserbeforethe
modelistrained.Theyshapethemodel’sbehaviouraspartofitsperformanceoptimisa‐
tionbyimpactingitsstructureorcomplexity[20].Theyareappliedinthetrainingloopin
theformofdifferenttrainingconfigurationstotunethemodeloralgorithmbeingtrained.
Thisissubclassifiedintotwotypes[14]:
1. modelhyperparametersthatfocusonthearchitectureofthemodel(e.g.,numberof
layersintheCNN);and
2. traininghyperparametersthatdeterminethetrainingprocess(e.g.,thetrainingrate).
TheseabovestepsformatrainingloopthatallowstheCNNmodeltolearngeneral‐
isedandaccuratefunctionsfromthetrainingsets.Atthesametime,progressisintermit‐
tentlyvalidatedthroughthevalidationdatasets.Finally,themodelwillbeexaminedona
testdatasetoncetheperformanceisfullyoptimisedandvalidated.Thisisanentirely“un‐
seen”setusedattheendofthedevelopmentoftheCNNmodeltoconfirmitsgeneralised
performanceonthesefinalsetsofdatasamples.
Inthetrainingloop,theperformanceoftheCNNisassessedbycomparingthepre‐
dictedoutputproducedbytheCNNagainstthetrueoutput.Low‐valuelossisdesirable
inmachinetraining.Therefore,thetrainingloopaimstodiscoverafunctionwiththebest‐
fittedparameterstominimisethelossacrossallthetrainingdatasets.Thiscanbeillus‐
tratedinasimplestatisticallinearregressionexample,asshowninFigure5[14].
Figure5.Thisusessimplelinearregressionmodelstodemonstratehighandlowlosswhencom‐
paringthepredictedoutputfromAIagainstthetrueoutput.
3.4.ConsequencesofOverfittingandUnderfittingData
Duringmachinelearning,abalanceinthelossneedstobefoundwhenconductinga
trainingloop.WhentheCNNisovertrained(e.g.,inanextendedtrainingperiod),itleads
tooverfitting.ThisisduetotheCNNmodelmemorisingirrelevantfeatures,including
thebackgroundnoisefromthetrainingset,whichiscommonforthesespecificpatients
butnotrelevanttothefinding.Then,theoverallaccuracyofthevalidationsetstartsto
deteriorate.Thesolutionstoovercomeoverfittinginclude
1. theapplicationofalargerdataset,althoughinmedicalimaging,thatmightnotbe
possibleorverycostly;
2. modificationofthemodeltoasimplerversion;and
Diagnostics2023,13,103811of19
3. theutilisationoftechniquessuchasregularisationanddataaugmentation.These
methodsempowertheAImodeltolearnandpreservethegeneralobservationsonly,
allowingtheextrapolationofwhatithaslearnedtonewunseendata.
Ontheotherhand,underfittingisequallydamagingasthisarisesfromneedingto
beabletocapturetheunderlyingfunctionofthedataduetoalackofexposuretothe
trainingsets(inadequatetraining,seeFigure6A)orbecauseofalowcomplexityofthe
model(seeFigure6B).Therefore,achievinganappropriatefittingremainsoneofthesig‐
nificantchallengesinthisfield.
Figure6.Graphsdemonstrateoverfittingandunderfitting.Graph(A)comparestheoverallerror
againstseveralloopsconductedinthetrainingsets.Itshowstheerrorinthevalidationsetuptrends
whenoverfittingoccurswhiletheerrorinthetrainingsetcontinuestodowntrendasthefunction
memorisesallthebackgroundnoiseandnonspecificdetailsinthetrainingsets.Finally,graph(B)
demonstratestheunderfitted,bestfit,andoverfittedconceptsbyusingasimplebest‐fittrendlines
model.
ThefinalstepoftraininganAIisusingtheacompletelynewtestsettodetermine
theAImodel’soverallperformance.Inaclassificationproblem,measuressucharesensi‐
tivity,specificity,accuracy,andprecisionareusuallyused.Moreover,otherglobal
measuressuchasthereceiveroperating(ROC)curveorareaunderROCcurve(AUC)are
widelyusedtocomparemethodsbecausetheydonotdependonanythreshold.
4.ProcessofColonCapsuleEndoscopyVideoAnalysis
TheAmericanSocietyofGastrointestinalEndoscopy(ASGE)and,morerecently,its
Europeanequivalent(ESGE),havesuggestedcredentialingstandardsandcurriculumfor
CEreadingearlyon[21].Atthesametime,itisknownthatnotonlyexperienceinGI
endoscopybutconcentrationcapacityandfatiguecanalsointerferewiththeoutcomesof
Diagnostics2023,13,103812of19
CEreading[21].Althoughthereis,todate,noscientificprooforguidelinestoindicatethe
optimalwaytoreadaCCEvideo,reviewingCCEvideosposesextrachallengesthatare
absentinsmallbowelCE(SBCE)reading.Prolongedsegmentaldelayscompoundbythe
to‐and‐forth,spiralingmovementofthecapsuleinthecaecumandproximalcolon,the
capsule’sbullet‐typepropulsioninmore“muscular”distalcolonicsegmentscompound
withthecolourandturbidityoftheluminalfluid,requirestime,focusedattention,and
accuratelandmarkingforproperevaluationandvideoreview[22].
