Improving
burn
depth
assessment
for
pediatric
scalds
by
AI
based
on
semantic
segmentation
of
polarized
light
photography
images
Marco
Domenico
Cirillo
a,b,c,*
,
Robin
Mirdell
d,e,f,
Folke
Sjöberg
d,e,f,
Tuan
D.
Pham
g,**
a
DepartmentofBiomedicalEngineering,LinköpingUniversity,Linköping,Sweden
b
CentreforMedicalImageScienceandVisualization,LinköpingUniversity,Linköping,Sweden
cCenterforMedicalImageScienceandVisualization,LinköpingUniversity,Linköping,Sweden
dTheBurnCentre,LinköpingUniversityHospital,Linköping,Sweden
eDepartmentofPlasticSurgery,HandSurgery,andBurns,LinköpingUniversity,Linköping,Sweden
fDepartmentofClinicalandExperimentalMedicine,LinköpingUniversity,Linköping,Sweden
g
CenterforArtificialIntelligence,PrinceMohammadBinFahdUniversity,Khobar,SaudiArabia
a
b
s
t
r
a
c
t
Thispaperillustratestheefficacyofanartificialintelligence(AI)(aconvolutionalneural
network,basedontheU-Net),fortheburn-depthassessmentusingsemanticsegmentationof
polarizedhigh-performancelightcameraimagesofburnwounds.Theproposedmethodis
evaluatedforpaediatricscaldinjuriestodifferentiatefourburnwounddepths:superficial
partial-thickness(healingin07days),superficialtointermediatepartial-thickness(healing
in813days),intermediatetodeeppartial-thickness(healingin1420days),deep
partial-thickness(healingafter21days)andfull-thicknessburns,basedonobservedhealingtime.
Intotal100burnimageswereacquired.Seventeenimagescontainedall4burndepthsand
wereusedtotrainthenetwork.Leave-one-outcross-validationreportsweregeneratedand
anaccuracyanddicecoefficientaverageofalmost97%wasthenobtained.Afterthat,the
remaining83burn-woundimageswereevaluatedusingthedifferentnetworkduringthe
cross-validation,achievinganaccuracyanddicecoefficient,bothonaverage92%.
Thistechniqueoffersaninterestingnewautomatedalternativeforclinicaldecisionsupportto
assessandlocalizeburn-depthsin2Ddigitalimages.Furthertrainingandimprovementofthe
underlyingalgorithmbye.g.,moreimages,seemsfeasibleandthuspromisingforthefuture.
©2021TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBY
license(http://creativecommons.org/licenses/by/4.0/).
a
r
t
i
c
l
e
i
n
f
o
Articlehistory: Availableonlinexxx Keywords: Artificialintelligence DeeplearningConvolutionalneuralnetworks
U-Net
Semanticsegmentation
Paediatricburns
1.
Introduction
Burnwoundsoccurwhentheskincomesincontactwithfire,
hotwater,electricity,orchemicals.Dependingontemperature
and contact duration with the skin, different burn depths
develop.Burndepthmaybeclassifiedintoseparatelevels[1]:
superficial partial-thickness (I), superficial to intermediate
partial-thickness (II), intermediate to deep partial-thickness
(III), deep partial and full-thickness burns (IV). Importantly,
* Correspondingauthorat:DepartmentofBiomedicalEngineering,LinköpingUniversity,Linköping,Sweden.
** Correspondingauthor.
E-mailaddresses:marco.domenico.cirillo@liu.se(M.D.Cirillo),tpham@pmu.edu.sa(T.D. Pham).
https://doi.org/10.1016/j.burns.2021.01.011
0305-4179/©2021TheAuthors.PublishedbyElsevierLtd.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.
org/licenses/by/4.0/).ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/).
Available
online
at
www.sciencedirect.com
ScienceDirect
burns of deep partial or full-thickness depth benefit from excision
andskingraftingtohealappropriately.Patientslessthan4years
oldwhogetaburnduetohotwater(Scald)representthe3040%
ofthepatientsarrivingataBurnCentre.Beinganagedefined
andahomogenousgroupfacingburnsmainlyonthetrunkand
armstheywerechosenforthisevaluation.
