Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
11
,
No.
3
,
Septem
ber
201
8
, pp.
1
083
~
1093
IS
S
N:
25
02
-
4752
,
DOI: 10
.11
591/
ijeecs
.
v1
1.i
3
.pp
1083
-
1093
1083
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Autom
ated
Det
ection
o
f Mi
croan
eurysms
u
sing Probabil
ist
ic
Cascade
d Neural
Networ
k
Jeyapri
ya J,
Umadevi
K
S,
Jagadee
sh
K
an
n
an
R
School
of
Co
m
p
uti
ng
Sci
ence &
Engi
ne
eri
ng,
V
e
ll
ore
Instit
u
te of Te
chno
log
y
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
May
1,
2018
Re
vised Ju
n 2
1, 20
18
Accepte
d
J
un
28, 201
8
The
dia
gnosing
fea
tur
es
for
Diabe
t
ic
Re
ti
no
pat
h
y
(DR)
co
m
prises
of
fea
tur
es
occ
urr
i
ng
in
and
aro
un
d
the
reg
ions
of
blood
vessel
z
one
which
will
result
int
o
exude
s,
hemorr
hage
s,
m
ic
roa
n
e
ur
y
sm
s
and
gene
ration
of
te
xtur
es
on
the
al
bum
en
reg
ion
of
e
y
e
balls.
I
n
thi
s
stud
y
we
pre
senta
proba
bil
ist
ic
con
volut
ion
neu
ral
net
work
base
d
a
lgori
thms
,
utilized
for
the
ext
ra
ct
ion
of
su
ch
fe
at
ure
s
fro
m
the
retina
l
i
m
age
s
of
pat
i
en
t’s
e
y
eb
al
ls.
The
class
ifi
c
at
i
ons
profic
ie
nc
y
of
var
ious
DR
sy
st
ems
is
ta
bula
t
ed
and
exa
m
ine
d
.
The
m
aj
ority
of
the
rep
orte
d
s
y
st
ems
are
profoundl
y
adva
n
ce
d
reg
ard
ing
th
e
a
naly
z
ed
fundus
images
is
ca
tc
hing
up
to
the
hum
an
ophtha
lmologist
’s c
har
ac
t
eri
z
ati
on
ca
p
acities.
Ke
yw
or
d
s
:
Bl
ood vessel
Detect
ion
of
diabeti
c
reti
nopathy
Re
ti
nal n
er
ve h
e
m
or
r
hag
e
s
m
ic
ro
ane
ur
ys
m
s
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Jey
a
pr
iy
a J,
School
of Com
pu
ti
ng
Scie
nce
&
E
ngineeri
ng,
Vell
or
e
Insti
tute o
f
Tec
hnolog
y, I
nd
ia
.
Em
a
il
:
pr
iy
acse
m
e@g
m
ail.co
m
1.
INTROD
U
CTION
Ra
pid
ad
va
nce
m
ent
of
diabet
es
has
bee
n
th
e
m
os
t
pr
om
inent
fun
dam
ent
al
diff
ic
ulti
es
of
a
vaili
ng
pro
per
h
eal
th
protect
io
n
ser
vice. Th
e m
easur
e o
f
ind
i
vidual
s inf
lue
nced
with the d
ise
ase keeps o
n
devel
op
i
ng
at
a
distur
bing
rate.
The
WHO
(
World
Healt
h
Orga
niza
ti
on)
ex
pe
ct
s
the
qu
a
ntit
y
of
ind
i
vidua
ls
with
diabeti
cs
to
inc
rem
ent
fr
om
m
or
e
tha
n
125
m
il
li
on
to
over
350
m
i
ll
ion
thr
ough
ou
t
t
he
fo
ll
ow
i
ng
tw
o
dec
ades
[1
]
.
T
he
ci
rcum
s
ta
nce
is
exa
cerb
at
e
d
by
th
e
way
that
on
e
an
d
on
ly
50%
of
the
patie
nts
kn
ow
ab
ou
t
there
pro
gr
essi
on
of
disease.
Wh
at
'
s
m
or
e,
worse
ns
the
sit
uatio
n
is
as
thera
pe
utica
ll
y,
diabetes
prom
pts
to
caus
e
extrem
e
com
pl
ic
at
ion
s
as
it
pro
gr
e
sses
furth
er.
T
hese
i
ntri
caci
es
inco
rpo
rate
la
rg
e
scal
e
an
d
sm
al
le
r
scale
vasc
ular
cha
nges
wh
ic
h
resu
l
t
in
cor
ona
ry
il
lness,
renal
issues
an
d
reti
nopat
hy.
For
in
sta
nce,
surv
ey
in
the
USA
al
one
de
m
on
strat
e
that
diabetes
is
co
nsi
der
e
d
t
o
the
5
th
m
os
t
de
adl
ie
st
disease,
an
d
ti
ll
now
the
r
e
is
no
cur
e
[
2].
Diabe
ti
c
reti
nopathy
(D
R)
as
sho
wn
in
Fig
ur
e
1
is
a
ty
pical
aftera
ff
ect
of
diabete
s.
Und
oubtedl
y,
it
is
com
m
on
to
the
point
t
hat
it
is
the
m
ai
n
so
urce
of
vis
ua
l
def
ic
ie
ncy
i
n
the
work
i
ng
popu
la
ce
of
we
ste
rn
nations
[
3],
[4
]
.
The
rate
of
diabetes
is
ex
pandin
g
at
an
al
ar
m
ing
rate.
Re
gret
ta
bly,
m
os
t
of
the
natio
ns
of
te
n
una
ble to
r
ec
or
d DR cases
and th
us nee
d fun
dam
ental
r
ecordin
g
strat
e
gies
for regist
eri
ng
DR cases
[5
].
Early
disco
very
of
DR
is
cr
it
ic
al
,
in
l
igh
t
of
the
fact
th
at
treatm
ent
st
rategies
can
de
rail
ed
the
adv
a
ncem
ent
of
t
he
disease
.
Ma
jority
of
treatm
ent
strat
eg
ie
s
de
pe
nd
on
la
se
r
ba
sed
m
et
ho
ds.
Laser
photo
c
oa
gu
la
ti
on
cl
os
e
s
up
visu
al
bloo
d
ve
ssels,
wh
ic
h
adequate
ly
stop
s
their
sp
il
la
ge.
The
ce
ntra
l
la
ser
treatm
ent strate
gy
dim
inishes r
et
inal t
hic
keni
ng
[
6
]
-
[
8
]
. T
his m
ay
co
un
te
ra
ct
d
ecl
inin
g of
reti
nal sw
el
li
ng.
T
o
be
par
ti
cula
r,
t
his
treat
m
ent
le
ssens
t
he
da
nger of
visio
n
m
isfort
une b
y
ha
lf.
F
or
a
li
tt
le
nu
m
ber
of
cases
,
with
aggre
gate v
isi
on m
isf
or
tu
ne, c
hange is
conce
ivable
[9
].
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
11
, N
o.
3
,
Se
ptem
ber
2
01
8
:
1
0
8
3
–
1
0
9
3
1084
Figure
1. Illust
rati
on of
d
iffe
r
ent f
eat
ur
es
in
vo
l
ved in
DR i
nf
l
uen
ce
d reti
na
l
i
m
ages
An
al
ysi
s
of
r
et
inal
pictures
is
a
gr
owin
g
rese
arch
area
that pr
ese
ntly
dr
a
w
s
in
bunch
e
s
of
enthu
sia
sm
from
bo
th
rese
arch
e
rs
an
d
do
ct
or
s.
T
he
goal
of
this
fiel
d
is
to
create
com
pu
ta
ti
onal
m
e
t
hods
w
hich
will
help
m
easur
em
ent an
d re
pr
ese
ntati
on of i
ntri
gu
i
ng
path
ology an
d
a
natom
ic
al
st
ru
ct
ur
es
. T
hese
an
al
ysi
s f
ram
ework
works
with
a
dv
a
n
ce
d
f
undus
im
ages
of
the
ey
e
[1
0
]
.
The
strat
egy
of
ta
ki
ng
f
un
du
s
im
ages
beg
ins
by
wide
ning
th
e
pupil
as
sho
wn
i
n
Fi
gure
2
with
ph
a
rm
aceutical
ey
e
dro
ps
.