CCEreadingshouldbeperformedduringprotectedtimeslotstomaintainhigh
standardsandremainathoroughanddiligentprocess,justlikeanyotherendoscopicpro‐
cedure.Admittedly,amassingreadingexperiencecanreducereadingtimes;however,the
officialtimeallocatedforreview/landmarkingofaCCEvideoshouldbeatleast45–65 min
forthefirst/prereadersandatleast25–35 minforthevalidatorsonaverage[21].TheCCE
readingtimerequireddependsonthecleansinglevel,colonanatomy,andtransittimes.
Unfortunately,thesefactorsarenotpredictable.However,itbecomesevidentthatmeth‐
odstoreducethosetimesandefforts,suchasAI,havetobefoundtoreducetheburden
onexpertsand,morebroadly,adoptCCE.
Withoutthosemethods,thefirststepshouldbeaquickpreviewoftheentirevideo:
Thiscanbedonebyusingafastreading(QuickView)modewithbothcameraheadssim‐
ultaneously.Next,oneshouldlookatthetotallengthoftimethecapsuleneededtogo
throughthecolon.Then,theyneedtoidentifythelandmarks(caecum,hepaticandsplenic
flexures,andrectum/anus/excretionofcapsule).
Thesecondstepisaproperreviewoftheimages.Asthecoloncapsuleisdesignedto
havetwocameras,theyarerepresentedbyyellowandgreen.Startingfromtheyellowor
greencamerabutonecameraalone,followedbytheothercameraataframeratebetween
8and15picturespersecond.Adifferentapproachisadvisableifthepassagetimeisshort
ortoolong.Often,thecapsuletendstostagnateincolonicsegmentsasthecolon’smus‐
cular,propulsivemechanismisusuallyweakerthantheSB’spropulsivemechanism.If
thatoccurs,theframeratecouldbeincreased.Onthecontrary,ashortvideomeansthat
thecapsulehasgonethroughthecolonquickly,andtherearefewerframestoview,so
ourrateofframesperminutecouldbedecreasedbyusingthescrollwheel(scrollbutton)
onthecomputermouse.Thisoftenhappensinthetransversecolon,wherethepassage
timecanbequick;alower(pre)readingspeedisadvisedinthissegmenttoavoidmissing
lesions.
Thelaststepisreportingthefindings.Adetailedreviewofthemarkedsuspected
lesionsimages(thumbnails)thatuseswhitelightandvirtualchromoendoscopyforchar‐
acterisationisusedwhenapplicable.Eachrelevantimageisannotatedandattachedby
usingthehospitalreportingordocumentationsystem.Thereportshouldfinaliseallthe
findings,colonicandextracolonic,transittimes,significantfindings,diagnosis,andrec‐
ommendations[22,23].
Theoptimalframerateforathoroughcoloninvestigationwithoutanyriskofmissing
lesionsremainsunanswered.Introducingprucaloprideaspartoftheboosterregimento
improvetheoverallprocedurecompletionrateisbeingexamined.Thisregimenreduces
boththetransitandreadingtimes.However,thisalsopotentiallyincreasestheriskof
missinglesionsasthecapsulespeedsthroughthecolon.Morerobustdataontheframe
rateortheminimumlengthofthevideoisundoubtedlyrequiredinfuturestudies[24,25].
5.Evidence‐BasedLiteratureReviewofAIandCCE
5.1.AIinColonCapsuleEndoscopyintheLiterature
AIincoloncapsuleendoscopyisanewfieldofinterest.Recently,Afonsoetal.[26]
analysed24CCEexams(PillCamCOLON2)performedatasinglecentrebetween2010
and2020.Fromthesevideorecordings,3635framesofthecolonicmucosawereextracted,
770containingcoloniculcersorerosionsand2865showingnormalcolonicmucosa.After
optimisingtheneuralarchitectureoftheCNN,theirmodelautomaticallydetectedcolonic
Diagnostics2023,13,103813of19
ulcersanderosionswithasensitivityof90.3%,specificityof98.8%,andanaccuracyof
97.0%.Theareaunderthereceiveroperatingcharacteristiccurve(AUROC)was0.99.The
meanprocessingtimeforthevalidationdatasetwas11sec(approximately66frames/s).
Saraivaetal.[2]usedCCEimagestodevelopadeeplearning(DL)toolbasedona
CNNarchitecturetodetectprotrudingcoloniclesionsautomatically.ACNNwascon‐
structedbyusingananonymiseddatabaseofCCEimagescollectedfrom124patients.
