Segmen a ion and classi ica ion o bu n images by colo
and ex u e in o ma ion
Begon
˜a Acha
Ca men Se ano
Jose
´I. Acha
A
´ ea de Teo ı
´adelaSen
˜al y Comunicaciones
Escuela Te
´cnica Supe io de Ingenie os
Uni e si y o Se ille
Camino de los Descub imien os s/n
41092 Se illa, Spain
Lau a M. Roa
G upo de Ingenie ı
´a Biome
´dica
Escuela Te
´cnica Supe io de Ingenie os
Uni e si y o Se ille
Camino de los Descub imien os s/n
41092 Se illa, Spain
Abs ac . In his pape , a bu n colo image segmen a ion and classi-
ica ion sys em is p oposed. The aim o he sys em is o sepa a e bu n
wounds om heal hy skin, and o dis inguish among he di e en
ypes o bu ns (bu n dep hs). Digi al colo pho og aphs a e used as
inpu s o he sys em. The sys em is based on colo and ex u e in o -
ma ion, since hese a e he cha ac e is ics obse ed by physicians in
o de o o m a diagnosis. A pe cep ually uni o m colo space
(
L
*
u
*
*)was used, since Euclidean dis ances calcula ed in his
space co espond o pe cep ual colo di e ences. A e he bu n is
segmen ed, a se o colo and ex u e ea u es is calcula ed ha se es
as he inpu o a Fuzzy-ARTMAP neu al ne wo k. The neu al ne wo k
classi ies bu ns in o h ee ypes o bu n dep hs: supe icial de mal,
deep de mal, and ull hickness. Clinical e ec i eness o he me hod
was demons a ed on 62 clinical bu n wound images, yielding an
a e age classi ica ion success a e o 82%. ©
2005 Socie y o Pho o-Op ical
Ins umen a ion Enginee s.
[DOI: 10.1117/1.1921227]
Keywo ds: colo images; bu n; image segmen a ion; bu n classi ica ion.
Pape 04076 ecei ed May 12, 2004; e ised manusc ip ecei ed Jan. 11, 2005;
accep ed o publica ion Jan. 11, 2005; published online May 11, 2005.
1 In oduc ion
Fo a success ul e olu ion o a bu n inju y i is essen ial o
ini ia e he co ec i s ea men .1To choose an adequa e one,
i is necessa y o know he dep h o he bu n, and a co ec
isual assessmen o bu n dep h highly elies on specialized
de ma ological expe ise. As he cos o main aining a bu n
uni is e y high, i would be desi able o ha e an au oma ic
sys em o gi e a i s assessmen in all he local medical cen-
e s, whe e he e is a lack o specialis s.2,3 The Wo ld Heal h
O ganiza ion demands ha , a leas , he e mus be one bed in
a bu n uni o each 500000 inhabi an s. So, no mally, one
bu n uni co e s a la ge geog aphic ex ension. I a bu n pa-
ien appea s in a medical cen e wi hou bu n uni , a ele-
phone communica ion is es ablished be ween he local medi-
cal cen e and he closes hospi al wi h bu n uni , whe e he
nonexpe doc o desc ibes subjec i ely he colo , shape, and
o he aspec s conside ed impo an o bu n cha ac e iza ion.
The esul in many cases is he applica ion o an inco ec i s
ea men 共 e y impo an o a co ec e olu ion o he
wound兲, o unnecessa y displacemen s o he pa ien , in ol -
ing high sani a y cos and psychological auma o he pa ien
and amily.
Wi h he as ad ances in echnology, compu e aided di-
agnosis 共CAD兲sys ems a e gaining widesp ead accep ance.
Howe e , nowadays, he esea ch in he ield o skin colo
images is de eloping slowly due o he di icul y o ansla -
ing human colo pe cep ion in o objec i e ules, analyzable by
a compu e . Gene ally speaking, one can ind wo main appli-
ca ions abou skin colo image p ocessing in he li e a u e:4
he assessmen o he healing o skin wounds o ulce s,5–9 and
he diagnosis o pigmen ed skin lesions such as
melanomas.10–15 The analysis o lesions in ol es mo e adi-
ional image p ocessing echniques such as edge de ec ion and
objec iden i ica ion, as well as an analysis o he colo , i -
egula i y, and shape o he segmen ed lesion. In wound
analysis, he analysis o he colo s wi hin he wound si e is
o en mo e impo an han he de ec ion o he wound bo de
o he calcula ion o i s a ea. Pa icula ly, in he case o bu n
dep h de e mina ion, ocusing on he shape o he bu n is
i ele an o p edic ing i s dep h. The main cha ac e is ics o
his pu pose a e colo and ex u e in o ma ion, as hey a e he
ea u es obse ed by physicians in o de o gi e a diagnosis.
Au oma ic bu n wound diagnosis is s ill a la gely unex-
plo ed ield. In he ela ed bibliog aphy, one can ind ha
he e is a endency o in es iga e objec i e me hods o de e -
mining he dep h o he bu n in o de o educe he subjec i -
i y and he high expe ience equi emen ha isual inspec ion
demands. Some esea ch in o he ela ionship be ween dep h
and supe icial empe a u e16 has been de eloped. The e a e
also o he wo ks ying o e alua e bu n dep h by using he -
mog aphic images,17 in a ed and ul a iole images,18 adio-
ac i e iso opes19 and lase Dopple lux measu emen s20
On he o he hand, he e is ha dly bibliog aphy abou bu n
dep h de e mina ion by isual image analysis and p ocessing.