Af
te
r
that
t
he
patie
nt
is
requ
est
ed
to
aff
ix
gaz
e
with
a
sp
eci
fic
en
d
go
al
t
o
recor
d
the
reti
na
in
m
ot
ion
le
ss
pos
it
ion
.
Wh
il
e
ta
king
the
photos,
the
patie
nt
will
see
a
pr
og
ressio
n
of
re
peated
f
la
sh
es.
T
he
w
hole
proce
dure
ta
kes
ar
ound
fi
ve
to
te
n
m
inu
te
s.
T
o
gu
a
ra
ntee
tha
t
DR
treat
m
ent
is
pe
rfor
m
ed
on
ti
m
e,
the
ey
e
fundus
im
ages
of
dia
betic
patie
nts
m
us
t
be
analy
zed
in
an
y
even
t
once
e
ver
y
ye
ar
[
11
]
.
In
cl
ude
e
xtrac
ti
on
te
ch
nique
s
an
d
e
xam
inati
on
Im
age
ha
nd
li
ng
can
do
both
de
crease
the
w
orkloa
d
of
scree
ner
s
a
nd
ass
um
e
a
fo
c
al
p
art
in
qu
al
it
y
aff
irm
ation
unde
rt
akin
gs
.
In
this
way,
t
her
e
has
bee
n
an
expansi
on
in
the
us
e
of
com
pu
te
rize
d
pictu
re
ha
ndli
ng
proce
dur
es
for
pro
gr
am
m
ed
identific
at
io
n
of
DR
[
12
]
.
F
or
instance,
s
had
i
ng
feat
ur
es
on
Ba
ye
sia
n
m
eas
ur
a
ble
cl
assifi
e
r
wer
e
ut
il
iz
ed
to
c
ha
racteri
ze
e
ver
y
pix
el
into
l
esi
on
or
non
-
le
si
on
cl
asses
[13
]
.
T
he
acc
om
pan
yi
ng
areas
portray
set
s
of
reti
na
l
scenari
os
su
c
h
as:
hae
m
or
rh
age
s,
e
xudes
,
m
acul
op
at
hyan
d
m
i
cro
a
ne
ur
ysm
s.
Su
c
h
identific
at
ion
proce
dures
yi
e
ld
the
m
ajo
rity
of
t
he
feat
ur
e
s
w
hich
are
util
iz
ed
as
a
pa
rt
of
com
pu
te
rize
d
DR
detect
ion f
ram
ewor
k.
Figure
2. Sam
ple D
R i
nfl
ue
nc
ed
im
ages f
or t
hr
ee
stages
i.e
., (a)
No
rm
al
,
(b
)
Mo
de
rate, (
c
)
Sev
e
re
Adva
nced
f
un
du
s
photogra
phy
f
ro
m
the
hum
an
ey
e
giv
e
s
cl
ear
im
ages
of
the
bl
oo
d
vessels
i
n
the
reti
na.
T
his
te
c
hn
i
qu
e
gi
ves
a
bri
ll
ia
nt
wind
ow
to
t
he
record
the
tr
ue
sta
te
of
a
patie
nt'
s
reti
na
i
nf
l
uenced
by
DR
[
14
]
.
T
he
blood
ve
ssel
st
ru
ct
ur
e
was
ex
tract
ed
by
s
ubje
ct
ing
t
he
gr
ee
n
par
t
of
the
RGB
f
undus
pic
ture
t
o
var
i
ou
s
im
ag
e
proc
essi
ng
a
lgorit
hm
s
[1
4
]
.
Bl
ood
vess
e
ls
wer
e
ide
ntifie
d
util
iz
ing
two
-
dim
e
ns
ion
al
coor
din
at
ed
c
ha
nn
el
s
[
15
]
.
T
he
Dark
le
vel
prof
il
e
of
cr
oss
area
of
bloo
d
vessels
that
are
been
ap
pro
xim
at
ed
by
Ga
us
sia
n
filt
ers.
The
idea
of
c
oor
din
at
ed
channel
locat
ion
of
si
gn
al
s
was
util
iz
ed
to
recogn
iz
e
piece
wis
e
strai
gh
t
sect
io
ns
of
blood
vess
el
s
after
the
ve
ssel
cl
assifi
er.
Vessel
f
ocuses
in
a
cro
ss
se
gm
ent
are
found
wit
h
a
fu
zzy
-
c
-
m
eans
cl
assifi
er
[
16
]
,
[
17
]
.
This
t
echn
i
qu
e
f
ound
a
nd
s
ketc
hed
ou
t
t
he
bounda
ries
of
blood
vessels
in
the
give
n
i
m
ages
by
the
util
iz
at
ion
of
a
novel
strat
egy
to
sect
ion
blood
vessel
s
that
com
pli
m
ents
neig
hbour
hood
vessel
pro
per
t
ie
s
with
area
ba
sed
qual
it
ie
s
of
the
r
et
inal
structu
re.
T
he
com
pu
te
r
suppo
rted
determ
inati
on
syst
e
m
been
de
velo
ped
t
o
he
lp
doct
ors
in
identify
in
g
ir
regularit
ie
s
co
nn
ect
e
d
with
f
undus
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Autom
ated
det
ect
ion
of m
ic
r
oane
ur
ys
ms usi
ng p
r
obabil
ist
ic
ca
sc
aded
(
Jeyapri
ya
J
)
1085
i
m
ages
of
the
reti
na
[18
]
.
Their
pr
opos
e
d
syst
e
m
can
recog
nize
bl
ood
vessel
co
nver
ge
nces
an
d
it
can
disti
nguish
pre
ci
se
widths
in
blood
ves
sel
s.
Ele
ct
ro
nic
syst
e
m
fo
r
both
e
xt
racti
on
a
nd
qua
ntit
at
ive
de
pi
ct
ion
of
t
he
f
unda
m
ental
vascu
l
ar
sym
pto
m
at
i
c
sign
s
from
fun
du
s
im
ages
in
hy
per
te
ns
i
ve
reti
nopa
thy
was
exh
i
bited
[
19
]
.
The
feat
ur
es
they
hav
e
c
onside
red
a
re
ve
ssel
tortuosit
y,
su
m
m
ed
up
and
ce
ntral
vesse
l
narrowin
g,
nea
rn
ess
of
Gun
n
or
Salus
si
gns.
Anot
her
syst
e
m
was
pr
op
ose
d
f
or
the
a
utom
at
ed
extracti
on
of
the
vasc
ular
st
ru
ct
ur
e
in
reti
nal
i
m
ages,
in
view
of
a
sca
nty
f
ollo
wing
m
et
ho
d
was
pro
posed
[20
]
.
Bl
ood
vessel
f
ocu
se
s
in
a
cro
ss
area
are
found
by
m
et
ho
d
f
or
a
f
uzzy
-
c
-
m
eans
cl
assifi
er.
I
n
th
e
wak
e
of
f
ollow
i
ng
te
chn
iq
ue,
the
vessels
disti
nguish
e
d
fr
a
gm
e
nts
we
re
ass
oc
ia
te
d
in
util
iz
i
ng
ava
rici
ous
com
pu
ta
ti
on
.
A
t
la
s
t
bif
ur
cat
io
ns
an
d
intersect
io
ns
wer
e
disti
ng
uis
hed
dissect
ing
vessel
en
d
ind
i
cat
es
with
difference
of
the
ve
ssel
structu
re.
B
lo
od
vessel
trac
ker
cal
culat
io
n
was
create
d
t
o
deci
de
the
r
et
inal
vascu
la
r
syst
e
m
is
ca
pture
d
util
iz
ing
an
ad
van
ce
d
cam
era
[21
]
.
T
he
trac
ker
cal
c
ulati
on
identifie
s
opti
c
ci
rcle,
le
sio
ns,
f
or
exam
ple,
cott
on
fleece
s
spots,
a
nd
fine
nerve
s
or
es
,
f
or
e
xam
ple,
hem
or
r
ha
ge
s.
This
cal
c
ulati
on
rec
ogniz
es
supp
ly
r
ou
t
es
an
d
veins
w
it
h an
e
xactness
o
f 7
8.4% a
nd
66.5
%
separ
at
el
y [
22
].
Figure
3. Re
su
l
ts of reti
nal
blood ve
ssel detec
ti
on
A
strat
egy
is
been
ad
opte
d
for
aut
om
at
ed
locat
ion
a
nd
arr
a
ng
em
ent
of
vasc
ular
i
rr
e
gu
la
ri
ti
es
in
diabeti
c
reti
nopathy
[23
]
.