Thisdatabaseincludedimagesofpatientswithprotrudingcoloniclesions,normalcolonic
mucosa,orotherpathologicfindings.Atotalof5715images(2410protrudinglesions,3305
normalmucosaorotherfindings)wereextractedforCNNdevelopment.Theareaunder
thecurve(AUC)fordetectingprotrudinglesionswas0.99.Thesensitivity,specificity,
PPV,andNPVwere90.0%,99.1%,98.6%,and93.2%,respectively.Theoverallaccuracyof
thenetworkwas95.3%.ThisDLalgorithmaccuratelydetectedprotrudinglesionsinCCE
images.
AtsuoYamadaetal.trainedadeepCNNsystembasedonaSingleShotMultiBox
Detectorbyusing15,933CCEimagesofcolorectalneoplasms,suchaspolypsandcancers
[27].Theyassessedperformancebycalculatingareasunderthereceiveroperatingcharac‐
teristiccurves,alongwithsensitivities,specificities,andaccuraciesbyusinganindepend‐
enttestsetof4784images,including1850imagesofcolorectalneoplasmsand2934normal
colonimages.TheAUCfordetectingcolorectalneoplasiabytheAImodelwas0.90.The
sensitivity,specificity,andaccuracywere79.0%,87.0%,and83.9%,respectively,ataprob‐
abilitycutoffof0.35.
HiroakiSaitoetal.[28]usedadatabaseof30,000VCEimagesofprotrudinglesions
from290patientexaminationstodevelopaCNNmodel.TheCNNmodeldevelopedfrom
thisdatabasewas90%sensitiveand79%specificwhenidentifyingtestimagescontaining
protrudinglesions.Inaddition,subsetanalysesevaluatingmodelperformancefordiffer‐
entlesionsdemonstratedthatthemodelwas86%sensitivefordetectingpolyps,92%sen‐
sitivefordetectingnodules,95%sensitivefordetectingepithelial‐basedtumours,77%
sensitivefordetectingsubmucosallesions,and94%sensitiveforidentifyingprotruding
venousstructures,suchasvarices.
Nadimietal.developedaCNNfortheautonomousdetectionofcolorectalpolyps;
theirCNNwasanimprovedversionofZF‐Net,aCNNusingacombinationoftransfer
learning,preprocessinganddataaugmentation[29].Theycreatedanimagedatabaseof
11,300capsuleendoscopyimagesfromascreeningpopulation,includingcolorectal
polyps(anysizeormorphology,N=4800)andnormalmucosa(N=6500).TheirCNN
modelresultedinanevenbetterperformancewithanaccuracyof98.0%,asensitivityof
98.1%,andaspecificityof96.3%.(SeeAppendixATableA1forthesummaryoftheabove
results)
5.2.AIAssessmentofCCEBowelCleansing
InapilotstudyconductedbyBuijsetal.,anonlinearindexbasedonthepixelanalysis
modelandamachinelearningalgorithmbasedonthesupportvectormachineswithfour
cleanlinessclasses(unacceptable,poor,fair,andgood)weredevelopedtoclassifytheCCE
videosof41screeningparticipants[30].Theresultsofbothmodelswerecomparedto
cleanlinessevaluationsbyfourCCEreaders.TheML‐basedmodelclassified47 %ofthe
videosinagreementwiththeaveragedclassificationbyCCEreaders,comparedto32 %
bythepixelanalysismodel.Inaddition,theML‐basedmodelwassuperiortothepixel
analysisinclassifyingbowelcleansingqualityduetoahighersensitivitytounacceptable
andpoorcleansingquality.
Inanotherstudy[31],aCACscore,definedasthecolourintensities’redovergreen
(R/G)ratioandredoverbrown(R/(R + G)ratio,wasdeveloped.Bowelcleansingevalua‐
tionforeachCCEframewasdefinedaseitheradequatelyorinadequatelycleansed.Four‐
hundred‐and‐eightframeswereextracted.Twohundredsixteenstillframeswerein‐
cludedintheR/Gsetand192intheR/(R + G)set.RegardingtheR/Gratio,athreshold
valueof1.55wascalculated,withasensitivityof86.5 %andaspecificityof77.7%.
Diagnostics2023,13,103814of19
RegardingtheR/(R + G)ratio,athresholdvalueof0.58wascalculatedwithasensi‐
tivityof95.5 %andaspecificityof62.9 %.ThetwoproposedCACscoresbasedontheratio
ofcolourintensitieshavehighsensitivitiesfordiscriminatingbetween“adequately
cleansed”and“inadequatelycleansed”CCEstillframes.TheirstudyshowedthatCAC
scorestoassessbowelpreparationqualitybasedonacolourintensityratioofredand
greenpixelsonstillimagesisfeasibleandrapid(seeAppendixATableA2forthesum‐
maryoftheaboveresults).