Burndepthsarecorrectlyclassifiedbyexpertclinicianswith
anaccuracyaround6476%andaround50%bynon-expert
clinicians[27].Today,onetoolthathasbeenusedsuccessfully
asadecisionsupportforcliniciansarebasedonlaserDoppler
[8,9]andonitsmostrecentdevelopment:laserspecklecontrast
imaging(LSCI)[1012].Suchinstrumentshavebeenadvocated
inordertoimproveburndepthsassessmentandtheyareused
occasionallybycliniciansasadecisionsupport device[13].
Thesetechniquesprovideperfusionimagesoftheinjuredskin.
Shortcomingsinclude that they require training and knowledge
tobefullyoperationalandmostimportantlyisthattheimage
generatingprocedureischallengingandthustimeconsuming.
Thishasledtothelimitedclinicaluseofthemethodology.From
anaccuracyperspective,thetechnique also requires at leasttwo
consecutivemeasurementstobeabletoclassifytheburndepth
withreliableaccuracy[10].
For thesereasons, an automatic, fast, objective,and accurate
methodissoughttoevaluatesuchtypesofinjuriesandwiththe
goaltohelpclinicians(decisionsupport),decideifapatientwill
bemefitfromsurgicaltreatmentoftheburnwoundornot.
1.1. AIbasedburndepthassessmentbysemanticimage segmentation
Artificial intelligence based on Convolutional neural networks for
semantic image segmentation as fully convolutional neural
networks[14],SegNet[15],U-Net[16],etc.becameveryattractive
modelsinmedicinebecausetheycombinelocalandglobalimage
information after which a pixel-wise based classification is
provided[17].Theonlydisadvantageisthatthesemodelsrequire
ademandinglearningandtrainingprocessatthebeginning(bya
large computer calculating capacity), but after that, they compute
separateimagesegmentationinafewseconds.Duringthelast
years,theU-Nethasbecomequitepopularinthemedicalfield,so
thatmanymodifiedU-Netswerecreatedandappliedinmedical
applications.Forexample:V-Net[18],tosegmenttheprostate;
DUnet[19],tosegmentretinal vessels;H-DenseUNet[20],for
segmentliverandtumoursinit;AttentionU-Net[21],tosegment
thepancreas;andNonew-Net[22](2ndplacewinnerinBraTS
2018challenge),tosegmentbraintumours.Inthispaperweused
amodifiedU-Netwithresidualstosegmentfourdifferent
burn-depths(superficialpartial-thickness(I),superficialto
intermedi-atepartial-thickness(II),intermediatetodeeppartial-thickness
(III),deeppartialandfull-thickness(IV))inimagesgeneratedbya
high-performancelightcamerawithpolarisationfilterswiththe
aimtoprovideautomatedandobjectiveimagestobeusedbythe
burnsurgeonfortheburn-woundassessmentsupport.
2.
Method
2.1. Patientpopulation
Consecutivelyarrivingchildren,intheagerange04years,at
the outpatient clinic at the Linköping Burn Centre were
included. Laser Dopplerand laserDopplerSpeckleimaging
datafromthiscohorthaspreviouslybeenpresentedinaseries
ofpublications [2,3,10,11,23,24].Inshort, thepatientswere
anesthetised rectallywithketamin[25]andthewoundbed
wasproperlycleanedpriortoimagecapture.Imagecapturing
was doneinaclimatecontrolledroomwithregularindoor
lightning (no windows). For this study, based on a
high-performancelightcamera,imagesweretakeninparalleltothe
onepresentedinthepreviouspublication[3].
2.2. Data
Onehundredburnwoundimageswereacquiredfrompatients
withageequalorless4yearsoldusingaTiVi700,whichisa
tissue viability imagingdevice(WheelsBridgeAB, Sweden).
TiVi700isahigh-performancedigitalcameraequippedwith
polarisationfiltersandflashlightsallarounditslenstoavoid
the reflectingartefact duetoroom lightand/orthecamera
flashandburnwoundfluid.