Th
ey
reco
gniz
ed
vasc
ular
va
riat
ion
s
f
ro
m
the
no
rm
util
izing
scal
e
and
intr
oduc
ti
on
par
ti
cula
r
Ga
bor
c
ha
nn
el
banks.
T
he
pro
po
s
ed
strat
egy
ord
ers
reti
nal
im
a
ges
as
ei
the
r
ge
ntle
or
serio
us
case
s
in
li
gh
t
of
the
Gabo
r
channel
yi
el
ds
.
The
m
ic
ro
a
neurysm
s
in
reti
nal
fluoresciena
ng
i
ogr
a
m
s
was
reco
gniz
ed
by
fir
st
fi
nd
i
ng
the
fovea
by
s
ub
-
ins
pecti
ng
picture
by
fig
ure
of
f
our
e
very
m
easur
em
ent
[
24
]
.
C
onseq
ue
ntly
,
the
picture
wa
s
su
bject
e
d
to
m
idd
le
separ
at
ing
with
a
5
by
5
veil
to
di
m
inish
hi
gh
-
rec
urren
ce
se
gm
e
nts.
A
t
that
po
int
the
picture
was
r
el
at
ed
with
a
two
-
dim
ensional
ci
rcu
la
rly
s
ymm
et
ric
tria
n
gu
la
r
capaci
ty
with
disp
la
ye
d
n
et
sh
adi
ng
of
the
m
acula.
Bl
oo
d
-
vess
el
locat
ion
as
show
n
in
Figure
3
cal
cu
la
ti
on
in
view
of
th
e
te
rr
it
or
ia
l
rec
ur
si
ve
var
i
ou
s
le
veled
disin
te
gr
at
i
on
util
i
zi
ng
quadt
ree
s
an
d
post
-
fil
trat
ion
of
ed
ges
t
o
con
ce
ntrate
blo
od
vessels
was
co
ns
i
der
e
d
[
25
]
.
T
his
strat
egy
co
ul
d
dim
inish
bogus
re
j
ect
io
ns
of
pr
e
dom
inate
l
y
huge
ed
ges
and
quic
ke
r
i
n
c
on
t
rast
w
it
h
the
cu
rr
e
nt
appr
oach
with
le
ssene
d
sto
c
kp
il
in
g
pr
e
requisi
te
s
f
or
the
e
dg
e
delineat
e.
The
w
ork
util
iz
ed
the
arteriolar
-
to
-
ve
nu
la
r
distance
acro
s
s
pro
portion
of
reti
nal
bloo
d
ve
ssels
as
a
poi
nter
of
disease
relat
ed
c
hang
es
in
t
he
reti
na
l
blo
od
vessel
tre
e
[
26
]
.
T
hei
r
tria
l
resu
lt
s
sho
w
a
97.1%
achieve
m
ent
rate
in
th
e
reco
gniz
a
ble
pro
of
of
DR
ve
ssel
beg
in
ning
sta
ges,
an
d
a
99.
2%
achievem
ent
r
at
e
in
the
f
ollo
wing
of
reti
nal
vessels
.
Anot
her
te
ch
nique
for
s
urface
bas
ed
vessel
div
is
ion
t
o
beat
this
issue
was
pr
opos
e
d
[27
]
.
The
Fu
zz
y
CM
eans
(F
C
M)
bunc
hing
c
al
culat
ion
wa
s
util
iz
ed
to
ord
er
the
com
po
ne
nt
ve
ct
or
s
into
ves
sel
or
non
-
ve
ssel
in
view
of
the
surfac
e
prop
e
rtie
s.
They
co
ntrast
ed
their
te
chn
iq
ue
a
nd
handlabele
d
gr
ound
tr
uth
div
i
sion
for
fi
ve
im
ages
an
d
a
cc
om
plished
83.
27%
af
fectabil
it
y
and
99.62%
s
pecifi
ci
ty
.
The
m
e
tho
d
div
i
des
the
data
into
tw
o
pa
rts,
one
is
f
or
le
arn
in
g,
a
nd
a
no
t
her
is
f
or
te
sti
ng
.
Fo
r
t
he
purpo
se
of
i
den
ti
fyi
ng
f
undus
im
ages
incl
ud
i
ng
the
norm
al
c
la
ss
or
glauc
om
a
set
of
cl
asses.
Applic
at
ion
of
su
pp
or
t
vecto
r
m
achines
(SVM)
was
hi
ghly
us
e
d
is
show
n
in
Fig
ur
e
6
[
28
]
,
[29
]
.
The
determ
inati
on
of
sig
n
was
m
ade
from
the
ver
te
x
co
ordinat
es
based
on
th
e
deg
ree
s
of
th
e
diff
ere
nt
dire
ct
ion.
The
syst
em
was
te
ste
d
ba
sed
on
t
he
pe
rcen
ta
ge
of
s
uccess
rate
from
the
Har
ris
po
i
nt
detect
ion
a
nd
avail
abili
ty
fo
r
detect
ing
sig
n
on
diff
e
re
nt
ra
ng
e
.
T
he
res
ulted
on
e
sho
wn
that
not
al
l
Ha
rr
is
po
i
nt
in
t
he
was
i
m
age
detect
ed
al
thou
gh
m
os
t
of
the
im
ages
wer
e
possib
le
in
recogn
iz
i
ng
t
he
sig
n
di
recti
on
of
it
[
30
].
A
su
bpi
xel
-
accu
r
acy
ed
ge dete
c
ti
on
alg
or
it
hm
was
e
xp
l
or
e
d,
base
d
on t
he
w
avelet
tran
s
f
orm
at
ion
w
it
h
t
he
cubic
sp
li
ne
inter
po
l
at
ion
of
the
le
n'
s
m
od
ule
app
e
aran
ce qu
al
it
y
insp
ect
io
n
syst
e
m
.
It
do
es
firs
tl
y
ca
lc
ulati
on
of
the
m
axi
m
u
m
wavel
et
m
od
ul
us
,
and
de
te
ct
ion
of
a
pix
el
-
a
ccu
rate
edg
e.
Fi
na
ll
y
ind
us
t
rial
m
easur
em
ent
m
et
ho
ds
us
in
g
t
his s
ubpi
xel
-
accu
rate e
dg
e
d
et
ect
io
n
ba
sed
al
gorithm
w
as
stu
died [
31
].
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
11
, N
o.
3
,
Se
ptem
ber
2
01
8
:
1
0
8
3
–
1
0
9
3
1086
2.
PROBLE
M
I
DENTIFI
C
A
TION
Mi
cro
ane
ury
sm
s
reco
gnit
ion
is
i
m
per
at
ive,
on
the
gr
ounds
that
these
structur
e
s
co
ns
ti
tu
te
the
early
conspic
uous
el
e
m
ent
of
DR.
The
m
ai
n
repo
rts
w
hich
co
nn
ect
these
str
u
ct
ur
es
to
DR
go
bac
k
to
18
79
[32],
[33
]
.
All
the
m
or
e
as
of
la
t
e,
it
hav
e
anal
yz
ed
the
a
pp
e
a
ran
ce
a
nd
van
i
sh
in
g
of
m
ic
roaneury
sm
s
in
var
i
ou
s
per
i
od
s
of
flu
or
escei
n
a
ngio
gr
a
phy
[34
]
.
I
n
a
com
par
at
ive
stu
dy
both
arr
a
ng
em
ent
rate
an
d
va
nis
hing
of
m
ic
ro
ane
ur
ys
m
s
in
early
D
R
wer
e
analy
zed
[35].
The
m
ic
ro
ane
ur
ys
m
s
turnov
e
r
we
re
reg
ist
ere
d
r
el
ia
bib
ly
from
sh
adin
g f
undus im
ages [36
]
.
They
util
iz
ed
ano
t
her
strat
e
gy
cal
le
d
MA
-
tr
a
cker
to
num
ber
m
ic
ro
ane
ury
sm
s
as
sh
own
i
n
Fig
ur
e
4.
They
dem
on
str
at
ed
that
the
m
ic
ro
a
neurysm
s
sta
y
sta
ble
after
so
m
e
tim
e,
howe
ver
just
29%
sta
y
at
a
sim
il
ar
place.