6.ChallengesofUtilisingAIinEndoscopyVideoSettings
6.1.UnderstandingtheInputDataUsedbytheAI
Oneofthemainchallengesofthedeepneuralnetworkistheneedtounderstand
whatsignalsthemodelhasextractedfromtheinputtodrawuptheassociationbetween
theinputdataandthepredictedoutput.AstheAIcreatesitsproblem‐solvingmethods,
theprocessisentirelyindependentoftheprogrammer.AnexamplewouldbeutilisingAI
topredictfracturesonanklex‐rays.TheAIcancorrectlypredictfracturesbasedoniden‐
tifyingthearrowsthattheradiographersdrewtoindicatetheareaofinterestinsteadof
detectingthediscontinuationoftheoutlineofthebone.However,themodeldrewacon‐
clusionbasedonnonmedicalsignals,andtheoutcomewasconsideredaccurateeven
thoughitwasentirelyincorrect.Thisisaclassicexampleofaccidentallyfittingconfound‐
ersratherthanthetruesignal[32,33].
6.2.The“BlackBox”orUninterpretableAIAlgorithm
Withthecomplexityofthedeeplearningneuralnetwork,itisverychallengingto
interprettheAI’sprocessingmethodsbeforearrivingatthefinaloutput,whichisreferred
toastheneuralnetworkblackbox.Themorecomplextheneuralnetworkis,themore
accuratebutlessinterpretableitbecomes.Forexample,astudentcouldcomeupwiththe
answertoamathematicalquestionwithoutshowinganyworkingsteps.Asaresult,itis
noteasytounderstandhowthestudentreachesthesolutionintheend,whichleadsto
concernaboutunderstandingtheprinciples.Theneedformoreclarityandinterpretability
intheseneuralnetworksbecomesasignificantobstacleintheprogressionofAIdevelop‐
mentinthemedicalfield(seeAppendixAFigureA2foragraphicalrepresentationofthe
concept).
Moreover,poorinterpretabilityimpliesmorechallengingadjustmentstothemodel
forimprovement.Toovercomethis,approachessuchasinvolvingamultidisciplinary
teamtoreviewthefalsepositiveandfalsenegativeresultspredictedbythemodeland
testingthemodelonanexternaldatabaseareadopted[34].
6.3.PoorDifferentiationBetweenCorrelationVersusCausation
Inaddition,theAImodelwillnotbeabletodifferentiatethecorrelationorcausation
associationbetweentheinputandoutputdata.AgoodexampleisanAImodelcorrelating
theincreasingnumberofdrowningcasesintheswimmingpoolwiththegrowingice
creamsalesattheentranceinthesummer.Therefore,itconcludesthatgrowthinicecream
salescausesanincreaseindrowningincidentsintheswimmingpoolwhenweknowthat
bothofthesefactorscorrelatetothehotweatherinthesummer.
6.4.TheImportanceofDataQuality
InthecontextofartificialintelligenceinCCE,dataqualityismoreimportantthanthe
neuralnetworkalgorithmordataengineeringtechniques.“Garbagein,garbageout”is
commonlyusedinartificialintelligenceengineering.Thisreferstothefactthatthechosen
datashouldbehighquality,reliable,consistentandreproducible.Unfortunately,inCCE,
awidevariationinquantifyingthequalityofbowelpreparationandthebubbleeffect
amongexpertsisagoodexample.Thelackofaccuratedefinitionsforthesecomponents
Diagnostics2023,13,103815of19
compromisesthedataqualityandremainsproblematicforAIdevelopmentinthefieldof
CE.
6.5.Generalisabilityand“NoOneSizeFitsAll”
Inaddition,samplingstrategiesandtrainingpractices,suchassingle‐institution
data,smallgeographicareasampling,orotherapproaches,cancreateunintentionalbias
andreducegeneralisation.Forexample,aCCE’sAIdevelopedbasedonanEnglishpop‐
ulation’scolonimagesmightnotapplytoanAsianpopulation.Thisconceptisequivalent
tosamplingerrorinstatisticalterms.Therefore,thefeasibilityandaccuracyoftheAIto
adapttovariousmedicalimagingtechniquesindiversegeographicalandracialpopula‐
tionsstillrequiresfurtherexplorationandexaminationinfuturestudies.
Oneofthepotentialsolutionsisthepossibilityofsharingdatasetsamongdifferent
countriestocontributetobuildinganAIwithalarge,heterogenous,andmultinational
superalgorithmthatallowsaccuratedataprocessingfromanydatasetintheworld.How‐
ever,theharmonisationofimagesissimilarlyessential.Moreover,inthecontextofmul‐
tiinstitutionaldatasources,thereisapotentialriskforvariableequipment,protocols,etc.,
whichcanequallyaffecttheaccuracyoftheAIoutputs[35].