Al hough some esea ch g oups apply segmen a ion algo-
i hms o bu n images,5,7,8,21,22 hey y o gi e an assessmen
o he healing o he bu n, so hey ocused on calcula ing
di e ences among se e al aspec s such as a ea, shape, and
appea ance in o de o gi e a p edic ion o he healing e olu-
Add ess all co espondence o Begon
˜a Acha. Tel: +34-954487333; E-mail:
[email p o ec ed] 1083-3668/2005/$22.00 © 2005 SPIE
Jou nal o Biomedical Op ics 10(3), 034014 (May/June 2005)
034014-1Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
Fig. 1 Di e en appea ances ha could p esen a bu n: (a) supe icial de mal (blis e s), (b) supe icial de mal ( ed), (c) deep de mal, (d) ull hickness
(beige), (e) ull hickness (b own).
Fig. 4 Examples o he di e en 49⫻49 bu n images used o ain he classi ie : (a) supe icial de mal (blis e s), (b) supe icial de mal ( ed), (c) deep
de mal, (d) ull hickness (beige), (e) ull hickness (b own).
Acha e al.: Segmen a ion and classi ica ion...
034014-2Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
ion o he wound. To ou knowledge, only he g oup o A o-
mowi z e al.21,22 ies o gi e a diagnosis o he bu n dep h.
F om his assessmen , hey es ima e he numbe o days ha
he wound will ake o heal. They measu e he op ic e lec i -
i y in he ed, g een, and in a ed bands, hypo hesizing ha i
is highly co ela ed wi h bu n healing ime, and hey o m a
alse colo image ha indica es he ime o healing, o equi a-
len ly, he dep h o he bu n. The main disad an age o he
me hod is he complexi y and cos o he image acquisi ion
sys em 共 ideo came a, il e wheel, mo o d i e , e c.兲.
The main con ibu ion o his wo k is he design o a clini-
cally easible sys em o au oma ic bu n wound classi ica ion
based on isual digi al images. Fi s , a p o ocol o he s an-
da diza ion o he bu n image acquisi ion was designed. This
i s s ep was equi ed due o he no el y o he applica ion.
Second, a new segmen a ion algo i hm is p oposed, which has
been p o en e ec i e in segmen ing bu n wound images.
Thi d, once he bu n pa is segmen ed, ep esen a i e colo
and ex u e desc ip o s a e ex ac ed om i . Finally, a neu al
ne wo k classi ie p ocesses hese desc ip o s o gi e an es i-
ma ion o he bu n dep h.
2 Ma e ials and Me hods
2.1
Bu n Cha ac e iza ion
The e a e h ee main ypes o bu n wounds.1共1兲Supe icial
de mal bu n: when he epide mis and pa o he de mis a e
des oyed. The p esence o blis e s 共usually b own colo 兲
and/o a b igh ed colo cha ac e ize i . I is pain ul. 共2兲Deep
de mal bu n: i is cha ac e ized by i s pink-whi ish colo . 共3兲
Full- hickness bu n: all he skin hickness is des oyed and
skin g a s a e needed. A beige-yellow o a da k b own colo
cha ac e izes i . I is no pain ul.
Al hough a bu n wound is classi ied in h ee classes, i can
p esen i e di e en appea ances. 共A兲Blis e s: hey a e su-
pe icial de mal bu ns wi h a b igh ex u e and a ose-b own
colo . 共B兲B igh ed: hey a e supe icial de mal bu ns wi h
b igh ed colo s and we appea ance. 共C兲Pink-whi e: hey a e
deep de mal bu ns wi h a do ed appea ance. 共D兲Yellow-
beige: i s appea ance o ull- hickness bu ns. 共E兲B own:
second appea ance o ull- hickness bu ns. Examples o each
appea ance a e shown in Fig. 1.
2.2
Image Acquisi ion and Calib a ion
The image acquisi ion was ca ied ou by means o a digi al
pho og aphic came a, he Canon EOS 300D 共Canon Inc., To-
kyo, Japan兲. Any nonspecialized pe son should be able o ac-
qui e da a om he pa ien , because i is no possible o ha e
an expe in each cen e . A digi al pho og aphic came a is easy
o u ilize and people a e used o hem.
The p oblems we ound ha had o be sol ed when using a
digi al pho og aphic came a o his applica ion a e explained
in he ollowing subsec ions.
2.2.1
Illumina ion in luence
The mos impo an sou ce o in o ma ion o ou sys em in
o de o classi y bu n dep hs is colo , which is ex emely in-
luenced by he illumina ion. In hospi als he ligh ing condi-
ions can change depending on he oom whe e he pa ien is.
Then, measu ed pixel alues depend on he illuminan s and
wi h mul iple illuminan s he measu ed alues canno be ac-
cu a ely con e ed o a known colo space wi hou some ad-
di ional in o ma ion. The e o e, a s udy abou he in luence o
he di e en sou ces o illumina ion is needed. To pe o m
his s udy, we pho og aphed he Macbe h Colo Checke DC
cha 共G e ag-Macbe h GmbH, Ma ins ied, Ge many兲unde
h ee di e en illumina ions: in a da k oom wi h he buil -in
lash 共guide numbe ⫽13 m a ISO 100兲, in a da k oom wi h
luo escen ligh , and in a oom unde di used sunligh . Unde
hese h ee di e en si ua ions, we ixed he ISO speed o 100,
he s op (A ) o 20 and we a ied he exposu e ime (T ).