AnexampleofsuchdataisgiveninFig.1a,whichshowsa
burnwoundimagecapturedbytheTiVi700;whereasFig.1b
showsitsground-truthlabelledmanuallybyaburnclinician
expertoftheLinköpingUniversityHospitalBurnCentre.The
ground-truths,astheoneinFig.1b,weredefinedbasedonthe
wound'shealingtime:asuperficialpartial-thicknesswound
healed within7days,a superficialtointermediate
partial-thicknesshealedbetween813days;anintermediatetodeep
partial-thicknesshealedwithin1420days;andadeeppartial
or full-thickness, which did not heal within 20 days and
underwent surgery. Importantly, surgery was always done
after day 20, which gives ground-truth a high degree of
reliability as all children were observed until day 20 and
healingearlierthanthatwasrecordedbyoneclinician.These
earlierhealingeventsweredividedinto re-epithelialization
within7,14or21days,respectively.
Thetargetofthisprojectisachievingasegmentationresult,
asinFig.1b,fromaburnwoundimage,asinFig.1a,using
artificialintelligence,morespecificallyaconvolutionalneural
network,similartotheU-NetproposedbyRonnebergeretal.
[16],but withdifferent depth,lossfunction,optimizer, and
applyingtheresidualstheoryonit.
Since each burn wound image has a really complex
backgroundrichofobjects(i.e.healthyskin,blanket,medical
tools,nurses’gloves,monitors,etc.),thisisremovedinorderto
lettheCNNonlyfocusesonsegmentingtheregionofinterest,
theburnwound,anddistinguishbetweenthefourdifferent
burn-depths.
Convolutional neural networks minimise the dice loss
[18,26]toachieveagoodsegmentationresultratherthanthe
moregenerallyusedcross-entropyloss,becausetheformer
doesnotcountinthetruenegatives(thebackground),which
normallyhavethemajornumberofpixelsintheimage.The
higherthedicecoefficientisthehighertheaccuracyis,butthe
contraryisnottrue.Thedicelossismathematicallydefinedas
DL¼1D¼12 PC c¼1wcPNngcnpcn PC c¼1wcPNngcnþpcn ; (1)
whereDLstandsforthediceloss,Dfordicecoefficient,Cfor
numberofclasses,Nfornumberofpixels,wcfortheweight
2
burns xxx (2021) xxx xxxassignedtoclassc,gcnandpcnforthen-thpixelwhichbelongsto
theground-truthandtothenetwork’spredictiononc-thclass
respectively.Whenwcisnotavectorofones,Eq.(1)represents
thegeneralizeddice-loss.Otherwise,wcisdefinedas
wc¼
Ns
CNc;
(2)
whereNs isthe numberofpixelsinthe imageand Ncthe
numberofpixelsthatbelongtotheclassc.Inthiswayallthe
classesarebalanced,becausethenetworkweighseachclass
accordingtotherespectiveweight.If,forexample,thereare
manypixelsbelongingtooneclassoverthewholedataset,its
weightwillbelow;vice-versa,iftherearefewpixelsbelonging
tooneclass,itsweightwillbehighaccordingtoEq.(2).So,the
networkwillpaymoreattentiontolearnaclassrepresentedby
fewpixelsratherthanaclassbymanypixels.
Sincetheburnimagedatabasehasimagesrepresentingall
orsomeofthe four burndepths,thesegmentation stepis
appliedonlytotheimagesthatrepresentburnswithallthe
burn depths (there are 17 in total)in order to enable the
convolutionalneuralnetworktolearnfromahomogeneous
dataset.Asimplifieddiagramofthesemanticsegmentationis
describedinFig.2.
Theaccuracy(Acc),F1coefficient,intersectionoverunion
(IoU), precision (P) and sensitivity (S) are calculated for
measuring the performance of the segmentation obtained
from the second convolutional neural network using the
groundtruth.Thesemetricsarecalculatedas:
Acc¼ TPþTPTNþþTNFPþFN (3) F1¼ 2TP 2TPþFPþFN (4) IoU¼ TP TPþFPþFN (5) P¼ TPTPþFP (6) S¼ TPTPþFN (7)
whereTP,TN,FPandFNrepresenttruepositive,truenegative,
falsepositive,andfalsenegative,respectively.Thesevalues
arecalculated onthe binaryclassimages,so,forexample
there is a TP when both the ground-truth and model’s
prediction segmentationhavevalue 1forthesamepixels.