7.FutureofAIinGastroenterology
Accuratelyanalysingcapsuleendoscopyisatime‐consumingtaskforcliniciansde‐
pendingonthecomfortandexpertiseofthereader[35,36].UsingAIcanreducethattime
byhelpingcliniciansduringanalysisandreducediagnosticerrorsduetohumanlimita‐
tionssuchasbiasesandfatigue.Thiswouldultimatelyleadtomoretimeforcliniciansto
focusontraininganddiagnosingpathologies.Thiswirelessandpatient‐friendlytech‐
nique,combinedwithrapidreadingplatformsandthehelpofartificialintelligence,will
becomeanattractiveandviablechoicetoalterhowpatientsareinvestigatedinthefuture
withingastroenterology[37].WiththegrowthoftelemedicinesteppedupbytheCOVID‐
19pandemic,alargepartofspecialistcarewillcontinuetobeperformedremotely.As
CCEbecomesmoreestablished,ithasenormouspotentialintelemedicinesettings.
Withthatinmind,thereareconcernsaboutfuturejobsinthegastroenterologysector
beingreplacedbyAIautomation.However,insteadofjobreplacement,weanticipatethe
shifttowardjobdisplacementbyfocusingmoreresourcesonthetasksthatarenoteasily
automated,suchasclinicianandpatientinteraction,morecomplexprocedures,complex
decision‐making,education,andtraining.Inaddition,newjobsorindustries,suchas
medicalmachinelearningengineering,mightberequiredinthefuturemedicalhealth
system.
Thehuman–AIpartnershipwouldsuggestthatthemachinecannotbeusedalone.
Furthermore,overdependenceonAIwouldundoubtedlyleadtodeskilling,especiallyin
theformofcognitivework,suchaspolypdetectionandrecognition.Therefore,thekeyto
integratingAIintogastroenterologyshouldfocusonbalancingAIautomationandthe
personalcarewevalueforourpatientstoprovideanefficientandcost‐effectiveendos‐
copyserviceinthefuture[38–40].
8.Conclusions
Inthefuture,AIisexpectedtooffermultiplebeneficialapplicationsinGIdiseaserisk
stratification,lesionrecognitionandassessment,diagnosis,andtreatment.Theprogress
inthelastdecadesuggeststhatAI‐aidedCCEwillbeavailablesoonandradicallytrans‐
formmedicalpracticeandpatientcare.Understandingthefundamentalsandthebasic
conceptsofmachinelearningtechnologywillnotonlystrengthenthetrustinAIamong
clinicalprofessionalsbutpreventanyunintendedpitfallsinAIapplicationsinfutureclin‐
icalpractice.ThismayallowfutureAIrefinementoroptimisationwithamultidisciplinary
teamapproach.
Diagnostics2023,13,103816of19
Withthecurrentethicaluncertaintyandchallenges,futuremulticentre,randomised
trials,whichvalidateAImodels,shouldfocusonansweringthefundamentalquestionof
whetherAImodelscanenhancephysicianperformancesafelyandreliably.Intheend,a
robustmultidisciplinarycollaborationamongphysicians,computerscientists,andentre‐
preneursisrequiredtopromoteAI’sclinicaluseinmedicalpractice[38‐40].
AuthorContributions:Conceptualization,I.I.L.,E.W.andG.J.N.;validation,S.S.,A.W.,H.W.,A.K.,
A.J.M.W.andR.P.A.;literaturereview,I.I.L.andG.J.N.;resources,S.S.;writing—originaldraftprep‐
aration,I.I.L.,E.W.,A.K.andG.J.N.;writing—reviewandediting,I.I.L.,S.S.,A.W.,H.W.,A.J.M.W.,
A.K.andR.P.A.;visualisation,I.I.L.Allauthorshavereadandagreedtothepublishedversionof
themanuscript.
Funding:Thisresearchreceivednoexternalfunding
DataAvailabilityStatement:Notapplicable.
ConflictsofInterest:Theauthorsdeclarenoconflictofinterest.
AppendixA
TableA1.SummaryofCNNperformancefordetectionofcoloniclesions.
StudyNo.of
Images
Colonic
Lesion
Normal
Colonic
Mucosa
SensitivitySpecificity
Accuracy
ofthe
Network
AUROCfor
Detectionof
Protruding
Lesion
Afonso
[26]3635770286590.3%98.8%97.0%0.99
Saraiva[2]57152410330590.0%99.1%95.3%0.99
Atsuo
Yamada
[27]
47841850293479.0%87.0% 0.902
Hiroaki
Saito[28]17,507750710,00090.0%79.0% 0.911
Nadimi,
E.S[14]16954800650098.1%96.3%98.0%
TableA2.SummaryofthetwostudiesonAIassessmentofCCEbowelcleanliness.