We de ine ha he exposu e ime is op imum unde a pa icu-
la illuminan when i is he maximum ime wi hou sa u a ing
any channel. The a io be ween he exposu e imes will gi e
us he in luences o he di e en sou ces o ligh . The op i-
mum exposu e imes we e 1/200, 0.6, and 1.6 s o he lash,
sunligh , and luo escen ligh , espec i ely. Tha means ha
he lash is 320 imes s onge han he luo escen and 120
imes s onge han he sunligh . In o he wo ds, i we choose
T ⫽1/200 and 8 bi s pe colo componen , he luo escen
ligh will no in luence e en he leas signi ican bi and he
sunligh will in luence he wo leas signi ican bi s. In ac ,
we ook a pho og aph unde bo h luo escen and sunligh
illumina ions wi h his pa ame e (T ⫽1/200)and only hese
wo leas signi ican bi s had alues di e en o 0.
We can conclude ha he xenon lash illumina ion is su -
icien ly s ong o domina e illumina ion. Tha is an impo an
esul because in his way we only ha e o calib a e he im-
ages once o each came a, and no o each oom whe e
pa ien s a e ea ed.
2.2.2
Calib a ion
An addi ional p oblem we encoun e ed is ha manu ac u e s
no mally do no publish ei he he ed 共R兲, g een 共G兲, blue 共B兲
p ima ies o he came a o he colo empe a u e o he lash.
The e o e we need o de e mine in some way a ans o ma ion
ma ix o con e om measu ed RGB coo dina es o a
de ice-independen colo ep esen a ion sys em.
Fo his pu pose, we ind he ma ix ans o ma ion be-
ween RGB and CIE 共Commission In e na ionale de
l’Eclai age XYZ 共de ice-independen colo space兲. In he li -
e a u e he e a e many ans o ma ion ma ices om RGB o
XYZ colo space, bu hey a e de ined o speci ic illuminan s
共D65, D50, e c.兲and speci ic RGB p ima ies 共CCIR Rec. 709,
FCC-NTSC, e c兲.23 We ha e de eloped a calib a ion me hod
based on he Macbe h Colo Checke DC cha , which is spe-
ci ically designed o calib a ion o digi al came as. The Mac-
be h Colo Checke DC cha has 240 colo chips and i is
supplied wi h da a gi ing he CIE XYZ ch oma ici y coo di-
na es o each chip unde D50 illuminan . The 240 chips oc-
cupy an a ea o 12 cm⫻20 cm. Ou me hod inds he ans-
o ma ion ma ix om RGB unde unknown illuminan o
XYZ unde D50, and co ec s he nonuni o mi y o he illu-
mina ion as well as he spa ial nonuni o mi y o he came a
sensi i i y. This algo i hm i e a i ely pe o ms he ollowing
s eps:
1. Wi hou co ec ing he illumina ion p o ile and using
only h ee colo pa ches, we calcula e he ini ial ma ix
M1 ha con e s om RGB unde an unknown illumi-
nan o XYZ unde D50.
2. In he i’ h s ep, using he 240 colo pa ches in he cha
Acha e al.: Segmen a ion and classi ica ion...
034014-3Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
and he ma ix Mi⫺1,we calcula e he p o iles,
PR,i(x,y),PG,i(x,y),and PB,i(x,y),so ha , o each
pa ch, he R,G,Bco ec ed wi h he p o iles and mul-
iplied by Mi⫺1a e he X,Y,Z alues speci ied by he
manu ac u e o he colo cha . Tha is, o each pa ch
kin he posi ion (xk,yk) he ollowing equa ion is pe -
o med:
冋
PR,i
PG,i
PB,i
册
⫽
冋
1/R共xk,yk兲
1/G共xk,yk兲
1/B共xk,yk兲
册
共Mi⫺1兲⫺1
冋
Xk
Yk
Zk
册
.共1兲
3. We calcula e he h ee ou h o de su aces, PR,i
⬘(x,y),
PG,i
⬘(x,y),and PB,i
⬘(x,y), ha ma ch bes he p o iles
PR,i(x,y),PG,i(x,y),and PB,i(x,y)calcula ed in s ep
2. P e iously, we ha e expe imen ally de e mined ha a
ou h o de su ace adequa ely app oxima es he sensi-
i i y o he came a and he nonuni o mi y o he lash
illumina ion al oge he .
4. Using his p o ile, we calcula e he ma ix Mi ha bes
maps he R,G,B alues in o he X,Y,Z alues speci ied
o all he pa ches in he colo cha . To de e mine his
op imum Mi he ollowing mean squa e e o is mini-
mized:
␧2⫽1
240 兺
k⫽1
240
共X k⫺Xk兲2⫹共Y k⫺Yk兲2⫹共Z k⫺Zk兲2,
共2兲
whe e X k,Y k,and Z ka e he X,Y, and Z alues o he
k’ h colo pa ch, in he posi ion (xk,yk),speci ied by
he manu ac u e .
5. Repea om s ep 2 un il he mean squa e e o ␧begins
o g ow.
I mus be emphasized ha he ma ix Mis he p oduc o
wo ma ices: he ans o ma ion om RGB o XYZ unde an
unknown illuminan and he linea ans o ma ion o pe o m
he ch oma ic adap a ion om an unknown illuminan o D50.
The ma ix ob ained wi h he p oposed me hod is
M⫽
冋
45 60 ⫺19
24 93 ⫺23
33739
册
when he R,G,B alues a e no malized o one. I should be
no ed ha his ma ix Mis speci ic o each came a, so cali-
b a ion should be pe o med o e e y came a used.
2.2.3
Acquisi ion p o ocol
The hi d p oblem consis s o ixing he acquisi ion p o ocol
so ha he pho og aphs a e use ul o diagnosis. A e ixing i
we ha e alida ed i s sui abili y.
