a
c
h
s
o
c
e
s
p
o
a
l
m
o
l
.
2
0
1
2;8
7(9):284–289
ARCHIVOS
DE
LA
SOCIEDAD
ESPAÑOLA
DE
OFTALMOLOGÍA
www.
e
lse ie .es/o almologia
O iginal
a icle
Au oma ed
de ec ion
o
mic oaneu ysms
by
using
egion
g owing
and
uzzy
a map
neu al
ne wo k夽,夽夽
S.
Jiméneza,b,∗,
P.
Alemanya,b,
F.J.
Nú ˜
nezc,
I.
Fondónc,
C.
Se anoc,
B.
Achac,
I.
Failded
aSe icio
de
O almología,
Hospi al
Uni e si a io
Pue a
del
Ma ,
Cádiz,
Spain
bDepa amen o
de
Ci ugía,
Facul ad
de
Medicina,
Uni e sidad
de
Cádiz,
Cádiz,
Spain
cDepa amen o
de
Teo ía
de
la
Se˜
nal
y
Comunicaciones,
Escuela
Técnica
Supe io
de
Ingenie ía,
Uni e sidad
de
Se illa,
Se illa,
Spain
dDepa amen o
de
Biomedicina,
Bio ecnología
y
Salud
Pública,
Facul ad
de
Medicina
de
Fisio e apia,
Uni e sidad
de
Cádiz,
Cádiz,
Spain
a
i
c
l
e
i
n
o
A icle
his o y:
Recei ed
28
Janua y
2012
Accep ed
8
Ap il
2012
A ailable
online
24
Oc obe
2012
Keywo ds:
Compu e
aided
diagnosis
Fundus
pho og aphy
Diabe ic
e inopa hy
sc eening
Mic oaneu ysm
Neu al
ne wo k
a
b
s
a
c
Objec i e:
To
assess
whe he
he
me hodological
changes
o
his
new
algo i hm
imp o es
he
esul s
o
a
p e iously
p esen ed
s a egy.
Me hods:
We
enhance
he
image
and
fil e
ou
he
g een
channel
o
he
digi al
colo
e inog-
aphy.
Mul i ole ance
h esholding
was
applied
o
ob ain
candida e
poin s
and
make
a
seed
g owing
egion
by
a ying
in ensi ies.
We
ook
15
cha ac e is ics
om
each
egion
o
ain
a
uzzy
A map
neu al
ne wo k
using
42
e inal
pho og aphs.
This
ne wo k
was
hen
applied
in
he
s udy
o
11
good
quali y
e inal
pho og aphs
included
in
he
diabe ic
e inopa hy
ea ly
de ec ion
sc eening
p og am,
wi h
ini ial
s ages
o
e inopa hy,
ob ained
wi h
he
Topcon
NW200
non-myd ia ic
e inal
came a.
Resul s:
Two
expe ienced
oph halmologis s
de ec ed
52
mic oaneu ysms
in
11
images.
The
algo i hm
de ec ed
39
mic oaneu ysms
and
3752
mo e
egions,
confi ming
38
mic oa-
neu ysm
and
135
alse
posi i es.
The
sensi i i y
is
imp o ed
compa ed
o
he
p e ious
algo i hm,
om
60.53%
o
73.08%.
False
posi i es
ha e
d opped
om
41.8
o
12.27
pe
image.
Conclusions:
The
new
algo i hm
is
be e
han
he
p e ious
one,
bu
he e
is
s ill
oom
o
imp o emen ,
especially
in
he
ini ial
de e mina ion
o
seeds.
©
2012
Sociedad
Española
de
O almología.
Published
by
Else ie
España,
S.L.
All
igh s
ese ed.
De ección
au oma izada
de
mic oaneu ismas
median e
c ecimien o
de
egiones
y
ed
neu onal
Fuzzy
A map
Palab as
cla e:
Diagnós ico
asis ido
po
o denado
Re inog a ía
De ección
de
e inopa ía
diabé ica
e
s
u
m
e
n
Obje i o:
Comp oba
si
las
modificaciones
me odológicas
de
es e
nue o
algo i mo
mejo an
el
esul ado
de
o a
es a egia
p esen ada
an e io men e.