StudyTypeofAI
Numberof
Videos/Frames
Analysed
Levelof
AgreementAI
withReaders,%
Sensitivityspecificity
Buijs[30]
Non‐linear
indexmodel
SVMmode
41videos
41videos
32%
47%
_
_
_
_
Becq[31]R/Gratio
R/(R+G)ratio
216frames
192frames
‐
‐
86.5%
95.5%
77.7%
62.9%
Diagnostics2023,13,103817of19
FigureA1.Anexampleofmappinglocationsintheimagestothepixelvalueaspartofmachine
analysisofthepicture.
FigureA2.TheintrinsicmethodorbehaviouroftheAIcodeinthemodelisuninterpretable,likea
blackboxwithnotransparency.
References
1. Bejnordi,B.E.;Veta,M.;VanDiest,P.J.;VanGinneken,B.;Karssemeijer,N.;Litjens,G.;VanDerLaak,J.A.W.M.;Hermsen,M.;
Manson,Q.F.;Balkenhol,M.;etal.DiagnosticAssessmentofDeepLearningAlgorithmsforDetectionofLymphNodeMetas‐
tasesinWomenwithBreastCancer.JAMA2017,318,2199–2210.https://doi.org/10.1001/jama.2017.14585.
2. Saraiva,M.M.;Ferreira,J.P.S.;Cardoso,H.;Afonso,J.;Ribeiro,T.;Andrade,P.;Parente,M.P.L.;Jorge,R.N.;Macedo,G.Artificial
intelligenceandcoloncapsuleendoscopy:Developmentofanautomateddiagnosticsystemofprotrudinglesionsincoloncap‐
suleendoscopy.Tech.Coloproctol.2021,25,1243–1248.https://doi.org/10.1007/s10151‐021‐02517‐5.
3. Bjørsum‐Meyer,T.;Koulaouzidis,A.;Baatrup,G.Commenton‘Artificialintelligenceingastroenterology:Astate‐of‐the‐artre‐
view’.WorldJ.Gastroenterol.2022,28,1722–1724.
4. Robertson,A.R.;Segui,S.;Wenzek,H.;Koulaouzidis,A.Artificialintelligenceforthedetectionofpolypsorcancerwithcolon
capsuleendoscopy.Ther.Adv.Gastrointest.Endosc.2021,14,26317745211020277.https://doi.org/10.1177/26317745211020277.
5. Dray,X.;Iakovidis,D.;Houdeville,C.;Jover,R.;Diamantis,D.;Histace,A.Artificialintelligenceinsmallbowelcapsuleen‐
doscopy—Currentstatus,challengesandfuturepromise.J.Gastroenterol.Hepatol.2021,36,12–19.
6. Qin,K.;Li,J.;Fang,Y.;Xu,Y.;Wu,J.;Zhang,H.;Li,H.;Liu,S.;Li,Q.Convolutionneuralnetworkforthediagnosisofwireless
capsuleendoscopy:Asystematicreviewandmeta‐analysis.Surg.Endosc.2021,36,16–31.https://doi.org/10.1007/s00464‐021‐
08689‐3.
7. Soffer,S.;Klang,E.;Shimon,O.;Nachmias,N.;Eliakim,R.;Ben‐Horin,S.;Kopylov,U.;Barash,Y.Deeplearningforwireless
capsuleendoscopy:Asystematicreviewandmeta‐analysis.Gastrointest.Endosc.2020,92,831–839.e8.
https://doi.org/10.1016/j.gie.2020.04.039.
8. Horie,Y.;Yoshio,T.;Aoyama,K.;Yoshimizu,S.;Horiuchi,Y.;Ishiyama,A.;Hirasawa,T.;Tsuchida,T.;Ozawa,T.;Ishihara,S.;
etal.Diagnosticoutcomesofesophagealcancerbyarti‐ficialintelligenceusingconvolutionalneuralnetworks.Gastrointest.
Endosc.2019,89,25–32.
9. Wang,P.;Berzin,T.M.;GlissenBrown,J.R.;Bharadwaj,S.;Becq,A.;Xiao,X.;Liu,P.;Li,L.;Song,Y.;Zhang,D.;etal.Real‐time
automaticdetectionsystemincreasescolonoscopicpolypandadenomadetectionrates:Aprospectiverandomisedcontrolled
study.Gut2019,68,1813–1819.
10. Aoki,T.;Yamada,A.;Aoyama,K.;Saito,H.;Tsuboi,A.;Nakada,A.;Niikura,R.;Fujishiro,M.;Oka,S.;Ishihara,S.;etal.Auto‐
maticdetectionoferosionsandulcerationsinwirelesscapsuleendoscopyimagesbasedonadeepconvolutionalneuralnetwork.
Gastrointest.Endosc.2019,89,357–363.e2.https://doi.org/10.1016/j.gie.2018.10.027.
11. Aoki,T.;Yamada,A.;Aoyama,K.;Saito,H.;Fujisawa,G.;Odawara,N.;Kondo,R.;Tsuboi,A.;Ishibashi,R.;Nakada,A.;etal.
Clinicalusefulnessofadeeplearning‐basedsystemasthefirstscreeningonsmall‐bowelcapsuleendoscopyreading.Dig.En‐
dosc.2019,32,585–591.https://doi.org/10.1111/den.13517.