I
n
il
lustr
at
ion
the
gree
n
segm
ent,
of
th
e
RGB
fundus
picture
,
was
gott
en
the
m
ic
ro
aneurysm
s.
Like
the
exudates
rec
og
niti
on
cal
culat
i
on,
first
the
c
on
s
pic
uous
str
uctu
res
insi
de
reti
na
im
ages,
for
exam
ple,
bloo
d
vessel
tree
an
d
opti
c
plate
are
to
be
e
xpel
le
d.
Af
te
r
th
at
an
adv
a
nce
d
su
cce
ssio
n
of
picture
pr
e
par
i
ng
al
gorithm
s
was
util
iz
ed
to
de
c
ide
the
z
on
es
inside
t
he
f
undus
im
ages
to
ge
t
m
ic
ro
ane
ury
s
m
s
see
in
Figure
4
[37]
,
[38].
T
he
autom
at
ed
disti
nguish
i
ng
featu
res
of
diabeti
c
reti
no
pathy
in
vie
w
of
the
nea
rness
of
m
ic
ro
ane
ur
ys
m
s
wer
e
co
ns
i
der
e
d
[
29]
.
Th
e
op
t
om
et
rists
accom
plished
97
%
a
ffec
ta
bi
li
ty
a
t
88
%
f
or
eve
ry
penny
sp
e
ci
fici
ty
and
the
m
echan
iz
ed
r
et
ino
pat
hy
in
dicat
or
acc
ompli
sh
e
d
85%
aff
ect
abili
ty
at
90
%
sp
eci
fici
ty
.
Figure
4. R
esu
l
ts of m
ic
ro
ane
ur
ysm
s d
et
ect
ion
3.
PROB
ABILI
STIC
C
ASCA
DED
NE
URA
L NETWO
R
K
Color
feat
ur
es
wer
e
util
iz
ed
on
Ba
ye
sia
n
m
e
asur
a
ble
cl
assi
fier
orde
r
e
very
pix
el
i
n
to
in
jury
or
non
-
so
re
cl
asses
.
They
hav
e
acc
om
pl
ished
10
0%
exactness
in
disti
nguish
in
g
al
l
the
reti
nal
i
m
ages
with
ex
ud
at
es
,
and
70%
prec
isi
on
i
n
order
i
ng
ty
pical
reti
nal
im
ages
as
ordi
nar
y.
DR
an
d
ordin
a
ry
reti
na
wer
e
gr
oupe
d
conseq
ue
ntly
ut
il
iz
ing
pictur
e
prepa
rin
g
a
nd
m
ul
ti
la
ye
r
per
c
eptr
on
ne
ural
s
yst
e
m
[3
9
]
.
Th
e
syst
e
m
yi
el
ded
a
n
aff
ect
abili
ty
of
80.
21%
a
nd
a
sp
eci
fici
ty
of
70.
66%.
C
om
pute
rized
a
naly
sis
of
NPDR,
in
li
gh
t
of
three
s
or
es:
hem
or
rh
a
ges
a
n
d
m
ic
ro
ane
ury
s
m
sh
ard
e
xudates,
an
d
cot
ton
fleec
e
spo
ts,
was
co
nce
ntrated
on
[
40
]
.
The
te
chn
iq
ue
c
ould
disti
ng
uish
t
he
N
PD
R
a
rr
a
ng
e
e
ff
ect
i
vely
with
an
exact
ness
of
81.
7%.
The
f
ollo
wing
ste
ps
represe
nt the a
dopted
m
et
ho
dl
og
y.
St
ep
1. N
orm
al
i
z
at
ion
Norm
al
i
za
ti
on
is
e
m
plo
ye
d
to
sta
nd
ar
dize
th
e
intensit
y
values
of
an
pictu
re
by
ad
j
us
ti
ng
the
var
y
of
it
s
gr
ey
-
le
vel
va
lues
in
order
that
they
li
e
a
m
on
g
a
desire
d
var
y
of
valu
es
e.g
.
zer
o
m
ean
an
d
unit
var
ia
nce.
Let
I(
i,j
)
de
no
te
s
the
gr
ay
-
le
vel
value
at
pi
ct
ur
e
e
le
m
ent
(i,j),
M
&
V
AR
de
no
te
the
cal
culable
m
e
an
&
var
ia
nce
of
I(
i
,
j
)
sev
e
rall
y &
N(
i,
j
)
d
e
-
note
s the
norm
al
iz
e
d gr
ay
-
le
vel va
lue at pict
ur
e e
lem
ent (
i,j
).
Algori
th
m
: A
n a
l
go
ri
th
m
f
or
Norm
aliz
at
ion
Inp
ut
:
I
nput
m
ic
ro
a
ne
ur
ysm
s
i
m
age I
(
x,
y),
wh
e
re
x &
y re
pr
ese
nts t
he pixel p
os
it
io
n.
Ou
tp
ut:
S
egm
ented ri
dge
re
gi
on
,
I
'
(x,
y).
St
ep
1.1:
Re
ad
Im
age I
(
x,
y).
St
ep
1.2
:
Ma
s
k
the
Ide
ntifyi
ng Regi
on M(
x,
y)
from
I
(x,
y
).
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Autom
ated
det
ect
ion
of m
ic
r
oane
ur
ys
ms usi
ng p
r
obabil
ist
ic
ca
sc
aded
(
Jeyapri
ya
J
)
1087
St
ep
1.3
:
Brea
k
m
asked
i
den
t
ifie
d
re
gion i
nt
o bloc
ks
b(
m
,
n)
×
b(m
,n
).
St
ep
1.4
:
Fi
nd
In
te
n
sit
y value
s
(
(
,
)
,
(
,
)
)
of the
ide
ntifie
d
m
asked
reg
i
on
al
blo
c
ks.
St
ep
1.5
:
E
val
uate Stan
da
rd
Dev
ia
ti
on
of i
ntensity
for ea
ch bloc
ks
(
(
,
)
,
(
,
)
)
.
St
ep
1.6
:
Re
pe
at
steps 1.
5
-
1.6
&ch
ec
k:
if (
<
)
{
I’(x,y
)
→
V(x,
y)
el
se
exit(0
)
}
wh
e
re,t is t
he
t
hr
es
hold li
m
it
&
V(
x,y)
is t
he
vecto
r
in
dices l
ocati
on w
it
hin
the ide
ntifie
d m
asked
reg
i
on.
St
ep
1.7
:
E
nd
Pr
oc
ess.
St
ep
2. Im
age
Orie
ntation
Or
ie
ntati
on
of
a
m
ic
ro
-
ane
ury
sm
s
is
cal
c
ulable
by
the
s
m
al
le
st
a
m
o
un
t
m
ean
sq
.
or
ie
ntati
on
est
i
m
ation
al
gorit
hm
ic
ru
le
gi
ven
by
H
ong
durin
g
is
s
how
n
in
Fig
ur
e
5
wh
ic
h
f
orem
os
t
a
blo
c
k
of
siz
e
w
X
w
(25X2
5)
is
cen
tred
at
picture
el
e
m
ent
(i,
j
)
within
the
no
r
m
al
iz
ed
m
ic
ro
-
ane
ur
ysm
s
i
mage
f
or
eve
r
y
pictur
e
el
e
m
ent
du
ri
ng
t
his
blo
c
k
rec
kon,
t
he
Ga
us
s
ia
n
gr
a
die
nts
x(i
,
j)
an
d
y(i
,
j)
f
or
eac
h
pi
xe
l
po
sit
ion,
that
area
un
it
the
gra
dient m
agn
it
ud
es
within t
he x
&
y direct
io
ns
se
ver
al
ly
.
Figure
5.
O
rien
ta
ti
on
Algori
th
m
: A
n a
l
go
ri
th
m
f
or
Orient
at
i
on
Inp
ut
:
I
nput
norm
al
iz
ed
im
age
I
’(x,
y)
,
wh
e
re
x
&
y
represe
nts
the
pix
el
po
sit
ion
,
ri
dg
e
l
ocati
on
(
(
,
)
),
gra
dient
devi
at
ion
,
′
blo
c
k gr
a
dient
&
or
ie
ntati
on
gradie
nt
Ou
tp
ut:
Direct
ion
of
the
rid
ge
s,
ang
le
,
+
(i
.e.,
1)
an
d
–
ve
(i.e.,
0)
re
pres
ents
the
cl
ockwise
or
cl
oc
kwise
directi
on
of
t
he
r
id
ges.
St
ep
2.1
:
Re
a
d Im
age I
’
(x,
y).
St
ep
2.2
:
Dete
rm
ine
Im
age
gr
adie
nt
′
blo
c
k
gr
a
dient
&
or
i
entat
ion
gradie
nt
re
sp
ect
ively
from
the
giv
e
n
gr
a
dient
de
viati
on
:
′
→
Der
i
vative
of
Gau
s
sia
nuse
d
t
o
c
om
pu
te
im
a
ge gra
dients.
→
Der
i
vative
of
Gau
s
sia
nS
i
gma
to
sm
oo
tht
he final
or
ie
ntati
on
vecto
r
fiel
d
(
v’)
St
ep
2.3
:
C
hec
k for
l
=
no. of
. ri
dge locat
io
n
(
,
)
:
for
i:l
{
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
11
, N
o.
3
,
Se
ptem
ber
2
01
8
:
1
0
8
3
–
1
0
9
3
1088
I
f
(
′
<
) {
return
→
+
el
se
return
→
−
}
E
nd
for
lo
op
St
ep
2.4
:
E
nd
Process
.
St
ep
-
3
Fe
at
ur
e Ext
r
ac
tion
Af
te
r
or
ie
ntati
on
im
age
m
ic
ro
ane
ury
sm
s
op
ti
on
s
(Ri
dg
e
D
ensity
,
Ri
dg
e
breat
h,
a
nd
nat
ural
de
pr
essi
on
width)
cal
culable.
Ridg
e
brea
th:
Ri
dg
e
breat
h
(
W
i
dth)
R
is
ou
tl
ined
as
thick
ness
of
a
ri
dg
e
it
’s
com
pu
te
d
by
inv
est
igati
ng
the
qu
a
ntit
y
of
pixe
ls
betwee
n
c
onsec
utive
m
axim
a
po
ints
of
pro
j
ect
ed
im
age,
va
riet
y
of
0’
s
betwee
n
2
cl
ust
ers
of 1’s ca
n offer
r
id
ge breat
h
e.g
.
11
110
000001
111
In on t
op of e
xam
ple, r
id
ge b
r
eat
h
(
widt
h)
is
6 pixels.
Va
ll
ey
bre
at
h
:
-
Vall
ey
br
eat
h
(
Widt
h)
V
is
ou
tl
ine
d
as
t
hi
ckn
ess
of
valle
ys
it
’s
com
pute
d
by
in
vestig
at
ing
the
quantit
y
of
pix
el
s
betwe
en
co
ns
ec
utive
m
ini
m
a
po
ints
of
pro
j
ect
ed
i
m
age,
var
ie
t
y
of
1’s
betw
een
2
cl
us
te
rs of 0
’s ca
n offer
n
at
ural
d
ep
ressi
on breat
h.
e.g
.
0000
11111110
00
In on t
op of e
xam
ple, n
at
ur
al
depressi
on bre
at
h
is
7 pixels.
Ridg
e
D
e
nsit
y:
Ri
dge
Densi
ty
is outl
ined
as
v
a
riet
y of
rid
ge
s in
a
ve
ry g
i
ve
n bloc
k.
e.g
.
00
111110
001111
1011
Abo
ve
stri
ng c
on
ta
in
s
3
ri
dg
e
s in
a
v
e
ry b
l
oc
k.
T
hus
rid
ge d
ensity
is 3
.
In
t
he
pr
ocess
involvin
g
cl
ass
ific
at
ion
of
m
i
cro
a
ne
ur
ysm
ss
rid
ge
brea
dth
and
w
hite
li
nes
area
un
it
cal
culat
ed.
Usin
g
t
he
pr
ob
a
bl
ist
ic
m
od
el
,
fir
st
we
create
a
gr
a
ph
m
od
el
fo
r
t
he
featu
re
set
s.
Let
us
s
uppose
that
a
giv
en
graph
netw
ork
da
ta
com
pr
ise
s
of
b+1
nu
m
ber
of
ve
rtic
es.
Th
us
,
it
can
be
m
od
el
le
d
with
th
e
hel
p
of
a
gr
a
ph
tre
e
,
wh
ic
h
is
giv
e
n
as:
G:
=(
N,
E
);
wh
e
re
N
represents
the
set
s
of
ver
ti
ces
i.e,
N=1,..
.,N
n
a
nd
E
is
the edge
set
w
i
th the ca
r
din
al
it
y i
s g
ive
n
as
|
|
=
.
Figure
6. Mult
i c
la
ss S
VM
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Autom
ated
det
ect
ion
of m
ic
r
oane
ur
ys
ms usi
ng p
r
obabil
ist
ic
ca
sc
aded
(
Jeyapri
ya
J
)
1089
Her
e
,
each
of
the
ed
ges
in
for
m
of
the
tree
i
s
rooted
f
or
in
dex
value
n
r
ul
ed
by
the
pro
ba
bili
ty
flow
from
on
e
e
dg
e
to
a
no
t
her
a
nd
i
s
der
i
vab
le
from
so
ur
ce
S
to
desti
nation
D
with
j
am
ount
of
dev
ia
ti
on
in
the
form
S
n
+jD
n
.
The
n
the
w
hole
branc
h
pro
ba
bili
ty
based
gr
aph
fl
ow
m
od
el
can
be
de
riv
ed
by
init
ia
li
sing
the
conditi
on
al
pr
obabili
ty
in
fo
r
m
of
a
sequ
e
nc
e
traver
se
d
f
rom
par
ent
to
chi
ld
nodes
of
tw
o
local
poste
rior
s
i.e,
pro
bab
il
it
y
of
app
ea
ra
nce
of
ver
ti
ces
(P
1
)
a
nd
that
of
e
dges
(P
2
)
w
hich
i
s
gi
ven
by:
1
=
(
|
1
1
:
)
&
2
=
(
|
2
1
:
)
.
This
is
repre
sented
i
n
the
form
of
seq
ue
ncized
finite
s
et
s
with
m
ulti
obj
ect
den
sit
i
es
of
1
:
ob
s
er
ved ed
ge si
te
s.
He
re,
t
he
sync
hron
iz
at
io
n betwee
n
s
uc
h po
ste
rio
r
is
m
ai
ntained
as:
(
|
1
:
,
2
:
)
=
(
|
1
:
∪
2
:
)
Now,
to
over
com
e
the
pr
oble
m
of
unknown
c
orrelat
ion
bet
ween
no
two
distri
bu
t
ion
s
of
in
de
pe
nd
e
nt
var
ia
bles the
s
olu
ti
on is:
(
|
1
:
,
2
:
)
∝
(
|
1
:
)
(
|
2
:
)
(
|
1
:
∪
2
:
)
Hen
ce
, th
e
g
e
ne
rali
zed
po
ste
rior relat
ionshi
p ca
n be
represe
nted
i
n
the
fo
r
m
o
f
ge
om
et
ric
m
ean:
(
|
1
:
,
2
:
)
=
(
|
1
:
)
1
(
|
2
:
)
2
∫
(
|
1
:
)
1
(
|
2
:
)
2
Wh
e
re,
1
,
2
(
1
+
2
=
1
)
the
par
am
et
ers
det
erm
ining
the
relat
ive
pro
ba
bili
ty
of
wei
ghte
d
distrib
ution
(
w
)
be
twee
n
a
s
pe
ci
fic
hiera
rch
i
cal
le
vel
of
c
hild
an
d
pa
ren
t
node
s.
Now,
in
order
to
a
nnon
ym
iz
e
the
pe
rco
la
te
d
m
od
el
of
gr
a
ph
da
ta
,
we
nee
d
a
ru
le
set
to
al
gorithm
ic
a
ll
y
el
i
m
inate
the
sensiti
ve
ver
ti
c
es
or
add
e
xtra
ed
ge
s
betwee
n
la
be
ll
ed
ver
ti
ces
to
disrupt
the
pr
ob
a
bili
ty
of
find
i
ng
se
ns
it
ive
inform
at
ion
.
In
this
way we
can
i
nduce
ano
nym
i
t
y i
n
a
qu
a
ntit
at
ive w
ay
over
the
gr
a
ph
datab
ase.
Algori
th
m
:
Pr
obabil
istic
CN
N
Inpu
t:
×
RGB st
and
a
r
d
Im
age
Wh
e
re,
m
&
n are the
row &
colum
n
of the
giv
e
n
im
age.
Out
p
ut:
CC
, c
lustere
d
fe
atu
r
es an
d
Pix
el
lo
cation of
micr
oane
ur
ys
m
i
mag
es
//
f
or brea
king
dow
n pixels i
nt
o
no
rm
alized ill
u
m
inati
on
r
e
f
le
ct
ance f
ie
ld
if (
≤
(
|
1
:
,
2
:
)
))
{
=
(
|
,
)
E
lse
=
(
|
,
)
}
*/
Wh
ere
,
M
is
the
m
od
el
of
ne
rv
e
c
olor
al
so
em
bar
ked
as
low
inte
nsi
ty
pix
el
s.
&
∑
are
m
ean
&
cov
a
riance
of t
he pixel
distrib
ution ba
sed
on i
ntensiti
es in
rg
col
or sch
em
e after
pr
e
-
pr
oce
ssing.
*/
//
Now,
(
1
,
1
),
(
2
,
2
),
…,(
,
)
∈
wh
il
e
l
≤
m
in
(
∑
(
,
)
=
1
)
//
D
is t
he displa
ce
m
ent v
ect
or
//
Evaluate t
he f
easi
ble v
al
ue o
f
ta
r
get no
des (
pix
el
s)
:
=
∑
(
,
)
=
1
Creat
e target
ve
ct
or
s
f
or
feasi
ble n
ei
ghbouri
ng no
des:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
11
, N
o.
3
,
Se
ptem
ber
2
01
8
:
1
0
8
3
–
1
0
9
3
1090
Lo
op
:
f
or
1
t
o
∑
(
,
)
=
1
≤
//
To
create
a
n associa
ti
ve
C
N
N netw
ork of
(
,
)
=
∑
∑
(
(
,
)
+
(
,
)
)
=
1
=
1
//
is t
he
la
gra
nge
m
ulti
plier.
end for
lo
op
end whil
e lo
op
END PR
OCES
S
4.
DISC
US
SI
O
N
Ex
ud
at
es,
he
m
or
rh
age
s,
an
d
m
ic
ro
ane
ur
y
sm
s
wer
e
util
iz
ed
for
scree
ning
of
DR
su
bject
s.
T
he
aff
ect
abili
ty
an
d
sp
eci
fici
ty
of
their
product
was
74.
8%
a
nd
82.
7%,
se
pa
r
at
el
y
in
separ
a
ti
ng
DR
a
nd
ty
pica
l
su
bject
s effect
ively
. Ear
ly
ide
ntific
at
ion
of
DR (near
ness of
m
ic
ro
ane
ur
y
sm
s)
was
propose
d
in v
ie
w
of ch
oice
e
m
otion
al
ly
su
pport
ive n
et
w
ork
in m
any exi
sti
ng
works.
B
ay
es o
ptim
a
li
t
y crit
eria wer
e u
ti
li
zed to
d
ist
inguis
h
m
ic
ro
ane
ur
ys
m
s.
Their
te
ch
nique
co
uld
di
sti
nguish
the
e
arly
ph
ase
of
DR
with
an
af
f
ect
abili
ty
of
10
0%
and
sp
eci
fici
ty
of
67%.
O
rd
i
nar
y
,
gen
tl
e,
dir
ect
,
extrem
e
and
pro
du
ct
ive
DR
sta
ges
is
sh
own
in
Fi
gure
2
were
conseq
ue
ntly
ch
aracte
rize
d ut
il
iz
ing
both
z
one a
nd borde
r of
t
he
R
GB p
a
r
ts of the
bloo
d vessels t
oget
he
r
with
a
fee
dforwar
d
neural
syst
em
.
I
n
the
prese
nted
syst
em
with
noisy
im
ages
we
ac
hieve
d
a
norm
al
arr
an
ge
m
ent
prof
ic
ie
ncy
ov
er
84%
an
d
af
fectabil
it
y,
sp
eci
fici
ty
wer
e
90
%
an
d
100%
ind
ivi
du
al
ly
.
We
ha
ve
al
so
util
iz
ed
exudates
a
nd
bl
ood
vessel
range
al
on
gs
ide
s
urface
par
am
eter
s
c
om
bin
ed
with
neural
sy
stem
to
orde
r
f
undus
i
m
ages.
W
e
a
c
qu
i
red
a
disco
ver
y
preci
sio
n
of
93%,
f
or
no
n
-
noisy
i
m
ages
with
aff
ect
a
bili
ty
and
sp
eci
fi
ci
ty
of
90%
an
d
100%
ind
i
viduall
y.
Fig
ur
e
7
a
nd
8
s
how
s
the
a
ver
a
ge
M
SE
an
d
t
he
fa
ct
or
s
i
nf
lue
nci
ng
th
e
recog
niti
on
pe
r
form
ances.
The
sp
eci
fici
ty
cou
ld
be
ex
pa
nd
e
d
sim
i
la
rly
as
98
.
9%
,
how
ever
this
ex
pa
ns
io
n
was
j
oi
ne
d
by
a
fall
in
af
fectabil
it
y
to
90.8%.
At
a
set
ti
ng
with
94.8%
af
fecta
bili
ty
and
52.
8%
sp
eci
fici
ty
,
no
insta
nces
of
sig
ht
unde
rm
ining
r
et
ino
pat
hy
we
r
e
m
issed.
It
ha
s
resea
rch
e
d
both
phot
ogra
phy
and
opti
c
pla
te
geog
raphy
m
et
ho
d
of
t
he
reti
nal
thick
ness
a
na
ly
zer.
Re
co
gnit
ion
of
reti
nopathy
was
accom
plished
via
aut
om
at
ed
ste
ps
evaluati
ng
wit
h
90.
5%
a
ff
ect
abili
ty
and
67.
4%
s
pecifici
ty
wh
e
n
the
pro
ba
bili
sti
c
CNN
us
e
d
to
est
im
a
t
e
no
isy
coeffic
ie
nts
as
sh
ow
n
in
F
ig
ure
9.
Our
w
ork
is
based
on
the
idea
of
the
fin
di
ng
the
to
po
l
ogy
of
the
featur
e
s
of
a
m
ic
ro
-
a
neur
is
m
in
the
gi
ven
f
unds
im
age
&
t
hen
c
od
i
ng
it
in
l
ogic
al
seq
uen
c
e
with
the
ca
scade
d
neural
netw
ork
.
Figure
7. A
verage MSE
acr
oss va
rio
us
Gr
i
d si
z
es for
R
GB
& con
ver
te
d gre
y sca
le
co
lo
r
s
pace
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Autom
ated
det
ect
ion
of m
ic
r
oane
ur
ys
ms usi
ng p
r
obabil
ist
ic
ca
sc
aded
(
Jeyapri
ya
J
)
1091
Figure
8. A
ggr
egate Pe
rfor
m
ance
of m
et
ho
d for
pr
e
-
pr
ocess
ing
on the
b
a
sis o
f perse
ve
rance
of se
ver
al
qu
al
it
ie
s of t
he
i
m
ages in
for
m
o
f
m
ean r
an
k
ac
ro
s
s
dataset
s
(a)
(b)
Figure
9. Per
f
orm
ance an
al
ysi
s of the
CPRA
MLPs (a
)
Pl
ot
Fo
r
the
ev
al
uation o
f
the
esti
m
at
ed
no
isy
coeffic
ie
nts
(b)
Estim
at
edco
ef
fici
ents from
proba
bili
sti
c CNN
a
nd th
os
e
of
act
ual
no
isy
c
oeffici
ents in
th
os
e
Im
ages
5.
CONCL
US
I
O
N
Exten
de
d
per
i
od
of
dia
betes
pr
om
pts
to
DR,
wh
e
re
the
r
et
ina
is
har
m
e
d
beca
us
e
of
f
luid
sp
il
li
ng
from
the
blo
od
vessels.
Mo
r
e
of
te
n
tha
n
th
at
,
the
ph
ase
of
DR
is
j
udge
d
in
li
gh
t
of
bl
ood
vessels,
e
xudes
,
hem
or
rh
a
ges
,
m
ic
ro
ane
ur
ys
m
s
and
su
r
fac
e.
I
n
this
pa
pe
r,
we
hav
e
e
xam
ined
disti
nctive
te
ch
niques
f
or
featur
e
s
extrac
ti
on
an
d
a
utom
at
ic
te
chn
iq
ue
s
for
ic
r
oan
e
ur
ysm
s
based
on
pro
bab
il
ist
ic
CNN.
As
of
la
te
adv
a
nces
i
n
i
m
agin
g
f
or
DR
s
creeni
ng
al
lo
w
s
top
no
tc
h
la
sti
ng
rec
ords
of
the
reti
nal
fea
tures,
w
hich
c
an
be
util
iz
ed
for
observ
i
ng
of
dia
gnos
is
a
nd
progno
sis
f
or
deter
m
ining
the
c
ou
rse
of
treat
m
ent,
an
d
w
hich
c
an
be
check
e
d
on
by
an
opht
halm
o
log
ist
,
ad
va
nce
d
i
m
ages
can
po
s
sibly
be
ha
nd
le
d
via
im
a
ge
proces
sin
g
base
d
autom
at
ed
exa
m
inati
on
syst
e
m
s.
A
portio
n
of
th
e
al
gorith
m
s
and
syst
em
s
in
vestigat
ed
in
this
pa
pe
r
ar
e
nea
r
app
li
cable
for
DR scre
eni
ng in
cl
inica
l
pr
act
ic
e.
REFERE
NCE
S
[1]
W
orld
Diabe
t
es,
A ne
ws
le
tter
fro
m
the
W
orld
H
e
al
th
Organi
z
atio
n,
4
,
1998
.
[2]
Cigna
h
ea
l
thcare
cove
r
age posit
i
on
-
A Re
por
t, 20
07.
L
ast
acce
ss
e
d
on
5th
Dec
ember
2007.
[3]
Ong,
G.
L.
,
Rip
ley
,
L
.
G.,
News
om
,
R.
S.,
Co
oper
,
M
.
,
and
C
assw
el
l,
A.
G.,
"S
cre
eni
ng
for
sight
-
thr
ea
t
eni
ng
dia
be
ti
c
re
ti
nop
at
h
y
:
compari
s
on
of
fundus
photogra
ph
y
wi
th
au
tomate
d
col
or
cont
r
ast
thre
shold
te
st",
Am.
J
.
Opht
hal
m
ol.
vo
l. 137, no.
3,
pp
.
445
–
452
,
2004.
[4]
Hunter
,
A
.
,
Lowe
ll,
J.
,
Ow
ens,
J.
,
and
Kenne
d
y
,
L,
"Q
uantifi
ca
t
i
on
of
di
abe
t
ic
r
e
ti
nopat
h
y
using
neur
al
net
works
and
sensit
ivi
t
y
a
naly
s
is",
In
Proc
ee
dings o
f Artif
i
ci
al
Neural
N
etwor
ks
in
Me
d
ici
ne
and
Bi
olog
y,
pp
.
81
-
86,
2000
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
11
, N
o.
3
,
Se
ptem
ber
2
01
8
:
1
0
8
3
–
1
0
9
3
1092
[5]
Kum
ar,
A.,
"D
iabetic
bli
ndn
ess
in
India
:
the
emergi
ng
sce
nar
io"
,
Indian
J.
Ophth
almol
,
vol
.
46,
n
o.
2,
pp
.
65
–
66,
1998.
[6]
Orbis.
La
st
a
cce
ss
ed
Dec
ember 2009.
[7]
W
at
kins,
J.
P.,
"
ABC of
dia
b
etes re
t
inopa
th
y
",
B
riti
sh Me
d
ic
al
J
ournal
326,
pp
.
924
–
926,
2003
.
[8]
Acha
r
y
a
,
U.
R
.
,
Li
m
,
C
.
M
.
,
Ng,
E
.
Y.
K
.
,
Ch
ee
,
C.
,
and
Tam
ura
,
T.,
"Com
pute
r
-
base
d
dete
ct
ion
of
dia
b
ete
s
ret
inop
at
h
y
st
ag
es
using di
g
it
a
l
f
undus i
m
age
s
",
ProcI
nstMec
h
En
g
H.
vo
l. 223, no
.
5
,
pp
.
545
–
553
,
2009
.
[9]
Shahidi
,
M
.
,
Og
ura
,
Y.
,
Blair,
N.
P.,
and
Ze
i
m
er,
R.
,
"Re
ti
n
al
thickness
change
after
foc
al
la
ser
tr
ea
tmen
t
of
dia
be
ti
c
m
ac
ul
ar
oede
m
a",
Br
J Ophthalmol
,
vo
l. 78, no.
11
,
pp
.
8
27
–
830,
1994
.
[10]
Li
,
H
.
,
and
Ch
uta
t
ape
,
O.,
"F
u
ndus
image
feat
ure
ext
r
action
"
,
Proce
ed
ings
2
2nd
Annual
EM
BS
Inte
rn
at
iona
l
Confe
renc
e, Chicago
,
pp
.
3071
-
3
073,
2000
.
[11]
Fong,
D.
S.,
Aiel
lo,
L
.
,
Gardne
r
,
T.
W
.
,
King,
G.
L.
,
Bla
nk
enship
,
G.,
Cavalle
r
an
o,
J.
D.,
Ferris,
F.
L.
,
and
Kle
in,
R.
,
"D
i
abe
t
ic re
t
i
nopat
h
y
"
,
Diab
et
es
Care
v
o
l. 26
,
no
.
1,
pp
.
226
–
229,
2003
.
[12]
Scre
eni
ng
for
Diabe
t
ic
Retino
pat
h
y
in
Europ
e
15
y
e
ars
after
the
St.
Vin
ce
nt
Dec
la
r
at
io
n.
The
Live
rpool
Dec
larati
on
200
5
.
L
ast acce
ss
ed on 20t
h
De
ce
m
b
er
2007.
[13]
W
ang,
H.,
Hs
u,
W
.
,
Goh,
K.
G.,
and
Lee,
M
.
,
"A
n
eff
ective
appr
o
ac
h
to
de
tect
le
si
ons
in
col
our
re
t
ina
l
images",
In
Proce
ed
ings o
f
t
he
IE
EE Conf
er
enc
e
on
Comput
er
Vi
sion
and
Pa
tt
ern
Re
cogn
it
io
n
,
pp
.
181
-
187
,
2000.
[14]
Acha
r
y
a
,
U.
R.
,
Li
m
,
C.
M
.
,
N
g,
E.
Y.
K
.
,
Ch
ee
,
C
.
,
and
Ta
m
ura
,
T
.
,
"Com
pute
r
base
d
d
et
e
ction
of
dia
be
te
s
ret
inop
at
h
y
st
ag
es
using di
g
it
a
l
f
undus i
m
age
"
,
J. E
ng.
Me
d
.
vo
l.
223,
no
.
H5,
pp
.
545
–
553,
2009
.
[15]
Chaudhuri
,
S.,
Chat
terje
e,
S
.
,
Katz
,
N.,
Nelso
n,
M.,
and
Gold
baum,
M
.
,
"D
e
t
ec
t
ion
of
blood
vessels
in
re
ti
n
a
l
images
using t
w
o
-
dimensional
m
at
ch
ed
fi
lters"
,
I
EE
E
Tr
ans.
Me
d
.
Imag.
vol
.
8,
no
.
3
,
pp
.
263
–
269,
1989.
[16]
Hoover,
A.
D.
,
Ko
uza
net
sova
,
V.,
and
Goldba
um
,
M
.
,
"Loc
ati
ng
blood
vessel
s
in
retina
l
ima
ges
b
y
p
iece
wis
e
thre
shold
prob
in
g
of
a
m
at
ch
ed
f
i
lt
er
response",
I
EE
E
Tr
ans.
Me
d
.
Imag.
vol
.
19
,
no.
3
,
pp
.
203
–
2
10,
2000
.
[17]
Sophara
k,
A.
,
a
nd
U
y
y
anonvara
,
B.
,
"A
utomat
ic
exuda
t
es
det
e
ction
from
dia
bet
i
c
retinopat
h
y
r
etinal
imag
e
using
fuz
z
y
C
-
m
e
ans
and
m
orphological
m
et
hods"
,
Proce
ed
ings
of
the
thi
rd
I
ASTED
i
nte
rnational
con
fe
renc
e
Adv
anc
e
s
in
Computer
S
cienc
e
and
Te
chno
logy
,
Th
ai
l
and, p
p
.
359
-
3
64
,
200
7.
[18]
Fleming,
D.
A.,
Phili
p,
S.,
Goa
t
m
an,
A.
K.,
W
il
l
ia
m
s,
J.
G.,
Olson,
A.
J.,
and
Sha
rp,
F.
P.,
"A
uto
m
at
ed
detec
t
ion
of
exuda
t
es
for
d
iabetic
re
ti
nop
at
h
y
scre
en
ing",
Phys
.
Me
d
.
B
iol.
vol
.
52,
no.
24
,
pp
.
7
385
–
7396,
2007
.
[19]
Ha
y
ashi,
J.
,
Ku
nie
da
,
T
.
,
C
ole,
J.
,
Soga,
R.
,
Ha
ta
nak
a,
Y
.
,
Lu,
M.,
Hara
,
T.,
an
d
Fujit
a
,
F.
,
“
A
deve
lopment
of
computer
-
ai
d
ed
dia
gno
sis
s
y
ste
m
using
fundu
s
i
m
age
s,”
Procee
ding
of
the
7th
I
nte
rnational
Co
nfe
renc
e
on
Vi
rt
ual
Syste
ms
and
Mu
l
ti
Me
dia
(
VSMM
2001)
,
pp
.
429
-
4
38,
2001
.
[20]
For
rac
chia,
M
.
,
Grisan,
M.
E.,
and
Ruggeri
,
A.
,
"Extr
a
ct
ion
an
d
quant
itati
v
e
desc
ription
of
vessel
fea
ture
s
i
n
h
y
per
te
nsiv
e
r
et
i
nopat
h
y
fundus
i
m
age
s"
,
Prese
nt
ed
at
CAFIA200
1
,
2001
.
[21]
Grisan,
I.
E
.
,
Pes
ce
,
A.
,
Giani,
A., Forac
chi
a
,
M.,
a
nd
Ruggeri
,
A.
,
"A
new
tra
cki
ng
sy
st
em for
the
r
obust
ext
racti
on
of
ret
ina
l
vessel
struct
ure
",
26th
Annual
Inte
rnational
Confe
ren
ce
of
the
IEEE
EMB
S
San
Franci
sco,
USA,
pp
.
1620
-
1623,
200
4.
[22]
Zha
ng,
X.
,
and
Chuta
ta
p
e,
O
.
,
"
Dete
c
ti
on
and
class
ifi
cation
of
b
right
le
sions
in
col
our
fundus
i
m
age
s"
,
Int.
Co
nf.
on
Image Proce
s
sing
vol. 1, pp
.
1
39
–
142,
2004
.
[23]
Vall
abh
a,
D.
,
Dorai
ra
j,
R
.
,
Nam
uduri,
K.
,
and
T
hom
pson,
H.,
"A
utomate
d
dete
ct
ion
and
class
ifi
cation
of
vasc
u
l
ar
abnor
m
al
ities
in
dia
betic
re
ti
no
pat
h
y
",
P
roc
ee
d
ings
of
13th
IEE
E
Signal
s,
Sys
te
ms
and
Comp
ute
rs
,
vol.
2,
pp
.
1625
-
1629,
200
4.
[24]
Cree
,
J.
M
.
,
L
eandro,
J.
J.
G.,
Soare
s,
J.
V.
B.
,
C
esa
r,
R.
M
.
Jr.,
Jeli
nek
,
H.
F.,
a
nd
Cornforth,
D.,
"Com
par
ison
o
f
var
ious
m
et
hods
to
del
ineate
bl
ood
vessels
in
r
et
in
al
images",
Proce
ed
ings
of
the
16th
Australi
an
Instit
ut
e
of
Phy
sics
Congres
s,
Canberra
,
200
5.
[25]
Kandir
aj
u
,
N.
,
Dua,
S.,
and
Th
om
pson,
H.
W
.
,
"D
esign
and
i
m
ple
m
ent
at
ion
of
a
uniqu
e
blo
od
vessel
de
te
c
t
ion
al
gorit
hm
towar
ds
ea
rl
y
d
ia
gno
sis
of
dia
bet
ic
ret
inop
a
th
y
",
Pr
oce
ed
ings
of
the
Inte
rnationa
l
Confe
renc
e
on
Information
Tec
hnology
:
Cod
ing
and
Computing
(
ITCC
’05
)
IEE
E
Computer
Soc
iety
,
pp
.
26
-
31
,
20
05.
[26]
Li
,
H
.
,
Hs
u,
W
.
,
Le
e
,
M
.
L.,
and
W
ong,
T.
Y.
,
"A
utomate
d
gr
adi
n
g
of
retina
l
vessel
c
aliber
",
IEEE
Tr
ans.
B
iomed.
Eng.
52
,
pp
.
135
2
–
1355,
2005
.
[27]
Bhui
y
an,
A.
,
N
at
h,
B
.
,
Chu
a,
J
.
,
and
Kotagi
r
i,
R.
,
"Blood
v
essel
segm
ent
a
ti
on
from
col
or
reti
nal
images
usin
g
unsupervise
d
t
ex
ture
cl
assifi
catio
n",
IE
EE
Int. Co
nf.
Image
Proc
essing,
ICIP
5
,
pp
.
521
–
524,
2007.
[28]
Os
are
h
,
A.,
M
i
rm
ehdi
,
M
.
,
Th
om
as,
B.
,
and
M
ark
ham,
R.
,
"Com
par
at
ive
ex
udat
e
c
la
ss
ifica
t
i
on
using
suppor
t
vec
tor
m
ac
h
ine
s
and
neur
al
ne
t
works
",
The
5th
Inte
rnational
C
onf.
on
Me
di
cal
Image
Computing
and
Computer
-
Assisted
Int
erv
e
nti
on,
pp
.
413
-
4
20,
2002
.
[29]
Anind
it
a
,
S.,
Ha
m
dani
,
D
y
na
,
M
.
K.,
"The
Conto
ur
Ext
racti
on
of
Cup
in
Fundus
I
m
age
s
for
Glauc
om
a
Dete
ct
ion"
,
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
.
6,
no.
6,
pp
.
2797
-
280
4,
2016
.
[30]
Hair
ol,
N.
M
.
S.,
M
ohd,
Z.
A.
R.
,
Za
l
ina,
K.,
M
o
hd,
S.
M.
A.,
Nurs
abi
ll
i
la
h
,
M.
A.,
Faizil
,
W
.
,
Tengku,
M.
M
.
T
.
A.,
"S
ign
Det
ection
Vision
Bas
ed
M
obil
e
Rob
ot
Plat
form
",
I
ndonesian
Jour
nal
of
E
lectric
a
l
Engi
ne
ering
an
d
Computer
Scien
ce
(
IJEECS)
,
vol
.
7
,
no
.
2
,
pp
.
52
4
-
532
,
2017
.
[31]
Zha
ng,
J.
Z.,
He
,
Y.
,
"A
New
M
et
hod
for
Appea
ran
ce
Qua
li
t
y
Det
ection
of
Le
ns
M
odule
Based
on
M
achine
Vision",
TEL
KOMNIKA
(
Tele
c
omm
unic
ati
on
C
omputing
Elec
tronic
s
and
Contr
ol)
Vol
.
14,
No.
2
A,
pp
.
343
-
350
,
June
2016
.
[32]
W
al
te
r,
T
.
,
M
assin,
P.,
Ergi
n
a
y
,
A.,
Ordone
z,
R.
,
Jeul
in,
C.
,
and
Klei
n
,
J.
C.
,
"A
utom
at
i
c
det
e
ct
ion
of
m
ic
roa
neur
y
sm
s in col
or
fundus
images",
Me
d
.
I
mage
Anal.
vo
l.
11,
no
.
6
,
pp
.
55
5
–
566,
2007
.
[33]
Hell
stedt,
T
.
,
a
nd
Immonen
,
I.
,
"D
isappe
ar
an
ce
and
form
at
i
on
rat
es
of
m
ic
roa
neur
y
sm
s
in
ea
rl
y
dia
b
et
i
c
ret
inop
at
h
y
",
Br.
J. Ophthal
mol
.
vol.
80
,
no
.
2
,
pp
.
135
–
139
,
1996
.
Evaluation Warning : The document was created with Spire.PDF for Python.