Mé odos:
Se
ealza
y
fil a
la
imagen
negada
del
canal
e de
de
la
e inog a ía
digi al
en
colo .
Se
aplica
una
umb alización
mul i ole ancia
pa a
ob ene
pun os
candida os
夽Please
ci e
his
a icle
as:
Jiménez
S,
e
al.
De ección
au oma izada
de
mic oaneu ismas
median e
c ecimien o
de
egiones
y
ed
neu onal
Fuzzy
A map.
A ch
Soc
Esp
O almol.
2012;87:284–9.
夽夽 This
pape
was
p esen ed
a
he
esea ch
session
o
he
86
SEO
Cong ess
held
in
Mad id
on
Sep embe
22,
2010.
∗Co esponding
au ho .
E-mail
add ess:
[email protected]
(S.
Jiménez).
2173-5794/$
–
see
on
ma e
©
2012
Sociedad
Española
de
O almología.
Published
by
Else ie
España,
S.L.
All
igh s
ese ed.
a
c
h
s
o
c
e
s
p
o
a
l
m
o
l
.
2
0
1
2;8
7(9):284–289
285
Mic oaneu ismas
Red
neu onal
y
en
cada
semilla
se
ealiza
un
c ecimien o
de
egiones
po
a iación
de
in ensidades.
Se
oman
15
ca ac e ís icas
de
cada
egión
y
en enamos
una
ed
neu onal
Fuzzy
A map
con
42
e inog a ías.
Se
aplica
la
ed
en
el
es udio
de
11
e inog a ías
del
p og ama
de
de ección
p ecoz
de
e inopa ía
diabé ica,
de
buena
calidad,
con
lesiones
iniciales,
ob enidas
con
el
e inóg a o
no
mid iá ico
Topcon
NW200.
Resul ados:
Dos
o almólogos
expe imen ados
de ec an
52
mic oaneu ismas
en
las
11
imágenes.
El
algo i mo
de ec a
39
mic oaneu ismas
y
3.752
egiones
más,
confi mando
38
mic oaneu ismas
y
135
alsos
posi i os.
La
sensibilidad
ha
mejo ado
espec o
al
algo-
i mo
an e io
del
60,53
al
73,08%.
Los
alsos
posi i os
has
disminuido
de
41,8
po
imagen
a
12,27.
Conclusiones:
El
nue o
algo i mo
p esen a
indudables
mejo as
espec o
al
an e io ,
pe o
aún
se
puede
pe ecciona ,
sob e
odo
en
la
de e minación
inicial
de
semillas.
©
2012
Sociedad
Española
de
O almología.
Publicado
po
Else ie
España,
S.L.
Todos
los
de echos
ese ados.
In oduc ion
Diabe ic
e inopa hy
is
he
mos
equen
cause
o
new
adul
blindness
cases
o
age
g oups
be ween
20
and
74
yea s.1Diabe ic
e inopa hy
de ec ion
p og ams
ocus
on
iden i ying
inju ies
ha
poin
o
he
disease
(among
o h-
e s,
mic oaneu ysms)
in
colo
digi al
pho og aphs
aken
wi h
non-mid ia ic
e inog aphs
in
he
a ge
popula ion.2The
size
and
colo
o
mic oaneu ysms
make
hem
easily
iden ifiable
by
p ima y
ca e
physicians
in ol ed
in
de ec ion
p og ams.
In
addi ion,
he
localiza ion
p ocess
is
edious
and
slow,
e en
o
specialis s.
Fo
hese
easons
i
is
necessa y
o
de ise
an
au oma ed
sys em
o
quickly
and
e ficien ly
de ec
mic oaneu ysms.
To
his
end,
algo i hms
we e
de eloped
o
enhance
mic oaneu ysms
candida e
inju ies,
elimina e
o he
ocula
undus
elemen s
and
es ablish
di e en
cha -
ac e is ics
o
classi y
candida es
as
ue
mic oaneu ysms.3–6
This
pape
p esen s
an
imp o ed
algo i hm
o
de ec ing
mic oaneu ysms
in
colo
digi al
pho og aphs
o
pa ien s
wi h
diabe ic
e inopa hy
and
he
diagnos ic
use ulness
he eo .