Diagnostics2023,13,103818of19
12. Moen,S.;Vuik,F.E.R.;Kuipers,E.J.;Spaander,M.C.W.ArtificialIntelligenceinColonCapsuleEndoscopy—ASystematicRe‐
view.Diagnostics2022,12,1994.https://doi.org/10.3390/diagnostics12081994.
13. Min,J.K.;Kwak,M.S.;Cha,J.M.OverviewofDeepLearninginGastrointestinalEndoscopy.GutLiver2019,13,388–393.
https://doi.org/10.5009/gnl18384.
14. VanderSommen,F.;deGroof,J.;Struyvenberg,M.;vanderPutten,J.;Boers,T.;Fockens,K.;Schoon,E.J.;Curvers,W.;deWith,
P.;Mori,Y.;etal.MachinelearninginGIendoscopy:Practicalguidanceinhowtointerpretanovelfield.Gut2020,69,2035–
2045.
15. Deo,R.C.MachineLearninginMedicine.Circulation2015,132,1920.
16. Tayefi,M.;Ngo,P.;Chomutare,T.;Dalianis,H.;Salvi,E.;Budrionis,A.;Godtliebsen,F.Challengesandopportunitiesbeyond
structureddatainanalysisofelectronichealthrecords.WIREsComput.Stat.2021,13,e1549.https://doi.org/10.1002/wics.1549.
17. Cumberlin,R.L.;Rodgers,J.E.;Fahey,F.H.Digitalimageprocessingofradiationtherapyportalfilms.Comput.Med.Imaging
Graph.1989,13,227–233.https://doi.org/10.1016/0895‐6111(89)90129‐8.
18. Yu,H.;Samuels,D.C.;Zhao,Y.‐Y.;Guo,Y.Architecturesandaccuracyofartificialneuralnetworkfordiseaseclassificationfrom
omicsdata.BMCGenom.2019,20,167.https://doi.org/10.1186/s12864‐019‐5546‐z.
19. Missert,A.D.;Yu,L.;Leng,S.;Fletcher,J.G.;McCollough,C.H.Synthesizingimagesfrommultiplekernelsusingadeep
convolutionalneuralnetwork.Med.Phys.2019,47,422–430.https://doi.org/10.1002/mp.13918.
20. Luo,G.Areviewofautomaticselectionmethodsformachinelearningalgorithmsandhyper‐parametervalues.Netw.Model.
Anal.HealthInform.Bioinform.2016,5,18.https://doi.org/10.1007/s13721‐016‐0125‐6.
21. Faigel,D.O.;Baron,T.H.;Adler,D.G.;Davila,R.E.;Egan,J.;Hirota,W.K.;Jacobson,B.C.;Leighton,J.A.;Qureshi,W.;Rajan,E.;
etal.ASGEguideline:Guidelinesforcredentialingandgrantingprivilegesforcapsuleendoscopy.Gastrointest.Endosc.2005,61,
503–505.https://doi.org/10.1016/s0016‐5107(04)02781‐6.
22. Beg,S.;Card,T.;Sidhu,R.;Wronska,E.;Ragunath,K.;Ching,H.‐L.;Koulaouzidis,A.;Yung,D.;Panter,S.;Mcalindon,M.;et
al.Theimpactofreaderfatigueontheaccuracyofcapsuleendoscopyinterpretation.Dig.LiverDis.2021,53,1028–1033.
https://doi.org/10.1016/j.dld.2021.04.024.
23. Koulaouzidis,A.;Dabos,K.;Philipper,M.;Toth,E.;Keuchel,M.Howshouldwedocoloncapsuleendoscopyreading:A
practicalguide.Ther.Adv.Gastrointest.Endosc.2021,14,26317745211001983.https://doi.org/10.1177/26317745211001983.
24. ElHajjar,A.;Rey,J.F.Artificialintelligenceingastrointestinalendoscopy:Generaloverview.Chin.Med.J.2020,133,326.
25. Pan,G.;Yan,G.;Qiu,X.;Cui,J.BleedingDetectioninWirelessCapsuleEndoscopyBasedonProbabilisticNeuralNetwork.J.
Med.Syst.2010,35,1477–1484.https://doi.org/10.1007/s10916‐009‐9424‐0.
26. Mascarenhas,M.;Afonso,J.;Ribeiro,T.;Cardoso,H.;Andrade,P.;Ferreira,J.P.S.;Saraiva,M.M.;Macedo,G.Performanceofa
DeepLearningSystemforAutomaticDiagnosisofProtrudingLesionsinColonCapsuleEndoscopy.Diagnostics2022,12,1445.
https://doi.org/10.3390/diagnostics12061445.