Fig.3illustratesthespaceofthedefinedmetricsforanimage
segmentation.
ThealgorithmwaswritteninPython3.6,usingtheKeras
library[27]functionsonasuper-computerwith512GBRAM,2
Intel(R)Xeon(R)CPUE5-2697v4@2.30GHz,18coreseach,and
3NvidiaGTX10808GB.
Fig.1–Originalburnwoundimage(a)anditsburndepthareasground-truth(b)drewbyaclinicianspecialist:whitefordeep
partialandfull-thicknessdepth;silverforintermediatetodeeppartial-thickness;greyforsuperficialtointermediate
partial-thickness;darkgreyforsuperficialpartial-thickness;and,intheend,blackforuninjuredskinandthebackground.
ThisstudywasapprovedbytheRegionalEthicsCommittee
inLinköpingandconductedincompliancewiththe“Ethical
principlesformedicalresearchinvolvinghumansubjects”of
theHelsinkiDeclaration.
2.3. Trainingofthealgorithm
Beforestartingthe training process, sincetherewere only
17images availablewithallthe burndepthspresent,data
augmentationisstronglyneeded.Inordertoevaluatethe
convolutional neural network, leave-one-out
cross-valida-tioniscomputed,so16imagesareusedforthetrainingand
validation set and just 1 for the testing set. On these
16 original images, rotations of0, 90, 180 and 270 are
applied and for each of these rotated images other new
40 images are created using the elastic deformation
technique[28].Intheend,3936imagesareaugmentedfrom
16 original ones and then split 9010%into training and
validation set respectively for the second convolutional
neuralnetworktrainingprocess.
FromTable1,itispossibletonoticethatthebestnetwork,
theone withthehighestdice coefficient,isthe number3.
Minimizing the dice-loss, accuracy and dice coefficient
convergeatalmostthesamevalueand,afterleave-one-out
cross-validation,the systemhasaverageaccuracyanddice
coefficientof96.81%.Moreover,theaverageweightsforeach
classtobalancethetrainingprocess,calculatedusingEq.(2)
afterimageaugmentationoneachleave-one-outfold,are:
w¼
w0ðBackgroundÞ
w1ðSuperficialIÞ
w2ðSuperficialPartialthicknessIIÞ
w3ðDeepPartialThicknessIIIÞ
w4ðFullThicknessIVÞ
;
wherew0 isthe weight which belongs to the background,
whereastheotherstotheburn-depthclassesI,II,IIIandIV
respectively(seeEq.(2)).Aswanted,thebackgroundweight
hasasmallvalueand,ontheotherhand,thefulland
deep-thicknessdepthweighthasahighvalue,whereasclassIIand
IIIhavesimilarweights,soprobablytheclassificationbetween
themmightbecomplicated.
Fig.4,herebelow,showsfourdifferentsemantic
segmen-tationresults,usingthenetworks3,10,12and16ofTable1.on
theirrespectivetestimages.Eachimageillustratestheburn
wound without the background, its ground-truth and the
convolutional neural network’s prediction. Moreover, it
reportstheaccuracy,F1coefficient,intersectionoverunion,
precisionandsensitivitymetricsextractedfromthe
ground-truthandtheconvolutionalneuralnetwork’spredictionfor
eachclass(seeEqs.(3)(7)).
ItispossibletoconcludethatFig.4illustratesfourgood
semanticsegmentationresultsbecausethemetricsreported
havereallyhighvalues.InTable2arereportedtheaverageof
thesamemetricsoverallthe17burn-woundimagesforeach
class,anditispossibletonoticethatclassIIandclassIIIarethe
oneswithlowermetricsvalues.Thiswasexpectedsinceit
happenedalsoin[3]andalsobecauseburnexpertclinicians
havemoredifficultiestodistinguishthoseclasses.
Neverthe-less, they have high accuracy and suitable F1 coefficient,
precision and sensitivity to help the burn clinicians and
surgeonstoachieveabetterdiagnosis.Therearenoproblems
todistinguishclassIandclassIVsincetheirmetricsvalues
have F1coefficient of93.46%and 86.77%,intersection over
unionof88.68%and78.53%,precisionof93.35%and83.96%,
sensitivityof93.86%and92.80%respectively.