The acquisi ion p o ocol was de eloped by an in e disci-
plina y g oup o med by bu n specialized physicians and
echnicians.24 The main poin s o he acquisi ion p o ocol
we e he ollowing: dis ance be ween came a and pa ien
should be abou 40–50 cm 共 o ix his pa ame e , physicians
ca ied ou a ca e ul analysis o pho og aphs aken o di e en
bu n wounds om di e en dis ances; in he end, hey chose
40–50 cm because hey could dis inguish ex u e om his
dis ance and, a he same ime, hey usually had a global i-
sion o he bu n兲, heal hy skin should appea in he image
when possible, he backg ound should be a g een/blue shee
共 he ones used in hospi als, because as he blue/g een colo is
so di e en om he skin colo s, he backg ound can be easily
ejec ed by he segmen a ion algo i hm兲, he lash mus be on
and he came a should be placed pa allel o he bu n. The
pa ame e s o he came a we e se o: ISO speed 100, expo-
su e ime 1/200 s and ape u e 共 s op兲20.
In o de o alida e he acquisi ion p o ocol, a su ey was
done.24,25 Fo his su ey, 38 pho og aphs o all e iologies,
loca ions, and cha ac e is ics o he mos equen lesions
we e aken ollowing he speci ied p o ocol. They we e p e-
sen ed o a panel o 12 expe s in bu n diagnosis. The expe s
had o answe abou he ce ain y in diagnosis 共1–5兲:
1⫽minimal, 3⫽mode a e, 5⫽maximum, ce ain y. A mean o
4.26 in su eness in diagnosis and 84.6% o diagnos ic accu-
acy was answe ed, whe eas diagnos ic accu acy o a ained
plas ic su geon when looking li e a he same 38 bu n wounds
was 84.3%.
2.3
Bu n Wound Segmen a ion
The segmen a ion app oach used he e is a supe ised pixel-
based algo i hm based on measu es in he CIE L*u* *colo
coo dina e space. L*u* *and L*a*b*colo ep esen a ion
sys ems a e called uni o m sys ems because Euclidean dis-
ances be ween colo s measu ed in hese spaces a e e y
much co ela ed wi h colo di e ences acco ding o human
pe cep ion. They a e pa icula ly use ul in colo image seg-
men a ion o na u al scenes using his og am-based ech-
niques, in which ou me hod is included. They a e sligh ly
di e en because o he di e en app oaches o hei o mula-
ion. Ne e heless, bo h spaces a e equally good in pe cep ual
uni o mi y and p o ide e y good es ima es o colo di e -
ence 共dis ance兲be ween wo colo ec o s.23 The e o e, we
could ha e chosen any o hese wo spaces, bu we p e e ed
he L*u* *one, because he colo componen s a*and b*
do no depend on he luminance, and i is known ha colo
pe cep ion is s ongly in luenced by he luminance.26
The ollowing s eps show he scheme p oposed:
2.3.1
Selec ion o a small egion in he bu n wound
by he use and p ep ocessing o he image
Fo a nonexpe physician 共in ac , o mos o he people兲i is
easy o di e en ia e bu n skin om no mal one. The e o e,
he bu n wound will be segmen ed using he colo in o ma ion
o a5⫻5 pixel a ea a ound he poin ha he use selec s wi h
he mouse.
Be o e segmen ing he image, i is con enien o p ep o-
cess i in o de o ge mo e homogeneous egions elimina ing
noise and small s uc u es. To pe o m his ask, an aniso-
opic di usion is applied o he colo image.27,28 The aim o
he di usion is o make he egions mo e homogeneous bu
p ese ing he edge in o ma ion. In o de o pe o m he an-
iso opic di usion, he app oach o sepa a ing he di usion o
he ch oma ic and ach oma ic in o ma ion was ollowed28 as
is shown in Fig. 2. Fi s , he image is con e ed in o L*u* *
colo coo dina e sys em acco ding o23
Acha e al.: Segmen a ion and classi ica ion...
034014-4Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
L*⫽
再
116
冉
Y
Y0
冊
1/3
⫺16, i Y
Y0
⬎0.008856
903.3
冉
Y
Y0
冊
,o he wise
.共3兲
Compu a ion o u*and *in ol es in e media e u⬘, ⬘,
u0
⬘,and 0
⬘quan i ies de ined as
u⬘⫽4X
X⫹15Y⫹3Z,
共4兲
⬘⫽9Y
X⫹15Y⫹3Z.
Finally,
u*⫽13L*共u⬘⫺u0
⬘兲,
共5兲
*⫽13L*共 ⬘⫺ 0
⬘兲.
Y0,u0,and 0co espond o he whi e e e ence poin ,
which depends on he illuminan 共D50 a e he calib a ion兲.
F om hese coo dina es, he hue and ch oma componen s
a e calcula ed as H⫽a c an( */u*)and C⫽
冑
(u*)2⫹( *)2,
espec i ely. A complex quan i y is calcula ed ha ela es he
hue and he ch oma as P⫽Cexp(jH).
The ach oma ic aniso opic di usion, applied o L*,is
ca ied ou by means o he disc e e o mula ion27 o he pa -
ial di e en ial equa ion
⳵
⳵
L*共x,y, 兲⫽di 关
␣
共x,y, 兲ⵜL*共x,y, 兲兴,共6兲
whe e di and ⵜdeno e he di e gence and he g adien op-
e a o s, espec i ely, and
␣
(x,y, )is a mono onically de-
c easing unc ion o he image g adien magni ude called he
conduc ance coe icien and is gi en by
␣
共x,y, 兲⫽1
1⫹
冉
兩
ⵜL*共x,y, 兲
兩
␥
p
冊
2.共7兲
The di usion cons an
␥
pwas selec ed as he 5% o he
maximum alue o
兩
ⵜL*(x,y, )
兩
a each , an a i icial ime
pa ame e ha deno es he numbe o di usion i e a ions,
which was ixed o 20.