27. Yamada,A.;Niikura,R.;Otani,K.;Aoki,T.;Koike,K.Automaticdetectionofcolorectalneoplasiainwirelesscoloncapsule
endoscopicimagesusingadeepconvolutionalneuralnetwork.Endoscopy2021,53,832–836.
28. Saito,H.;Aoki,T.;Aoyama,K.;Kato,Y.;Tsuboi,A.;Yamada,A.;Fujishiro,M.;Oka,S.;Ishihara,S.;Matsuda,T.;etal.Automatic
detectionandclassificationofprotrudinglesionsinwirelesscapsuleendoscopyimagesbasedonadeepconvolutionalneural
network.Gastrointest.Endosc.2020,92,144–151.e1.https://doi.org/10.1016/j.gie.2020.01.054.
29. Nadimi,E.S.;Buijs,M.M.;Herp,J.;Kroijer,R.;Kobaek‐Larsen,M.;Nielsen,E.;Pedersen,C.D.;Blanes‐Vidal,V.;Baatrup,G.
Applicationofdeeplearningforautonomousdetectionandlocalizationofcolorectalpolypsinwirelesscoloncapsule
endoscopy.Comput.Electr.Eng.2019,81,106531.https://doi.org/10.1016/j.compeleceng.2019.106531.
30. Buijs,M.M.;Ramezani,M.H.;Herp,J.;Kroijer,R.;Kobaek‐Larsen,M.;Baatrup,G.;Nadimi,E.S.Assessmentofbowelcleansing
qualityincoloncapsuleendoscopyusingmachinelearning:Apilotstudy.Endosc.Int.Open2018,6,E1044–E1050.
https://doi.org/10.1055/a‐0627‐7136.
31. Becq,A.;Histace,A.;Camus,M.;Nion‐Larmurier,I.;Ali,E.A.;Pietri,O.;Romain,O.;Chaput,U.;Li,C.;Marteau,P.;etal.
Developmentofacomputedcleansingscoretoassessqualityofbowelpreparationincoloncapsuleendoscopy.Endosc.Int.
Open2018,6,E844–E850.https://doi.org/10.1055/a‐0577‐2897.
32. Kelly,C.J.;Karthikesalingam,A.;Suleyman,M.;Corrado,G.;King,D.Keychallengesfordeliveringclinicalimpactwith
artificialintelligence.BMCMed.2019,17,195.https://doi.org/10.1186/s12916‐019‐1426‐2.
33. Meskó,B.;Görög,M.Ashortguideformedicalprofessionalsintheeraofartificialintelligence.NPJDigit.Med.2020,3,126.
34. Ting,D.S.W.;Pasquale,L.R.;Peng,L.;Campbell,J.P.;Lee,A.Y.;Raman,R.;Tan,G.S.W.;Schmetterer,L.;Keane,P.A.;Wong,
T.Y.;etal.Artificialintelligenceanddeeplearninginophthalmology.Br.J.Ophthalmol.2019,103,167–175.
35. Glocker,B.;Robinson,R.;deDou,Q.C.;Konukoglu,E.MachineLearningwithMulti‐SiteImagingData:AnEmpiricalStudy
ontheImpactofScannerEffects.arXiv2019,arXiv:1910.04597.
36. Zheng,Y.P.;Hawkins,L.;Wolff,J.;Goloubeva,O.;Goldberg,E.Detectionoflesionsduringcapsuleendoscopy:Physician
performanceisdisappointing.Am.J.Gastroenterol.2012,107,554–560.
37. ChetcutiZammit,S.;Sidhu,R.Capsuleendoscopy—Recentdevelopmentsandfuturedirections.ExpertRev.Gastroenterol.
Hepatol.2021,15,127–137.
38. Chen,M.;Decary,M.Artificialintelligenceinhealthcare:Anessentialguideforhealthleaders.HealthManag.Forum.2019,33,
10–18.https://doi.org/10.1177/0840470419873123.
Diagnostics2023,13,103819of19
39. Parasher,G.;Wong,M.;Rawat,M.Evolvingroleofartificialintelligenceingastrointestinalendoscopy.WorldJ.Gastroenterol.
2020,26,7287–7298.https://doi.org/10.3748/wjg.v26.i46.7287.
40. Dinga,R.;Penninx,B.W.;Veltman,D.J.;Schmaal,L.;Marquand,A.F.Beyondaccuracy:Measuresforassessingmachinelearning
models,pitfallsandguidelines.bioRxiv2019,743138.https://doi.org/10.1101/743138.
Disclaimer/Publisher’sNote:Thestatements,opinionsanddatacontainedinallpublicationsaresolelythoseoftheindividual
author(s)andcontributor(s)andnotofMDPIand/ortheeditor(s).MDPIand/ortheeditor(s)disclaimresponsibilityforanyinjury
topeopleorpropertyresultingfromanyideas,methods,instructionsorproductsreferredtointhecontent.