Afterhavingtrainedthealgorithmonthese17imagesthe
remaining83wereexamined.
3.
Results
Sincewedidnothaveaccesstootherthan 83burn-wound
images which unfortunately did not containall the
burn-depth,the17convolutionalneuralnetworkscreatedduring
the leave-one-out cross-validation needed to be used to
evaluate thefinalsetofimages(n=83). Ifaconvolutional
Table1–Accuracyanddicecoefficientvaluesobtained afterleave-one-outcross-validation.
Network Accuracy Dicecoefficient
1 0.8814 0.8812 2 0.8042 0.8040 3 0.9977 0.9977 4 0.9968 0.9967 5 0.9972 0.9972 6 0.9930 0.9930 7 0.9568 0.9567 8 0.9937 0.9936 9 0.9906 0.9906 10 0.9911 0.9911 11 0.9976 0.9976 12 0.9852 0.9852 13 0.9865 0.9864 14 0.9867 0.9867 15 0.9840 0.9480 16 0.9898 0.9898 17 0.9619 0.9617 Average 0.96810.0498 0.96810.0498
Fig.3–Illustrationoftruepositive(TP),truenegative(TN),false
positive(FP)andfalsenegative(FN)betweenabinary
ground-truthanditsprediction.
4
burns xxx (2021) xxx xxxFig.4–Semanticsegmentationresultsusingthenetwork3,10,12and16ontherelativeimages.Eachsegmentationresults
showtheburnwoundimage,theground-truthandtheconvolutionalneuralnetwork’sprediction.Moreover,accuracy(Acc),
neuralnetworklearnthowtodistinguishfourburn-depthsin
animage,itshouldbeabletodothatalsoinoneimagethat
doesnotpresentallofthem.Accuracyanddicecoefficientare
reportedinTable3foreachnetwork.FromTable3itispossible
to notice that all the networksreport accuraciesand dice
coefficientsabovethe90%,withthe4-ththebestonewith
approximately93%forbothofthem.
4.
Discussion
In this paper we used amodified U-Net with residuals to
segmentfourdifferentburn-depths(superficial
partial-thick-ness (I), superficial to intermediate partial-thickness (II),
intermediatetodeeppartial-thickness(III) anddeeppartial
and full-thickness (IV)) in images generated by a
high-performancelightcamerawithpolarisationfilters withthe
aimtotrainthenetworktopredictburndepth.Afteracquiring
100burnimages,seventeenimageswereusedfortraining.
Leave-one-outcross-validationreportsweregeneratedandan
accuracyanddicecoefficientaverageofalmost97%wasthen
obtained.Afterthat, the remaining83 burn-woundimages
wereevaluatedusingthedifferentnetworkduringthe
cross-validation,achievinganaccuracyanddicecoefficient,bothon
average92%.TheF1score, ordicescorecoefficient,isthat
metrictypicallyusedtoevaluateimagesegmentationresults
becauseitdoesnotconsiderinitsequation(seeEqs.(1)and(4))
the true negatives, whereas it focuses more on the true
positivesandwherethepredictioninthisclinicalsettingmost
oftenfails(falsenegativeandfalsepositives).Inotherwords,it
measureshowgoodapredictedsegmentationbythenetwork
overlaps with the “true” segmentation provided by the
clinicianspecialistinthisstudymadeatday20afterburn.
4.1. Relatedworks
Burn woundsassessmentmadebycomputervision
techni-quesareyetnotsopopularbuttherearesomescientistswho
tried to investigate this field. Pinero et al. [6] identified
16 texture features for the burn image segmentation and
classification. These features were then inspected by the
sequential forward and backward selection methods via
fuzzy-ARTMAP neural network. This method achieved an
averageaccuracyofabout83%using250images,4949pixels,
dividedin5burnappearanceclasses:blisters,brightred,
pink-white, yellow-beige, and brown. Wantanajittikul et al. [29]
used the Hilbert transform and texture analysisto extract
feature vectorsand thenappliedasupportvectormachine
(SVM)classifiertoclassifyburndepth.Thebestaccuracyresult
fora4-foldcross-validation was90%using5imagesasthe
validation set and 34 images as the training set,and 75%
correctclassificationonablindtestwasthenobtained.Acha
etal.[30]appliedmultidimensionalscaling(MDS)analysisand
k-nearest neighbour classifier for burn-depth assessment.