The ch oma ic aniso opic di usion is pe o med by apply-
ing Eq. 共6兲 o he complex quan i y P
⳵
⳵
P共x,y, 兲⫽di 关
␣
共x,y, 兲ⵜP共x,y, 兲兴,共8兲
whe e ⵜP(x,y, )is28
ⵜP共x,y, 兲⫽关ⵜC共x,y, 兲⫹jCⵜH共x,y, 兲兴exp关jH共x,y, 兲兴
共9兲
and sepa a ing eal and imagina y pa s o Eq. 共8兲i ollows
ha
⳵
⳵
C⫽di 共
␣
ⵜC兲⫺
␣
C
兩
ⵜH
兩
2,
共10兲
⳵
⳵
H⫽di 共
␣
ⵜH兲⫹2
␣
CⵜC•ⵜH,
whe e he spa ial and empo al dependencies ha e been omi -
ed o con enience.
To ob ain he coe icien
␣
o he complex quan i y Pwe
need o calcula e
兩
ⵜP(x,y, )
兩
,which is
兩
ⵜP共x,y, 兲
兩
⫽
冑
兩
ⵜC共x,y, 兲
兩
2⫹C2共x,y, 兲
兩
ⵜH共x,y, 兲
兩
2.
共11兲
2.3.2
Con e sion o single channel image
In his s ep a g ay scale image is ob ained om he di used
colo image. In his g ay scale image, di e ences be ween he
bu n skin selec ed by he use and o he pa s o he image
a e emphasized. Based on he obse a ion ha doc o s seg-
men bu n wounds by measu ing di e ences among colo s,
he selec ion box selec ed by he use is slid as a mask o size
5⫻5 pixels along he image and, o each pixel in he image
unde he cen e o he sliding mask, he ollowing ope a ion
is pe o med:29
共n,m兲⫽1
MAX 兺
i⫽n⫺⌬
n⫹⌬
兺
j⫽m⫺⌬
m⫹⌬
dE关p共i,j兲,w共i,j兲兴,
共12兲
whe e MAX is max
n,m(兺i⫽n⫺⌬
n⫹⌬兺j⫽m⫺⌬
m⫹⌬dE„p(i,j),w(i,j)…),⌬
⫽(L⫺1)/2 wi h L⫽5, p(i,j) ep esen s a pixel in he di -
used image o be segmen ed in L*u* *colo space, w(i,j)
is a pixel o he mask selec ed by he use , and dE(•), he
Euclidean dis ance be ween pixels p(i,j)and w(i,j),is de-
ined as
Fig. 2 Di usion il e ing sepa a ing ch oma ic and ach oma ic in o -
ma ion.
Acha e al.: Segmen a ion and classi ica ion...
034014-5Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx

dE共p共i,j兲,w共i,j兲兲⫽
兵
关Lp
*共i,j兲⫺Lw
*共i,j兲兴2⫹关up
*共i,j兲
⫺uw
*共i,j兲兴2⫹关 p
*共i,j兲
⫺ w
*共i,j兲兴2
其
1/2.共13兲
2.3.3
Th esholding ope a ion and pos p ocessing
The esul o he abo e s ep is a g ay-scale image whe e pix-
els wi h lowes alues a e hose in he egion o be segmen ed.
This image has been ca e ully designed o emphasize he
bu n egions, and a h esholding ope a ion should su ice o
ge a good segmen a ion. The his og am o his dis ance im-
age is mul imodal so a me hod o ind a h eshold o selec he
mode in he le o he his og am should be ound. This ask is
ca ied ou in wo s eps: 共1兲 he peaks 共maximum alues兲o
he di e en modes p esen in he his og am a e ound, and
共2兲 he h eshold which sepa a es he wo modes closes o he
le o he his og am is calcula ed applying O su’s hesholding
me hod.30
To pe o m he i s s ep, he ollowing algo i hm is ap-
plied o he his og am o he g ay-scale image: 共1兲 ind all
peaks in he his og am, ha is, all he alues in he his og am
which a e highe han hei wo neighbo s; 共2兲 o m a new
cu e wi h he peaks ound in he p e ious s ep and hen
selec again he peaks in he new cu e; 共3兲 emo e nonsig-
ni ican peaks, i.e., hose peaks whose alues a e less han 1%
o he maximum peak alue a e ejec ed; 共4兲 emo e nonsig-
ni ican alleys, ha is, i wo peaks ha e no a signi ican
alley be ween hem we main ain only he highes o he wo
peaks. To check i a alley is signi ican o no , he minimum
alue be ween wo peaks is ound. I his minimum alue is
g ea e han 75% o he lowes peak ou o he wo peaks, hen
he alley is conside ed nonsigni ican . These ou s eps a e
illus a ed in Fig. 3.
Once we ha e localized he main modes in he his og am,
we ha e o ind he h eshold which sepa a es he wo modes
closes o he le pa o he his og am. This ask is ca ied ou
by applying O su’s me hod,30 which is an adap i e h eshold-
ing echnique o spli a his og am in o wo classes, c1wi h
g ay le els 关1,...,k兴,and c2wi h g ay le els 关k⫹1,...,K兴.Le
mi(k)and mTbe he mean in ensi ies o he class ciand o
he whole image, espec i ely. The be ween-class a iance
was de ined by O su as
␴
b
2共k兲⫽
␻
1共k兲共m1共k兲⫺mT兲2⫹
␻
2共k兲共m2共k兲⫺mT兲2,
共14兲
whe e
␻
1(k)and
␻
2(k)a e cumula i e sums o he p obabili-
ies in each class, ha is,
␻
1(k)⫽兺j⫽1
kpj,
␻
2(k)
⫽兺j⫽k⫹1
Kpj,and pj⫽xj/Npixels ,whe e xjis he numbe o
pixels wi h g ay le el jin an image and Npixels is he numbe
o pixels wi h g ay le els om 1 o Kin he whole image, ha
is, he o al numbe o pixels in he image. The op imal h esh-
old k
ˆis chosen so ha he be ween-class a iance
␴
b
2is maxi-
mized.