Ma e ials
and
me hods
Algo i hm
de elopmen
The
me hod
u ilized
in
his
pape
is
summa ized
in
he
schema
illus a ed
in
Fig.
1.
Du ing
he
p ep ocessing
phase
he
g een
nega ed
channel
(gneg)
is
ob ained
om
he
RGB
image.
In
his
channel
he
shiny
elemen s
ha
ep esen
he
op ic
disk
and
exuda es
(i
any)
a e
segmen ed
by
his og am
h esholding,
ob aining
a
bina y
image
whe e
he
pixels
belonging
o
said
elemen s
ha e
a
alue
o
1
and
he
emainde
P ep ocessing Seed selec ion
R egion g ow h
Classi ica gion
Fig.
1
–
Algo i hm
flow
diag am.
will
ha e
a
alue
o
ze o.
Subsequen ly,
he
ascula
ee
is
seg-
men ed
wi h
a
mul i- ole ance
h esholding
me hod7whe e
a
h eshold
condi ion
depending
on
he
su ace
occupied
by
he
ascula
ee
is
added.
The
esul
is
a
bina y
image
whe e
he
pixels
belonging
o
he
ascula
ee
ha e
a
alue
o
1
and
he
es
o
he
image
has
a
alue
o
ze o.
Said
bina y
image
is
combined
wi h
he
shiny
elemen s
image
o
a ain
a
bina y
mask
(mbin).
A e
he
ascula
ee
segmen a ion,
gneg
is
submi ed
o
a
polynomic
con as
enhancing
p ocess
which
p oduces
he
C(gneg)
image.
Subsequen ly,
he
image
is
submi ed
o
2
successi e
fil e ings.
Fi s ly,
a
3
×
3
pixel
Gaussian
fil e
is
applied
ha ing
a
alue
o

=
2.
Then
his
image
is
submi ed
o
a
high
s ep
fil e ing
o
emo e
he
undus
and
main ain
he
elemen s
ha
cons i u e
a
sudden
change
in
image
in ensi y,
hus
ob aining
he
s anda dized
image
<HP(C(gneg))>.
Finally,
he
ascula
ee
and
he
shiny
elemen s
de ec ed
in
mbin a e
emo ed
om
he
image,
which
causes
ha
he
emaining
pixels
ha e
a
alue
o
ze o.
Fig.
2
illus a es
he
p ep ocessing
schema
and
Fig.
3
shows
an
example
wi h
some
o
he
images
ob ained
wi h
said
p ep ocessing.
The
nex
s ep
consis s
o
illumina ing
he
ascula
ee
and
shiny
elemen s
om
he
<HP(C(gneg))>
image,
so
ha
all
he
pixels
p esen
in
mbin a e
0.
In
he
Seed
Selec ion
phase,
a
new
mul i- ole ance
h esh-
olding
o
he
image
is
ca ied
so
ende
a
maximum
o
1000
egions
which
could
be
mic oaneu ysm
seeds.
Du ing
h esh-
olding,
he
pixel
wi h
highes
in ensi y
in
each
segmen ed
egion
is
ma ked
as
a
seed
candida e.
These
po en ial
seeds
a e
subsequen ly
alida ed
by
calcula ing
o
each
he
s a is-
ical
ail
a io
ha
p o ides
in o ma ion
abou
he
dis ibu ion
o
in ensi ies
in
he
his og am
o
a
window
cen e ed
on
he
pixel
being
s udied.8
In
each
selec ed
seed
a
egion
g ow h
is
applied
in
which
neighbo ing
pixels
a e
added
ha ing
an
in ensi y
abo e
he
mean
alue
plus
he
ypical
de ia ion
o
a
window
cen e ed
on
he
seed
ha ing
a
size
o
10
×
10
pixels.
Fo
he
g owing
egion
o
be
alid
i
mus
no
ha e
a
su ace
exceeding
30
pixels
because
in
ou
da abase
and
a
he
esolu ion
we
wo k
wi h
he e
is
no
mic oaneu ysm
ha ing
an
ex ension
abo e
said
alue
(Fig.
4).