Using 20 images as a training set and 74 for testing, 66%
accuracywasobtainedforclassifyingburnwoundsintothree
depths,and84%accuracywasobtainedforthosethatneeded
ordidnotneedgrafts.Serranoetal.[7]usedastrictselectionof
texturefeaturesofburnwoundsfortheMDSanalysisandSVM
andobtained80%accuracyinclassifyingthosethatneeded
graftsandthosethatdidnot.Chauhanetal.[31]usedAIto
classifybodypartsfrom109burn-woundimages(30portray
burnwoundsontheface,35onthehand,23onthebackand21
on theinner forearm)withsize350450 300400pixels,
achievingoverallclassificationaccuracyof91%and93%using
a dependent and an independent convolutional neural
networkResNet-50respectively.Weourselves[3],alsotried
AI,similarlyfortheburn-depthclassification.Wecollected676
samplesofsize224224pixelsfrom23burn-woundimages
(almost100samplesforeachclass:thefourburn-depthsplus
thenormalhealthyskinandthebackground)andachievedan
average,minimum,andmaximumaccuracyof82,72,and88%
respectivelyusingtheResNet-101after10-fold
cross-valida-tion.Moreover,theaverageaccuracy,sensitivity,and
speci-ficitywereextractedforthefourburn-depths:91,74,and94%,
respectively.
5.
Study
limitations
Constructingatrainingdataset,largevolumesofstudyimages
areneeded.Giventhefrequencyofscalds,thecollectionof
very large image databases for training purposes are not
feasibleandthereforethedatasetusedinthisstudymaybe
Table2–Averageofaccuracy(Acc),F1coefficient(F1),
intersectionoverunion(IoU),precision(P)andsensitivity (S)overallthe17burnwoundimagesforeachburn-depth afterleave-one-outcross-validation.
Class Acc F1 IoU P S
I 0.9925 0.9346 0.8868 0.9335 0.9389
II 0.9867 0.7890 0.6907 0.8423 0.7800
III 0.9763 0.7287 0.6177 0.7501 0.7464
IV 0.9806 0.8677 0.7853 0.8396 0.9280
Table3–Accuracyanddicecoefficientvaluesforthe remaining83burnwoundimageswhichdonotshowall theburndepths,butsomeofthem.
Network Accuracy Dicecoefficient
1 0.9202 0.9202 2 0.9160 0.9156 3 0.9173 0.9173 4 0.9306 0.9305 5 0.9218 0.9218 6 0.9070 0.9070 7 0.9257 0.9257 8 0.9079 0.9070 9 0.9143 0.9142 10 0.9191 0.9191 11 0.9207 0.9207 12 0.9207 0.9207 13 0.9251 0.9251 14 0.9147 0.9147 15 0.9239 0.9239 16 0.9218 0.9218 17 0.9152 0.9150 Average 0.91890.0059 1.91880.0060
6
burns xxx (2021) xxx xxxclaimedtoosmall.Thisalbeitthefactthatalmosttwoyears'
collectionofpatients weremade. To improvethis point, a
specificimageoptimizationtechniquewasused(theelastic
deformationtechnique [28]). Bythis measure the16 initial
trainingimageswereartificiallyexpandedto3936imagesand
thusimprovingthepredictionmetrics.Havingmoreimages
forthetrainingsetisimportantforthefurtherimprovementof
thetechnique.
Anotherstudylimitationisofcoursewhatisclaimed“the
final”healingresult,andespecially determiningthe dayof
totalre-epithelializationusedtotrainthepredictionmethod.
Inthisstudyweawaitedthehealingsituation atday20 to
reducetheriskofasubjectiveeffectontheoutcomepresented.
However,thisneedstobeaddressedfurtherincomingstudies.
6.
Conclusion
Inthispaper,wewantedtoextendtheambitionbeyondour
previouspublication[3],addingthelocalclassificationtothe
globalone.Asshownintheprevioussection,AIisapowerful
tool that can be used to for the burn-depths assessment,
achievingaglobaldicecoefficientof97%afterleave-one-out
cross-validation,andtheaverageoftheF1coefficientsoverall
the17testimagesof93%,79%,73%and87%forsuperficial
partial-thickness, superficial to intermediate
partial-thick-ness,intermediatetodeeppartial-thickness,anddeeppartial
and full-thickness burns respectively. These values are
suitableforabetterburndiagnosissincetheexpertclinicians
onburnsassessaburnwoundwith75%accuracycomparedto
the92%presentedinthispaper.Importantlyitthenneedstobe
stressedthatthepresentpaperisbasedonlightphotography
imagesratherthanlaserDopplerbasedimages.Nevertheless,
theconvolutionalneuralnetworkperformanceanditsmetrics
maysurelyincreasewiththeavailabilityoflargerburnimage
databases.Thisobstaclemightbeovertakenwiththeuseof
Generative Adversarial Nets (GANs) [3234] for the image
augmentationonthetrainingimages.Suchfuture
improve-mentsappearespeciallyinterestinggiventheaccuracyand
practicalsimplicityofthemethodpresented.
Declarations
of
conflicts
of
interest
None.
REFERENCES
[1]HettiaratchyS,PapiniR.ABCofburns:initialmanagementofa majorburn:II—assessmentandresuscitation.BMJ:BrMedJ 2004;329:101.
[2]CirilloMD,MirdellR,SjöbergF,PhamTD.Tensor decompositionforcolourimagesegmentationofburn wounds.SciRep2019;9(1):3291.
[3]CirilloMD,MirdellR,SjöbergF,PhamTD.Time-independent predictionofburndepthusingdeepconvolutionalneural networks.JBurnCareRes2019;40(6):85763,doi:http://dx.doi. org/10.1093/jbcr/irz103.
[4]JeschkeMG.Burncareandtreatment:apracticalguide. Springer;2013.
[5]JohnsonRM,RichardR.Partial-thicknessburns:identification andmanagement.AdvSkinWoundCare2003;16(4):17887.
[6]PineroBA,SerranoC,AchaJI,RoaLM.Segmentationand classificationofburnimagesbycolorandtextureinformation. JBiomedOpt2005;10(3):034014.
[7]SerranoC,Boloix-TortosaR,Gómez-CíaT,AchaB.Features identificationforautomaticburnclassification.Burns2015;41 (8):188390.
[8]WearnC,LeeKC,HardwickeJ,AllouniA,BamfordA, NightingaleP,etal.Prospectivecomparativeevaluation studyofLaserDopplerImagingandthermalimagingin theassessmentofburndepth.Burns2018;44(1):12433.
[9]ShinJY,YiHS.DiagnosticaccuracyoflaserDopplerimagingin burndepthassessment:systematicreviewandmeta-analysis. Burns2016;42(7):136976.
[10]MirdellR,FarneboS,SjöbergF,TesselaarE.Accuracyoflaser specklecontrastimagingintheassessmentofpediatricscald wounds.Burns2018;44(1):908.
[11]MirdellR,IredahlF,SjöbergF,FarneboS,TesselaarE. Microvascularbloodflowinscaldsinchildrenanditsrelation todurationofwoundhealing:astudyusinglaserspeckle contrastimaging.Burns2016;42(3):64854.
[12]LindahlF,TesselaarE,SjöbergF.Assessingpaediatricscald injuriesusinglaserspecklecontrastimaging.Burns2013;39 (4):6626.
[13]JaspersME,vanHaasterechtL,vanZuijlenPP,MokkinkLB.A systematicreviewonthequalityofmeasurementtechniques fortheassessmentofburnwounddepthorhealingpotential. Burns2019;45(2):26181.
[14]LongJ,ShelhamerE,DarrellT.Fullyconvolutionalnetworks forsematicsegmentation.ConferenceonComputerVision andPatternRecognitionProceedings2015.
[15]BadrinarayananV,KendallA,CipollaR.SegNet:adeep convolutionalencoder-decoderarchitectureforimage segmentation.IEEETransPatternAnalMachIntell 2017;39:248195.
[16]RonnebergerO,FischerP,BroxT.U-net:convolutional networksforbiomedicalimagesegmentation.International ConferenceonMedicalImageComputingand Computer-AssistedIntervention2015.
[17]TopolEJ.High-performancemedicine:theconvergenceof humanandartificialintelligence.NatMed2019;25(1):44.
[18]MilletariF,NavabN,AhmadiS-A.V-net:fullyconvolutional neuralnetworksforvolumetricmedicalimage
segmentation.2016FourthInternationalConferenceon3D Vision(3DV)2016.
[19]JinQ,MengZ,PhamTD,ChenQ,WeiL,SuR.DUNet:a deformablenetworkforretinalvesselsegmentation. KnowledgeBasedSyst2019;178:14962.
[20]LiX,ChenH,QiXaDQ,FuC-W,HengP-A.H-DenseUNet: hybriddenselyconnectedUNetforliverandtumor segmentationfromCTvolumes.IEEETransMedImaging 2018;37(12):266374.
[21]OktayO,SchlemperJ,FolgocLL,LeeM,HeinrichM,MisawaK, etal.AttentionU-Net:learningwheretolookforthepancreas. arXiv2018preprintarXiv:1804.03999.
[22]IsenseeF,KickingerederPaWW,BendszusM,Maier-HeinKH. NoNew-Net.InternationalMICCAIBrainlesionWorkshop. Springer;2018.p.23444.
[23]MirdellR,FarneboS,SjöbergF,TesselaarE.Interobserver reliabilityoflaserspecklecontrastimagingintheassessment ofburns.Burns2019;45(6):132535.
[24]ElmasryM,MirdellR,TesselaarE,FarneboS,SjöbergF, SteinvallI.Laserspecklecontrastimaginginchildren withscalds:itsinfluenceontimingofintervention, durationofhealingandcare,andcosts.Burns2019;45 (4):798804.
GrossmannB,NilssonA,SjöbergF,NilssonL.Rectal ketamineduringpaediatricburnwounddressing
[25]procedures:arandomiseddose-findingstudy.Burns2019;45 (5):10818.
[26]SudreCH,LiW,VercauterenT,OurselinS,CardosoMJ. Generaliseddiceoverlapasadeeplearninglossfunctionfor highlyunbalancedsegmentations.Deeplearninginmedical imageanalysisandmultimodallearningforclinicaldecision support.Springer;2017.p.2408.
[27]CholletF.Keras. [Online].Available:..https://keras.io.
[28]SimardPY,SteinkrausD,PlattJC.Bestpracticesfor convolutionalneuralnetworksappliedtovisualdocument analysis.7thInternationalConferenceonDocumentAnalysis andRecognition(ICDAR)2003.
[29]WantanajittikulK,AuephanwiriyakulS,Theera-UmponN, KoanantakoolT.Automaticsegmentationanddegree identificationinburncolorimages.The4th2011Biomedical EngineeringInternationalConference2012.
[30]AchaB,SerranoC,FondónI,Gómez-CíaT.Burndepth analysisusingmultidimensionalscalingappliedto psychophysicalexperimentdata.IEEETransMedImaging 2013;32(6):111120.
[31]ChauhanJ,GoswamiR,GoyalP.Usingdeeplearningtoclassify burntbodypartsimagesforbetterburnsdiagnosis.Sipaim— Miccaibiomedicalworkshop.Springer;2018.p.2532.
[32]GoodfellowI,Pouget-AbadieJ,MirzaM,XuB,Warde-FarleyD, OzairS,etal.Generativeadversarialnets.Advancesinneural informationprocessingsystems..p.267280.
[33]YiX,WaliaE,BabynP.Generativeadversarialnetworkin medicalimaging:areview.MedImageAnal2019;101552.
[34]Frid-AdarM,DiamantI,KlangE,AmitaiM,GoldbergerJ, GreenspanH.GAN-basedsyntheticmedicalimage
augmentationforincreasedCNNperformanceinliverlesion classification.Neurocomputing2018;321:32131.