The elec ion o O su’s me hod, among many exis ing
h esholding me hods, is due o i s simplici y in
compu a ion.31 In ac , many mode n segmen a ion algo i hms
a e based in O su’s me hod o use i o compa ison.32–34
Finally, by he applica ion o a 3⫻3 median il e , he seg-
men a ion esul is imp o ed by emo ing spu ious poin s
共1–4 pixel sized兲, ha is, poin s ha ha e been segmen ed and
do no ac ually belong o he bu n.
2.4
Classi ica ion
Once he bu n is segmen ed, i s dep h mus be es ima ed o
classi ica ion pu poses. I has been p o en ha physicians de-
e mine he dep h o a bu n based on colo pe cep ion, as well
as on some ex u e aspec s. As i has been p e iously said,
L*u* *space is a pe cep ually uni o m colo ep esen a ion
sys em. Also, he hue and he ch oma coo dina es a e in i-
ma ely ela ed o he way human beings pe cei e ch oma ic-
i y. Tha is why, in his s udy, a se o desc ip o s o med by
s a is ical momen s o he his og ams ob ained o each coo -
dina e o he L*u* *colo space, as well as o he hue and
ch oma image planes de i ed om hem, ha e been used.
Mo e speci ically, he desc ip o s chosen a e: mean o ligh -
ness (L*),mean o hue 共H兲, mean o ch oma 共C兲, s anda d
de ia ion o ligh ness (
␴
L),s anda d de ia ion o hue (
␴
H),
s anda d de ia ion o ch oma (
␴
C),mean o u*,mean o *,
s anda d de ia ion o u*(
␴
u),s anda d de ia ion o *(
␴
),
skewness o ligh ness (sL),ku osis o ligh ness (kL),skew-
ness o u*(su),ku osis o u*(ku),skewness o *(s )and
ku osis o *(k ).
A e wa ds i has been necessa y o apply a desc ip o se-
lec ion me hod o ob ain he op imum se o he subsequen
classi ica ion.
2.4.1
Fea u e selec ion
The disc imina ion powe o hese 16 ea u es is analyzed
using he sequen ial o wa d selec ion 共SFS兲me hod and he
Fig. 3 P ocess o de ec ing he main peaks in he his og am. (a) De-
ec ion o he peaks in he his og ams: peaks a e ma ked wi h ci cles.
(b) Finding he peaks in he his og am o he peaks: peaks om he
o iginal his og am a e ma ked wi h do s and new peaks wi h ci cles.
(c) Rejec ion o nonsigni ican peaks: peaks om Fig. (b) a e ma ked
wi h do s and peaks selec ed in his s ep a e ma ked wi h ci cles. (d)
Final peaks in he o iginal his og am a e he ejec ion o peaks wi h-
ou a signi ican alley be ween hem. In his case he h ee peaks in
he o me s ep a e accep ed.
Acha e al.: Segmen a ion and classi ica ion...
034014-6Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
sequen ial backwa d selec ion 共SBS兲me hod35,36 ia he
Fuzzy-ARTMAP neu al ne wo k which is de ailed in he ol-
lowing subsec ion.
SFS is a bo om-up sea ch p ocedu e whe e one ea u e a
a ime is added o he cu en ea u e se . A each s age, he
ea u e o be included in he ea u e se is selec ed among he
emaining a ailable ea u es which ha e no been added o he
ea u e se . So he new enla ged ea u e se yields a minimum
classi ica ion e o compa ing o adding any single ea u e.
The algo i hm s ops when adding a new ea u e yields an
inc ease o he classi ica ion e o . The SBS is he op-down
coun e pa o he SFS me hod. I s a s om he comple e se
o ea u es and, a each s age, he ea u e which shows he
leas disc imina o y powe is disca ded. The algo i hm s ops
when emo ing ano he ea u e implies an inc ease o he
classi ica ion e o .
To apply hese wo me hods, 50 49⫻49 pixel images o
each bu n appea ance ha e been used 共see Fig. 4兲. As he e
a e i e appea ances, in all we ha e 250 49⫻49 pixel
images.*One pho og aph has been aken pe bu n wound. In
gene al, we selec ed only one 49⫻49 pixel image pe pho o-
g aph, unless he e we e di e en appea ances in he same
wound. In his case, one 49⫻49 image pe appea ance was
selec ed.
The selec ion pe o mance is e alua ed by i e old c oss
alida ion 共XVAL兲.15 In his sense, he disad an age o sensi-
i i y o he o de o p esen a ion o he aining se , ha he
SBS and SFS me hods p esen ,35 is diminished. To pe o m
he XVAL me hod he 50 images pe bu n appea ance a e spli
in o i e disjoin subse s. Fou o hese subse s 共 ha is, 40
images pe appea ance兲se e as a aining se o he neu al
ne wo k, while he o he one 共 en images兲is used as alida-
ion se . Then, he p ocedu e is epea ed in e changing he
alida ion subse wi h one o he aining subse s, and so on
ill he i e subse s ha e been used as alida ion se s. The inal
classi ica ion e o is calcula ed as he mean o he e o s o
each XVAL un.