In
o de
o
classi y
egions,
15
cha ac e is ics
a e
calcula ed
o
each
g own
seed−numbe ed
om
1
o
15–o
which
4
a e
286
a
c
h
s
o
c
e
s
p
o
a
l
m
o
l
.
2
0
1
2;8
7(9):284–289
+
-
RGB image
G een
nega ed
channel
Bina y mask
( ee and shiny
elemen s)
Con as
enhancemen
Gaussian
il e ing
A e aging
il e ing
Fil e ed image
Fig.
2
–
P ep ocessing
schema.
calcula ed
on
he
basis
o
he
bina y
ep esen a ion
o
each
g own
egion,
4
on
he
basis
o
he
gneg
image,
4
on
he
basis
o
he
C(gneg)
image
and
3
on
he
basis
o
he
<HP(C(gneg))>
image.
The
bina y
image
only
coun s
wi h
pixels
ha ing
al-
ues
o
ze o
and
one,
wi hou
quan ifiable
in ensi y
o
con as
di e ences.
I
is
he
mos
adequa e
image
o
de e mine
mo -
phological
pa ame e s.
The
4
cha ac e is ics
a e:
1) Su ace
(Si):
sum
o
pixels
belonging
o
egion
i.
2)
Aspec
a io:
he
esul
o
di iding
he
la ge
diame e
by
he
smalle
diame e
o
he
egion.
3) Pe ime e
(Pi):
he
numbe
o
pixels
loca ed
a
he
edge
o
egion
i.
4)
Ci cula i y
(Ci):
his
cha ac e is ic
p o ides
in o ma ion
abou
he
shape
o
egion
i,
and
is
calcula ed
by
means
o
Ci=
P2
i/(4Si).
The
4
cha ac e is ics
aken
on
he
basis
o
egions
in
gneg
a e:
Fig.
3
–
(a)
gneg,
(b)
mbin,
(c)
C(gneg),
and
(d)
<HP(C(gneg))>.
a
c
h
s
o
c
e
s
p
o
a
l
m
o
l
.
2
0
1
2;8
7(9):284–289
287
Fig.
4
–
(a)
Cu ou
o
C(gneg).
(b)
Resul
o
he
egion
g ow h
in
he
cu ou :
he
ligh e
poin s
o
he
seed
pixels
de ec ed
by
he
algo i hm;
he
g ay
ex ensions
a e
he
candida e
poin s
wi h
he
final
egion
g ow h.
The
mic oaneu ysms
de ec ed
by
he
specialis
a e
shown
wi hin
ed
squa es.
1) Mean
in ensi y
A
(<IiA>):
he
sum
o
he
pixel
in ensi ies
o
egion
i
di ided
by
he
su ace
o
said
egion.
2)
Co ela ion
wi h
Gaussian
A:
p o ides
in o ma ion
abou
he
simila i y
o
egion
i
wi h
a
Gaussian
s uc u al
elemen
o
he
same
size
o
egion
i.
3) Con as
A:
i
measu es
he
di e ence
be ween
he
mean
in ensi y
o
egion
i
pick
cells
and
adjacen
pixels.
4)
Fundus
A
mean
in ensi y
(<I A>):
he
mean
o
pixels
belonging
o
a
window
ha ing
a
size
o
15
×
15
pixels
cen-
e ed
on
he
ini ial
seed
o
each
egion.
The
cha ac e is ics
calcula ed
on
he
basis
o
C(gneg)
a e:
1)
Mean
in ensi y
B
(<IiB>).
2)
Co ela ion
wi h
Gaussian
B.
3)
Con as
B.
4)
Fundus
B
mean
in ensi y
(<I B>).
In
o de
o
ob ain
he
<HP(C(gneg))>
image,
a
high
fil e
was
applied
o
elimina e
he
undus.
Fo
his
eason,
his
image
does
no
allow
an
assessmen
o
he
mean
undus
in ensi y.
The
de e mined
cha ac e is ics
in
his
image
a e:
1)
Mean
in ensi y
C
(<IiC>).
2)
Co ela ion
wi h
Gaussian
C.
3) Con as
C.
A
selec ion
was
made
o
de e mine
cha ac e is ics
wi h
g ea e
disc imina ing
po en ial
by
means
o
o -
wa ded
sequen ial
selec ion
and
backwa d
sick
when
she’ll
alimen a ion.
Fo
classifica ion,
a
Fuzzy
A map9 ype
neu onal
ne wo k
was
u ilized,
which
was
ained
wi h
a
g oup
o
images
o
subsequen ly
classi y
a
di e en
g oup.
Re inog aphs
The
baseline
was
a
da abase
comp ising
42
e inog aphs
o
he
A
Diabe ic
Re inopa hy
Ea ly
De ec ion
Plan,
he
ini ial
assessmen
o
which
mus
be
ca ied
ou
by
p ima y
heal h
ca e
physicians
(PHP)
o
he
hospi al
a ea
o
aining
he
neu onal
ne wo k.
Said
images
we e
e iewed
by
2
o
he
au ho s
(SJ
and
PA)
who
poin ed
ou
204
mic oaneu ysms.
Subsequen ly,
11
successi e
images
we e
aken
o
11
pa ien s
om
he
same
ea ly
de ec ion
p og am
diagnosed
wi h
dia-
be ic
e inopa hy
by
he
PHP
and
e iewed
by
he
au ho s
o
e alua ing
he
algo i hm.
The
images
we e
aken
wi h
a
Topcon
NW-200
e inog aph
and
s o ed
in
he
JPG
o -
ma .
Compu e
equipmen
The
algo i hm
was
de eloped
wi h
he
MATLAB
7.6.0
(R2008a;
Ma hWo ks)
ma hema ical
so wa e
on
an
HP
Compaq
dc7600
Con e ible
compu e
wi h
an
In el®Pen ium®4
p ocesso
unning
on
a
CPU
o
3.20
GHz
and
wi h
0.99
GB
o
RAM
mem-
o y.
S a is ics
The
a e age
sensi i i y
o
he
algo i hm
and
he
numbe
o
alse
posi i es
o
each
image
was
calcula ed.
Resul s
Ini ially,
42
images
wi h
204
mic oaneu ysms
we e
selec ed.
Du ing
he
seed
selec ion
p ocess,
seeds
we e
placed
in
167
mic oaneu ysms.
A e
egion
g ow h,
he
algo i hm
seg-
men ed
50,184
egions,
o
which
153
we e
mic oaneu ysms
(75%),
and
he
emaining
15,031
we e
alse
posi i es.
These
we e
he
egions
o
which
he
15
cha ac e is ics
discussed
abo e
we e
calcula ed.
Du ing
he
neu onal
ne wo k
aining
he
bes
esul s
we e
ob ained
wi h
cha ac e is ics
numbe
4,
7,
9,
13
and
15,
ob aining
an
a e age
sensi i i y
o
92.85%
on
he
classifica ion
g oup,
wi h
he
classifie
exhibi ing
a
hi
a e
o
92.45%.
Once
he
ne wo k
was
ained,
he
g oup
o
11
e inog aphs
con aining
52
diagnosed
mic oaneu ysms
was
classified.
288
a
c
h
s
o
c
e
s
p
o
a
l
m
o
l
.
2
0
1
2;8
7(9):284–289
A e
he
seed
selec ion
p ocess
and
egional
g ow h,
a
o al
amoun
o
32
segmen ed
mic oaneu ysms
we e
ob ained
in
addi ion
o
3752
segmen ed
egions
ha
we e
no
mic oaneu ysms.
A e
classi ying
said
egions,
38
segmen ed
mic oaneu ysms
we e
ob ained
oge he
wi h
135
alse
posi i es,
which
ansla es
in o
a
sensi i i y
o
73.08%
and
a
mean
alue
o
12.27
alse
posi i es
pe
image.
Discussion
In
he
pas
a ious
me hods
ha e
been
p oposed
o
au oma -
ically
de ec ing
mic oaneu ysms.10–12 Howe e ,
said
me hods
ha e
always
been
assessed
wi h
p op ie a y
image
da abases
which
we e
di e en
o
each
au ho
and
had
di e en
echni-
cal
cha ac e is ics
and
e e ence
pa ame e s.
Niemeije
e
al.
ca ied
ou
an
online
expe ience
in
which
hey
made
a ail-
able
o
a ious
esea ch
g oups
an
images
da abase
aken
wi h
non-mid ia ics
e inog aphs,
s o ed
in
JPEG
o ma
in
o de
o
apply
a ious
algo i hms
and
wi h
esul s
eco ded
in
a
homogeneous
manne
by
means
o
FROC
cu es.2In
said
s udy,
hey
di e en ia ed
4
ypes
o
mic oaneu ysms.
On
he
basis
o
hei
size
and
isibili y
hese
a e:
minimum,
medium
and
e iden .
Acco ding
o
loca ion,
mic oaneu ysms
close
o
essels
a e
conside ed
in
a
specific
manne .
A
specific
assess-
men
was
made
o
each
mic oaneu ysm
ype
and
an
o e all
assessmen
in
which
he
de ec ion
o
he
4
ypes
was
aken
join ly.
In
he
5
de ec ion
me hods,
execu ed
by
5
esea ch
g oups
ha
accep ed
he
online
challenge,
a
di ec
ela ion-
ship
was
app ecia ed
be ween
he
g ea e
sensi i i y
o
he
algo i hm
and
he
inc ease
in
he
numbe
o
alse
mic oa-
neu ysms
pe
image.
When
he
algo i hms
analyzed
he
o al
amoun
o
mic oaneu ysms,
o e
10
alse
mic oaneu ysms
we e
de ec ed
by
image
o
aise
sensi i i y
alues
be ween
50%
and
60%.
On
he
basis
o
he
images
supplied
in
his
s udy,
he
me hod
de eloped
by
ou
esea ch
g oup
p esen ed
in
he
2009
CASEIB7was
applied,
ob aining
sensi i i y
esul s
o
16.53%
and
an
a e age
o
41.8
alse
posi i es
pe
image.
This
ep esen s
a
significan
imp o emen
is-
à- is
he
p e ious
me hods
and
alidi y
simila
o
ha
p esen ed
in
he
a ious
me hods
o
he
online
chal-
lenge.
The
iden ifica ion
o
small
second
o
hi d
o de
a e-
ioles
and
enules
as
isola ed
segmen s
is
he
mos
usual
sou ce
o
alse
mic oaneu ysms.
Fo
his
eason,
he
co -
ec
segmen a ion
o
he
ascula
ee
is
one
o
he
key
s eps
in
au oma ic
de ec ion
algo i hms.
In
u n,
he
seg-
men a ions
depend
on
he
no o ie y
o
he
s uc u es
a e
comple ing
image
p ep ocessing
and
ini ial
p ocess-
ing.
The
nex
s ep
in
ou
esea ch
will
ocus
on
inc easing
he
sys ems
sensi i i y
by
means
o
analyzing
new
app oaches
in
egion
segmen a ion
because
his
is
whe e
sensi i i y
is
educed
he
mos .
Ou
esea ch
con inues
o
diminish
he
a e
o
alse
posi i es
pe
image,
including
new
p e-
and
pos -p ocessing
me hods
ha
a e
cu en ly
in
expe imen a-
ion.
Conclusions
Wi h
he
algo i hm
p esen ed
in
his
pape ,
significan
imp o emen s
ha e
been
ob ained
in
he
au oma ic
de ec ion
o
ini ial
diabe ic
e inopa hy
inju ies,
inc easing
he
sensi i -
i y
o
he
p e ious
algo i hm
o
mic oaneu ysms
om
60.53%
o
73.08%,
and
diminishing
he
numbe
o
alse
posi i es
pe
image
om
41.08
o
12.27.
Howe e ,
he
me hod
s ill
equi es
addi ional
imp o emen s
so
ha
he
design
ool
can
be
com-
pa ed
o
human
de ec ion
wi h
su ficien
eliabili y
o
become
a
use ul
ool
in
clinical
diabe ic
e inopa hy
de ec ion
p o-
g ams.
Funding
Funded
by
he
Heal h
Resea ch
Fund
o
Resea ch
P ojec s
FIS
ETES-PI07/90379,
ETES-PI07/90373.
Conflic
o
in e es s
No
conflic
o
in e es s
has
been
decla ed
by
he
au ho s.
e
e
e
n
c
e
s
1.
Fong
DS,
Aiello
L,
Ga dne
TW,
King
GL,
Blankenship
G,
Ca alle ano
JD.
Diabe ic
e inopa hy.
Diabe es
Ca e.
2003;26:99–102.
2.
Niemeije
M,
an
Ginneken
B,
C ee
MJ,
Mizu ani
A,
Quellec
G,
Sanchez
CI,
e
al.
Re inopa hy
online
challenge:
au oma ic
de ec ion
o
mic oaneu ysms
in
digi al
colo
undus
pho og aphs.
IEEE
T ans
Med
Imaging.
2010;29:185–95.
3.
Spence
T,
Phillips
RP,
Sha p
PF,
Fo es e
JV.
Au oma ed
de ec ion
and
quan ifica ion
o
mic oaneu ysms
in
fluo escein
angiog ams.
G ae es
A ch
Clin
Exp
Oph halmol.
1992;230:36–41.
4.
C ee
MJ,
Olson
JA,
McHa dy
KC,
Sha p
PF,
Fo es e
JV.
A
ully
au oma ed
compa a i e
mic oaneu ysm
digi al
de ec ion
sys em.
Eye.
1997;11:622–8.
5.
Sánchez
CI,
Ho ne o
R,
Mayo
A,
Ga cía
M.
Mix u e
model-based
clus e ing
and
logis ic
eg ession
o
au oma ic
de ec ion
o
mic oaneu ysms
in
e inal
images.
In:
Ka ssemeije
N,
Gige
L,
edi o s.
SPIE
medical
imaging
2009:
compu e -aided
diagnosis.
P oceedings
o
SPIE,
ol.
7260.
2009.
p.
72601M1–M7.
6.
Zhang
B,
Wu
X,
You
J,
Li
Q,
Ka ay
F.
Hie a chical
de ec ion
o
ed
lesions
in
e inal
images
by
mul iscale
co ela ion
fil e ing.
In:
Ka ssemeije
N,
Gige
L,
edi o s.
SPIE
medical
imaging
2009:
compu e -aided
diagnosis.
P oceedings
o
SPIE,
ol.
7260.
2009.
p.
72601L1–L6.
7.
Nú ˜
nez
FJ,
Se ano
C,
Acha
B,
Fondón
I,
Jiménez
S,
Alemany
P.
De ección
Au omá ica
de
Mic oaneu ismas
en
Re inog a ías
pa a
Diagnós ico
P ecoz
de
Re inopa ía
Diabé ica.
CASEIB.
2009;27:581–4.
8.
Acha
B,
Se ano
C,
Rangayyan
RM.
De ec ion
o
mic ocalcifica ions
in
mammog ams
using
2d
p edic ion
fil e ing
and
a
new
s a is ical
measu e
o
he
igh
ail
weigh .
P oc
Embec.
2005:3112–7.
9. Ca pen e
A.
Fuzzy
ARTMAP:
a
neu al
ne wo k
a chi ec u e
o
inc emen al
supe ised
lea ning
o
analog
mul idimensional
maps.
IEEE
T ans
Neu al
Ne w.
1992;3:698–713.

a
c
h
s
o
c
e
s
p
o
a
l
m
o
l
.
2
0
1
2;8
7(9):284–289
289
10. Niemeije
M,
an
Ginneken
B,
S aal
J,
Su o p-Schul en
MSA,
Ab amo
MD.
Au oma ic
de ec ion
o
ed
lesions
in
digi al
colo
undus
pho og aphs.
IEEE
T ans
Med
Imaging.
2005;24:584–92.
11.
Sin hanayo hin
C,
Boyce
JF,
Williamson
TH,
Cook
HL,
Mensah
E,
Lal
S,
e
al.
Au oma ed
de ec ion
o
diabe ic
e inopa hy
on
digi al
undus
images.
Diabe
Med.
2002;19:
105–12.
12.
Quellec
G,
Lama d
M,
Josselin
PM,
Cazuguel
G,
Cochene
B,
Roux
C.
Op imal
wa ele
ans o m
o
he
de ec ion
o
mic oaneu ysms
in
e ina
pho og aphs.
IEEE
T ans
Med
Imaging.
2008;27:1230–41.