In Fig. 5 he e olu ion o he classi ica ion e o is p e-
sen ed o bo h selec ion me hods. I can be obse ed ha
bo h cu es coincide a he beginning and a he end, bu hen
hey sepa a e ob aining a minimum classi ica ion e o wi h
se en o eigh desc ip o s 共2% e o 兲 o he SFS me hod, and
six desc ip o s 共1.6% e o 兲 o he SBS me hod. In ac , his
minimum e o is again eached wi h 12 desc ip o s, al hough
i is easonable o choose he se o six, because i will imply
less complexi y in he neu al ne wo k and sho e p ocessing
ime. The six desc ip o s p o ided by SBS me hod we e cho-
sen as he bes ea u e se : ligh ness, hue, s anda d de ia ion
o he hue componen , u*ch ominance componen , s anda d
de ia ion o he *componen , and skewness o ligh ness.
2.4.2
Fuzzy-
ARTMAP
neu al ne wo k
The classi ie used is a Fuzzy-ARTMAP neu al ne wo k. This
ype o ne wo k is based on he Adap i e Resonance Theo y
de eloped by G ossbe g and Ca pen e . Fuzzy-ARTMAP is a
supe ised lea ning classi ica ion a chi ec u e o analog-
alue inpu pai s o pa e ns.37 The easons o his choice a e
ha Fuzzy-ARTMAP o e s he ad an ages o well-unde s ood
heo e ical p ope ies, an e icien implemen a ion, clus e ing
p ope ies ha a e consis en wi h human pe cep ion, and a
e y as con e gence. I has also a ack eco d o success ul
use in indus ial and medical applica ions.38 O he s ong-
poin s o his ype o neu al ne wo k a e he small numbe o
design pa ame e s 共 he igilance pa ame e ,
␳
a苸关0,1兴,and
he selec ion pa ame e ,
␣
⬎0兲, and ha he a chi ec u e and
ini ial alues a e always he same, independen o he appli-
ca ion.
When he inpu pa ame e s a e he ea u es selec ed by he
SBS me hod abo e, he ne wo k classi ies he bu n dep h o
he segmen ed egion in o i e ypes: he i s and he second
belonging o supe icial de mal dep h, he hi d o deep de -
mal, and he ou h and i h o ull hickness. So, he ne wo k
has six neu ons in he inpu laye and i e neu ons in he
ou pu laye . In he Fuzzy-ARTMAP neu al ne wo k he a chi-
ec u e is dynamic, so he numbe o neu ons in he hidden
laye is ixed du ing he aining and acco ding wi h he igi-
lance pa ame e .
3 Expe imen al Resul s
The images used o es he bu n CAD ool we e 62 digi al
pho og aphs aken by physicians ollowing he acquisi ion
p o ocol. All he images we e diagnosed by a g oup o plas ic
su geons, a ilia ed wi h he bu n uni o he Vi gen del Rocı
´o
Hospi al, om Se ille 共Spain兲. The assessmen s we e ali-
da ed one week la e , as is he common p ac ice when han-
dling bu n pa ien s. The images we e 1536⫻1024 pixels and
hey we e s o ed as JPEG 共high quali y兲 iles.
The compu e used was a Pen ium IV, 1.7 GHz and 256
MB o andom access memo y. The a e age un ime was 4
min o an image and he p og amming ool was MATLAB 6.1
共The Ma hwo ks Inc., Na ick, Massachuse s兲.
*The 250 49⫻49 pixel images a e small images showing each one only
one bu n appea ance 共no heal hy skin o backg ound兲. Each 49⫻49 pixel
image has been alida ed by wo physicians as belonging o a pa icula
dep h. The e o e, hese 250 images o m a da abase used only o he
ea u e selec ion s ep.
Fig. 5 E olu ion o he classi ica ion e o o SFS me hod (䊉) and SBS
me hod (䊊).
Acha e al.: Segmen a ion and classi ica ion...
034014-7Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
3.1
Segmen a ion Resul s
The segmen a ion algo i hm p oposed in his pape was es ed
wi h 35 ou o he 62 images o he da abase. These 35 images
we e manually segmen ed by i e physicians.
The eason o using 35 pho og aphs ins ead o 62 is ha ,
al hough he p o ocol says ha i should appea as heal hy and
bu n skin, e y o en he ex ension o he bu n wound is so
la ge ha he e is only bu n skin in he image. The e o e, in
hese cases i is no meaning ul o compa e he segmen a ion
esul s pe o med by he physicians and by he algo i hm.
The segmen a ion gold s anda d was ob ained by applying
he o ing me hod o he egions segmen ed by he i e spe-
cialis s. In o he wo ds, one pixel was conside ed o belong o
he segmen ed egion in he gold s anda d i mos o he phy-
sicians had conside ed i in his way.
Once a gold s anda d was ob ained, wo pa ame e s we e
calcula ed o measu e he pe o mances o he segmen a ion
algo i hm. The i s pa ame e was he posi i e p edic i e
alue 共PPV兲, which measu es he a io be ween he numbe o
pixels segmen ed by he algo i hm which i he segmen a ion
gold s anda d and he o al amoun o pixels segmen ed. The
second pa ame e is called sensi i i y 共S兲, and i is he a io
be ween he numbe o pixels segmen ed by he algo i hm
which i he segmen a ion gold s anda d and he o al amoun
o pixels in he segmen a ion gold s anda d. In ui i ely i can
be seen ha he i s pa ame e measu es he o e segmen a-
ion, which would be null i PPV we e 1. Likewise, Smea-
su es he unde segmen a ion. In Table 1 he esul s o he 35
images a e p esen ed. As is shown in his able, almos all he
pho og aphs a e p ope ly segmen ed. I mus be emphasized
ha , al hough he sensi i i y ends o be only a ound 0.8, his
is because doc o s end o o e segmen he bu n egion.
The e o e, his should no be in e p e ed as a poo pe o -
mance o he algo i hm.
Figu es 6–8 show he segmen a ion esul s o some im-
ages o he h ee ypes o dep h. Figu es 共a兲 ep esen o iginal
images and Figs. 共b兲 ep esen he segmen ed ones. In he
segmen ed images we ha e ma ked wi h yellow colo he seg-
men ed egion. In all he cases, he bu n wound was seg-
men ed co ec ly om he no mal skin.
3.2
Classi ica ion Resul s
To es he classi ica ion pa we employed he 62 images o
he da abase used o alida ion 共di e en om he one used
o aining兲. The neu al ne wo k was ained wi h he 250
49⫻49 pixel images p e iously ci ed. The aining was pe -
o med wi h
␳
a⫽1and
␣
⫽0.001. A he end o he aining
he weigh s we e ixed o he subsequen classi ica ion es .
Fo his es he six ea u es we e ex ac ed om he seg-
men ed pa o he 62 images. Classi ica ion esul s a e sum-
ma ized in Table 2. We ha e used 22 images wi h supe icial
de mal bu ns, 18 wi h deep de mal bu ns, and 22 wi h ull-
hickness bu ns. The a e age success pe cen age was 82.26%.
All supe icial de mal bu ns misclassi ied we e classi ied by
he ne wo k as deep de mal ones. All deep de mal bu ns we e
misclassi ied as supe icial de mal ones. And, in he case o
misclassi ied ull- hickness bu ns, 80% o hem we e classi-
ied as supe icial de mal and 20% as deep de mal.
4 Discussion and Conclusions
The classi ica ion o bu n dep hs based on isual inspec ion is
a di icul ask, which needs a lo o aining. Tha is why in
bu n ela ed li e a u e he e is a cons an sea ch o objec i e
me hods o de e mine he dep h o a bu n. A p o o ype o one
in asi e echnique is he acquisi ion o biopsies and hei his-
ological s udy o he bu n dep h diagnosis.39 This echnique,
al hough i can be conside ed as ‘‘gold s anda d,’’ is no ex-
emp om p oblems ela ed o loss o de mis in he bu n, o
he exis ence o conside able a iabili y depending on whe e
he biopsy was acqui ed, and o he ac ha his echnique is
a snapsho iew o he lesion, apa om he esidual sca s
p o oked by he biopsy acquisi ion. These incon eniences
ha e di ec ed e o s owa ds he design o nonin asi e p o-
cedu es. Some nonin asi e echniques analyze he pe usion
o he bu n wound based on he ac ha issue damage is
in e sely p opo ional o he ascula iza ion a e he
lesion.40–42 Ne e heless, in hese p ocedu es i is necessa y o
supply a i al colo an o he pa ien by in a enous me hod
and i is essen ial o ha e an eme gency sys em. O he expe i-
men al echniques analyze he changes in op ical p ope ies o
he skin ela ed o he changes o i s ascula iza ion,43 al-
hough hei applica ion en i onmen is, o he momen , ex-
clusi ely expe imen al. In ano he ype o app oxima ion o
he p oblem being s udied, he emission-op ical measu emen
exploi s he di e en spec al backsca e ing e ec s o bu ned
Table 1 Quan i ica ion o segmen a ion esul s (PPV: posi i e p edic-
i e alue; S: sensi i i y).
Image PPV S Image PPV S
Image 1 0,9309 0,8093 Image 19 0,8303 0,9280
Image 2 0,9314 0,6969 Image 20 0,9627 0,9005
Image 3 0,9391 0,8684 Image 21 0,9196 0,7418
Image 4 0,9302 0,8324 Image 22 0,8752 0,7789
Image 5 0,9614 0,9015 Image 23 0,9559 0,9725
Image 6 0,9741 0,8853 Image 24 0,8622 0,9107
Image 7 0,8807 0,7297 Image 25 0,9082 0,9069
Image 8 0,8984 0,8108 Image 26 0,9646 0,7989
Image 9 0,9618 0,7772 Image 27 0,9320 0,9364
Image 10 0,9737 0,8206 Image 28 0,8711 0,8457
Image 11 0,7928 0,8190 Image 29 0,9569 0,9482
Image 12 0,9624 0,7452 Image 30 0,9571 0,8814
Image 13 0,9806 0,7248 Image 31 0,9134 0,8318
Image 14 0,9424 0,7820 Image 32 0,9327 0,8698
Image 15 0,9384 0,8457 Image 33 0,6990 0,7588
Image 16 0,8327 0,8066 Image 34 0,9192 0,5174
Image 17 0,6420 0,8539 Image 35 0,7701 0,8530
Image 18 0,8788 0,9646 A e age 0,9023 0,8301
Acha e al.: Segmen a ion and classi ica ion...
034014-8Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx
Fig. 6 Segmen a ion esul o a supe icial de mal bu n. (a) O iginal image whe e he selec ion made by he use is shown wi h an a ow. (b)
Segmen ed image.
Fig. 7 Segmen a ion esul o a deep de mal bu n. (a) O iginal image whe e he selec ion made by he use is shown wi h an a ow. (b) Segmen ed
image.
Fig. 8 Segmen a ion esul o a ull hickness bu n. (a) O iginal image, which has bo h supe icial de mal bu n ( he ed pa ) and ull- hickness bu n
( he c eamy pa ). (b) Segmen ed image. In his case he use has made he selec ion in he c eamy pa in o de ha he algo i hm segmen s all he
ull- hickness pa o he bu n. I segmen s co ec ly all he ull- hickness pa s o he image ega ding wha physicians said.
Acha e al.: Segmen a ion and classi ica ion...
034014-9Jou nal o Biomedical Op ics May/June 2005 䊉Vol. 10(3)
Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/22403/ on 03/29/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx