Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
1
3
,
No.
3
,
Ma
rch
201
9
, p
p.
962
~
9
73
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
3
.pp
9
62
-
9
73
962
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Develop
ment of
a hybrid
framew
or
k
to ch
aracteri
ze
re
d l
esions
for early
detecti
on
of
di
abetic retin
opath
y
Deep
as
hree
Deva
r
aj, Pr
asa
nna Kum
ar
S.C.
Depa
rtment
o
f
E
le
c
troni
cs
and
In
strum
ent
at
ion
E
ngine
er
ing, RVCE,
B
enga
lu
ru
-
5
9,
Indi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
ul
08, 2
018
Re
vised
N
ov
1, 20
18
Accepte
d
Nov
2
8
, 2
018
Diabe
t
ic
r
et
inop
at
h
y
(DR)
is
one
of
the
dr
ivi
ng
r
ea
sons
for
visual
def
icie
n
c
y
,
aff
ecting
people
globa
lly
.
Curre
ntly
,
the
ophth
a
lmologists
nee
d
to
inspec
t
enor
m
ous
num
b
er
of
images
w
it
h
a
spe
ci
f
ic
e
nd
goal
to
per
f
orm
m
as
s
scre
eni
ng
of
Dia
bet
i
c
re
ti
nopa
th
y.
In
thi
s
pap
er,
a
n
eff
i
ci
en
t
Com
pute
r
ai
de
d
s
y
stem
base
d
on
a
H
y
brid
fra
m
e
work
is
propose
d
for
th
e
ea
r
l
y
d
ia
gnosis
of
DR
b
y
ex
tra
c
ting
the
e
arly
DR
le
sions
such
as
m
ic
roa
ne
ur
y
sm
s
and
hemorrhage
s.
T
he
developm
ent
of
such
a
scre
e
ning
s
y
stem
wo
uld
dec
r
ea
se
the
workload
of
the
oph
tha
lmol
ogists,
as
th
e
y
now
nee
d
to
lo
ok
at
th
ose
ret
in
al
i
m
ag
es
t
hat
ar
e
an
aly
z
ed
b
y
the
s
y
s
te
m
,
as
irre
gul
ari
t
ie
s.
The
r
et
in
al
images
obta
in
e
d
from
standa
rd
ret
in
al
da
ta
ba
ses
and
Hos
pit
al
s
are
pr
e
-
proc
essed
fol
lo
wed
b
y
the
de
tecti
on
and
el
imin
at
ion
of
blood
v
essels,
opt
ic
disk
and
exuda
t
es.
Quick
propa
gat
ion
Neu
ra
l
Network
is
used
for
tra
ini
n
g
and
te
sting
of
t
he
ret
in
al
fundu
s
images
since
it
has
the
f
aste
s
t
execut
ion
ti
m
e.
L
inear
Cl
assific
a
ti
on
and
Multi
class
cl
a
ss
ifi
ca
ti
on
of
re
ti
nal
fundus
images
are
pe
rf
orm
ed
for
the
c
la
ss
ifi
c
at
ion
and
gra
ding
of
re
tinal
fun
dus
images
int
o
nor
m
al
and
abnor
m
al
using
Al
y
uda
Neuro
-
Inte
l
li
gen
ce
softwar
e
.
A pa
tient
d
at
ab
a
se
is c
r
ea
t
ed
usin
g
M
y
SQ
L
to
stor
e
th
e
req
u
ire
d
d
e
ta
il
s of
th
e
pat
i
ent
and
a
gr
aphi
c
al
user
int
e
rfa
ce
is
develop
ed
for
an
eff
ic
i
e
nt
usage
of
the
s
y
st
em.
The
exe
cu
ti
on
ti
m
e
o
f
the
s
y
stem
is
found
to
be
7
-
9
sec
onds
and
is
te
sted
on
270
ret
in
al
fundus
i
m
age
s.
The
pr
e
ci
sion
and
ac
cu
racy
of
th
e
al
gorit
hm
is 92.
5
%
and
93
.
9%
,
r
e
spec
ti
v
ely
.
Ke
yw
or
ds:
Hem
or
rh
a
ges
Local e
ntr
op
y
Thr
e
sholdi
ng
Mi
cro
ane
ury
sm
s
Mor
phology
Qu
ic
k
pro
pag
a
ti
on
Neural
Netw
ork
Copyright
©
201
9
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
:
Deep
a
shree
De
var
a
j
,
Dep
a
rtm
ent o
f El
ect
ro
nics
and
In
st
ru
m
entat
i
on E
ng
i
neer
i
ng,
RVCE, Be
ng
al
uru
-
59, I
nd
ia
.
Em
a
il
:
deep
ash
ree@
rv
ce
.edu.i
n
1.
INTROD
U
CTION
Diabetes
esse
nt
ia
lly
happen
s
wh
e
n
t
he
pa
nc
reas
fail
s
to
se
crete
suffici
ent
insu
li
n
f
or
m
et
abo
li
sm
.
It
is
a
lon
g
te
rm
conditi
on
that
causes
ver
y
hi
gh
gl
uco
s
e
le
vels.
Diab
et
ic
Re
ti
no
pat
hy
(D
R)
is
a
le
ading
ey
e
relat
ed
disorde
r
af
fected
by
di
abetes.
Si
nce,
it
is
asym
pto
m
at
ic
in
the
init
ia
l
sta
ge;
a
n
e
ffi
ci
ent
scree
ning
will
pr
e
ve
nt
blin
dness.
Va
rio
us
I
m
age
processi
ng
al
gorithm
s
are
use
d
f
or
t
he
detect
ion
of
Mi
cro
ane
ury
sm
s
and
Hem
or
rh
a
ges
,
wh
ic
h
are
t
he
early
cl
inica
l
sign
s
of
DR.
The
detect
ion
of
the
early
DR
le
si
on
s
a
nd
cl
assifi
cat
ion
will
h
el
p
i
n
c
om
bat
ing
blin
dness acr
os
s t
he worl
d.
It
is
hig
hly
im
po
rtant
for
diabeti
c
patie
nts
to
ha
ve
r
egu
la
r
ey
e
ch
eck
-
ups.
Cu
rrent
reti
nal
exam
inati
on
syst
e
m
s,
wh
ic
h
are
util
iz
ed
for
locat
ing
a
nd
re
viewi
ng
diabeti
c
reti
nopath
y,
inc
or
porate
Ophthalm
os
co
py
(which
m
ay
be
ind
irect
and
direct)
,
Fluorescei
n
a
ngiogra
ph
y,
a
nd
Fundus
photogra
phy.
These
m
et
ho
ds
of
as
sessm
e
nt
a
nd
detect
i
on
of
DR
a
re
cu
rr
e
ntly
m
a
nu
al
,
e
xp
e
ns
iv
e
an
d
re
qu
ire
s
trai
ne
d
ophth
al
m
olo
gi
sts.
The
re
is
sign
ific
a
nt
n
ee
d
of
op
hth
al
m
ologist
s
across
the
world
,
es
pecial
ly
in
the
rural
areas.
A
c
om
pu
te
r
ai
de
d
diag
no
sis
(CA
D
)
s
yst
e
m
wo
uld
ta
ckle
the
pro
ble
m
by
increasi
ng
the
m
ass
screeni
ng
of
t
he
ea
rly
D
R
patie
nts
befor
e
co
nsult
ing
the
op
hth
al
m
olo
gists.
T
his
in
tur
n
re
duces
t
he
tim
e
con
s
umpti
on
and inc
reases t
he
e
ff
ic
ie
ncy
of the
d
ia
gnos
is
of the
disease.
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
Develo
pm
e
nt
of
a
hybri
d
fr
amew
or
k to
char
ac
te
rize red
le
si
on
s
for
earl
y de
te
ct
ion
… (
De
epashree
D
ev
araj
)
963
The
reti
nal
fea
tures
are
m
acu
la
,
bl
ood
vesse
ls,
f
ov
e
a
a
nd
opti
c
disc
(
OD)
.
A
ny
var
ia
ti
ons
i
n
the
se
featur
e
s
m
ay
resu
lt
in
dif
fere
nt
reti
nal
ab
norm
al
ities.
DR
is
of
t
wo
ty
pe
s,
nam
ely
Non
proliferati
ve
DR
(N
P
DR)
an
d
prolife
rati
ve
D
R
(PDR).
Ma
ny
featur
e
s
s
uc
h
as
Mi
cr
oa
ne
ur
ysm
s
(MAs),
Hem
or
r
hag
e
s
(H
Ms
),
Ex
ud
at
es
(
hard
and
cotto
nwo
ol
sp
ots
)
an
d
I
ntrar
et
in
al
Mi
c
rovasc
ular
A
bnorm
al
i
ti
es
(I
RM
A)
exists
in
NPDR.
The
ve
ry
first
s
ign
re
la
te
d
to
e
arly
DR
are
m
i
cro
a
ne
ur
ysm
s
(MAs
)
asso
ci
at
ed
with
t
he
loc
al
exp
a
ns
io
ns
o
f
the
reti
nal
capil
la
r
ie
s
an
d
that
re
su
lt
in
hem
or
r
hag
e
s
(
H
A)
w
hen
r
uptur
ed
.
In
P
DR,
ver
y
sm
a
ll
blo
od
ve
ssels
dev
el
op fro
m
the
reti
nal surf
a
ce. T
he
reti
nal
fun
du
s
im
age w
it
h d
iffe
re
nt fea
tures
a
re as
s
how
n
in
Fig
ure
1
.
Figure
1. A
bnorm
al
Re
ti
nal Fu
nd
us
Im
age
This
pa
per
pro
po
s
es
a
CAD
syst
e
m
fo
r
the
early
detect
ion
of
DR
to
dete
ct
and
char
act
e
rise
the
red
le
sion
s
ba
sed
on
a
Hybr
i
d
te
chn
i
qu
e
cl
ubbe
d
with
neural
netw
orks,
a
G
UI
that
is
us
e
r
fr
ie
ndly
an
d
a
patie
nt
database
c
reated
us
i
ng
My
S
Q
L.
The
pa
per
is
organ
ise
d
in
t
his
fas
hion.
A
con
ci
se
st
ud
y of
the r
esea
rch
li
nk
ed
to
the
propose
d
syst
e
m
is
presented
in
Sec
ti
on
2.
A
pro
fici
ent
early
DR
pr
edict
io
n
syst
e
m
is
pr
esent
ed
in
Sect
ion
3.
T
he
com
pr
ehe
ns
iv
e
resu
lt
s
an
d
di
scussions
are
giv
e
n
in
Sect
i
on
4.
C
on
cl
us
i
on
s
a
re
pr
ese
nt
ed
in
Sect
ion
5
t
hat s
um
s u
p
t
he
e
xplore
d work.
2.
RELATE
D
W
ORK
Most
early
sc
reen
i
ng
pro
gr
a
m
s
us
e
no
n
-
m
ydriat
ic
dig
it
al
fun
du
s
cam
eras
to
capt
ur
e
the
reti
na
l
i
m
ages.
Re
ti
nal
sp
eci
al
ist
s
then
exam
ine
these
i
m
ages
fo
r
the
pr
esenc
e
of
le
sio
ns
that
are
ind
ic
at
ive
of
DR.
Howe
ver,
the
weak
li
nk
is
that
the
num
ber
of
e
xp
e
rts
ar
e
co
ns
ide
rab
ly
le
ss
an
d
e
valuati
ng
th
ou
sa
nds
of
i
m
ages
is
pr
ac
ti
cal
ly
diff
ic
ult.
Th
us
,
a
utom
at
ic
detect
io
n
via
a
ppli
cat
i
on
of
im
age
pr
oces
sin
g
te
ch
nique
s
reduces
t
he
burd
e
n
of
the
hu
m
an
ex
per
ts.M
A
a
nd
H
A
a
re
the
ea
rlie
st
sign
s
of
DR.
Bl
ood
vessels
a
nd
OD
needs to
b
e
eli
m
inate
d
f
or
suc
cessf
ul d
et
ect
io
n o
f
ea
rly
D
R
.
Bl
ood
vessels
needs
to
be
el
i
m
inate
d
from
t
he
reti
nal
im
ag
es
since
it
has
si
m
il
ar
intensity
as
that
of
m
ic
ro
ane
ur
ys
m
s.
An
nie
et
.
al
[1
]
pro
pos
ed
a
n
al
gorith
m
fo
r
reti
nal
bloo
d
vessel
segm
entat
ion,
w
hich
consi
ste
d
of
pr
e
-
proce
ssin
g
st
ages,
en
h
a
nce
m
ent
us
in
g
fas
t
discrete
c
urv
el
et
transfor
m
and
m
ulti
structur
al
el
e
m
ent
m
or
ph
ology t
o detec
t reti
nal
blood
ve
ssels. Cu
rv
el
e
t t
ran
sf
orm
w
as then ap
plied
on the
reti
nal fund
us
i
m
ages.
The
li
m
it
a
ti
on
s
incl
ude
the
te
dious
cal
culat
ion
of
the
vasc
ular
para
m
et
ers.
So
a
re
s
et
.al
[
2]
sug
ge
ste
d
a
m
e
tho
d
f
or
t
he
vessel
e
xtra
ct
ion
us
in
g
Ga
bor
wa
velet
an
d
Ba
ye
sia
n
cl
assifi
er.
The
ac
cur
acy
was
f
ound
t
o
be
96%.
T
he
m
et
ho
d
produ
ced
segm
entat
ion
by
cat
egori
zi
ng
e
ver
y
im
age
pi
xel
(as
vessel
or
nonv
essel
),
centere
d
on
pi
x
el
'
s
featur
es.
Usm
an
et
.al
[3]
su
ggest
ed
a
wav
el
et
ba
sed
m
et
ho
d
for
the
enh
a
ncem
ent
of
the
vessel
al
ong
w
it
h
the
segm
en
ta
ti
on
to
ob
ta
i
n
the
reg
i
on
of
interest
.
2
-
D
Gabo
r
wav
el
et
to
aug
m
ent
the
le
ss
visible
vessels
wer
e
use
d.
Se
nsi
ti
vity
was
f
ound
t
o
be
ar
o
und
94%
.
Sim
ilar
wor
ks
wer
e
carried
ou
t
by
[4
]
,
[
5]
and
[
6].
Be
n
et
.al
[7
]
propose
d
a
m
et
ho
d
ba
sed
on
m
or
phol
og
ic
al
scal
e
sp
ace.
Line
str
uc
turing
el
em
en
t
was
ro
ta
te
d
ar
ound
the
s
eed
point
to
obta
in
t
he
curvatu
re
of
th
e
vess
el
s.
It
w
as
pe
rfo
rm
ed
on
few
im
ag
es,
wh
ic
h
was
on
e
of
th
e
m
a
in
lim
i
ta
t
ion
.
Zh
u
et
.al
[8
]
propose
d
a
super
vised
ap
proac
h
f
or
ide
ntifyi
ng
t
he
r
e
ti
nal
vessels
base
d
on
ext
rem
e
lear
ni
ng
m
achine
f
or
pix
el
cl
assifi
cat
ion
.
A
ve
rag
e
accu
rac
y
was
a
rou
nd
96%
.
Hybr
i
d
cl
assi
fiers
c
ould
be
use
d
t
o
im
pr
ov
e
the
pe
rfor
m
ance.
Vasa
nth
i
et
.al
[9
]
ca
rr
ie
d
ou
t
sim
i
la
r
works.
Akha
van
et
.al
[1
0]
sug
gested
a
m
et
ho
d
f
or
the
ide
ntific
at
ion
of
vess
el
s
us
ing
f
uzz
y
seg
m
entat
ion
.
Th
e
m
et
ho
d sh
owe
d
c
onsist
ent p
e
rfor
m
ance for
norm
al
as w
el
l as ab
norm
al
i
m
ages.
Detect
ion
a
nd
rem
ov
al
of
O
D
is
al
so
essen
ti
al
fo
r
the
earl
y
detect
ion
of
DR.
Va
rio
us
m
et
ho
ds
have
been
pro
posed
for
the
el
im
i
nation
of
O
D.
An
a
et
al
.
[11]
pro
po
s
ed
a
te
chn
iq
ue
for
the
extracti
on
of
t
he
vasc
ular
tree
usi
ng
a
te
c
hn
i
que
base
d
on
graph
-
c
ut
ap
proa
ch.
Thi
s
in
for
m
at
ion
was
use
d
t
o
locat
e
th
e
O
D.
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.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
9
6
2
–
9
7
3
964
The
se
ns
it
ivit
y
was
f
ound
to
be
le
ss,
w
hich
was
a
draw
bac
k.
A
kyol
et
.
al
.
[12]
propose
d
a
te
chn
iq
ue
f
or
the
identific
at
ion o
f op
ti
c
disk
t
ha
t i
nclud
e
d
i
m
age
-
processi
ng,
ke
y
-
point
extracti
on,
t
extu
re
-
a
naly
sis,
vis
ua
l
-
dicti
on
ary,
a
nd
cl
assifi
er
te
ch
ni
qu
e
s.
Accuracy
of
th
e
m
et
ho
d
w
as
fou
nd
to
be
ar
ound
94%.
T
he
te
chn
iq
ue
work
e
d
well
f
or
i
m
ages
with
no
ise
and
le
sion
s.
M
ur
ti
et
.al
[13]
sugge
ste
d
a
m
et
ho
d
f
or
t
he
rec
ogniti
on
of
OD
usi
ng
le
ast
squa
re
fitt
ing
al
go
r
it
h
m
.
En
h
ancem
ent
was
done
usi
ng
histo
gr
am
equ
al
iz
at
io
n
a
nd
th
res
ho
l
ding.
Acc
ur
acy
wa
s
f
ound
t
o
be
arou
nd
97.5%. Le
ss se
ns
it
ivit
y and sp
eci
fici
ty
w
ere i
ts m
ajo
r
dra
wback.
Ra
m
a K
ri
sh
na
n
et
al. [1
4]
d
id sim
il
ar w
orks.
Mi
cro
ane
ury
sm
s
and
Haem
orrh
a
ges
we
re
detect
ed
a
ft
er
the
el
i
m
inati
on
of
th
e
Bl
oo
d
ve
ssels
a
nd
Op
ti
c d
isk.
Dif
fer
e
nt
m
et
ho
ds
wer
e
ad
opte
d
for
the d
et
ect
io
n
of
t
hese
earl
y
DR
le
sions. Pr
eey
ap
orn
et
al.
[15]
pro
po
se
d
a
m
et
hod
that
us
e
d
a
m
ix
of
HSV
te
chn
i
que,
identific
at
io
n
of
area
al
ong
with
ecce
nt
rici
ty
te
chn
iq
ue.
A
colo
r
bar
was
fra
m
ed
with
the
colo
rs
of
arb
it
r
aril
y
identifie
d
MAs.
The
pos
it
ion
of
the
M
A
wa
s
recog
nized
if
the
dot
in
the
i
m
age
exists
within
the
range
of
the
ta
r
get
colo
r.
The
acc
uracy
was
93%.
It
was
fou
nd
to
hav
e
le
ss
sensiti
vity
.
Ruk
m
ini
e
t
al.
[16]
pro
pose
d
an
oth
e
r
te
ch
nique
in
li
gh
t
of
f
ractal
exa
m
inati
on
for
the
locat
io
n
of
MAs
.
Fra
ct
al
m
easur
e
m
ent
was
figured
util
iz
ing
Box
-
Co
unti
ng
syst
e
m
.
If
the
fr
act
al
dim
ension
of
a
giv
e
n
im
age
was
lo
wer
t
ha
n
the
th
res
ho
l
d
value,
it
was
consi
der
e
d
as
norm
al
or
vice
-
ve
rsa.
The
se
ns
it
ivit
y an
d
the
s
pecifi
ci
ty
w
ere
89.5% an
d 8
2.1%, wh
ic
h
is t
he
m
ai
n
lim
it
ation
.
Ak
a
ra
Sop
har
a
k
et
al
.
[
17]
pr
opos
e
d
a
syst
em
in
view
of
m
or
phologica
l
op
e
rati
ons
for
fin
ding
M
A
.
Af
te
r
pre
-
proc
essing,
c
oar
se
segm
entat
ion
was
pe
r
form
ed
to
disti
ng
uish
the
MA.
W
e
ka
inf
or
m
at
ion
m
inin
g
pro
gr
am
m
ing
a
nd N
ai
ve
Ba
ye
sia
n
ar
rangem
ent w
e
re
util
iz
e
d.
T
he
est
im
at
i
on
s
of acc
ur
ac
y and
preci
sion were
83.34
%
a
nd
99.99
%,
res
pe
ct
ively
and
s
ensiti
vity
was
found
to
be
le
ss.
More
im
pro
ved
resu
lt
s
we
r
e
achieve
d
by
[
18
]
.
Istva
n
et
.
al
[19]
pro
pos
ed
a
syst
em
fo
r
t
he
recog
niti
on
of
m
ic
ro
a
neurysm
ta
kin
g
int
o
account
nea
r
by
ro
ta
ti
on
al
cr
os
s
-
segm
ent
prof
il
e
analy
sis.
Peak
recog
niti
on
was
done
on
the
ac
qu
i
red
cro
ss
-
sect
ion
al
prof
il
es an
d
afte
rw
a
r
ds
factual m
e
a
su
res
w
e
re c
om
pu
te
d
to
get the MA
. Nai
ve B
ay
es (
NB)
classi
fier
was
util
iz
ed.
T
he
ad
va
ntages
wer
e
high
se
nsi
ti
vity
with
low
false
po
sit
iv
e
rates.
Ma
r
wa
n
an
d
Es
wa
ran
[20]
pro
po
se
d
a
sys
tem
in
wh
ic
h
MAs
an
d
H
As
wer
e
rec
ogniz
ed
f
or
the
ea
rl
y
analy
sis
of
DR
ut
il
iz
ing
C
LAHE
and
h
-
m
ini
m
a
trans
form
.
Cl
a
ssifie
rs
wer
e
not
us
e
d,
w
hich
was
on
e
of
th
e
m
ajo
r
li
m
it
a
t
ion
.
Nila
dri
et
al
.
[21]
pro
po
se
d
a
te
c
hn
i
qu
e
f
or
the
identific
at
ion
of
MA,
w
hich
c
om
pr
ise
d
of
div
isi
on
of
the
re
ti
nal
i
m
age
into
f
our
qu
a
drants,
fo
ll
ow
e
d
by
pre
processin
g
in
al
l
the
four
qua
drants,
a
nd
la
te
r
was
co
ncaten
at
ed.
Exec
utio
n
tim
e
was one
of
t
he m
ajo
r
lim
it
a
ti
on
s.
The
survey
of
the
li
te
ratur
e
w
orks
dem
on
str
at
es
that
no
te
w
or
t
hy
work
s
ha
s
been
do
ne
in
the
fiel
d
of
identific
at
ion
of
DR.
An
y
wa
y,
a
fr
am
e
wo
r
k
with
a
n
a
utom
at
ic
and
early
DR
diag
nosis
syst
e
m
i.e.,
MA
a
nd
HA
as
early
D
R
sign
s
is
sti
ll
no
t
pro
du
ct
i
ve.
The
m
ai
n
goa
l
of
t
his
w
ork
i
s
de
velo
pm
ent
of
a
f
ram
ewo
r
k
f
or
the
early
rec
og
niti
on
o
f
DR.
I
t
exp
ect
s
t
o
disti
ng
uis
h
M
As,
H
As
f
ro
m
the
norm
al
i
m
a
ges
pr
eci
sel
y, gro
up
a
nd
gr
a
de
t
hem
util
iz
ing
the
ne
ural
net
wor
k
sy
stem
in
co
njuncti
on
with
st
or
a
ge
of
the
de
ta
il
s
in
the
pa
ti
ent
database
an
d
a
Gr
a
phic
al
use
r
inter
face
(
GUI)
f
or
t
he
prof
ic
ie
nt
a
nd
si
m
ple
acce
ss
of
the
fr
am
ework.
It
fo
c
us
es
f
unda
m
ental
ly
o
n
th
e
dev
el
opm
ent
of
a
hybri
d
f
r
a
m
ewo
r
k
that
can
eff
ect
iv
el
y
locat
e
the
early
DR
featur
e
s
with a
n
e
nh
a
nce
d
se
nsi
ti
vity
, s
pecifici
ty
, p
recisi
on
and accu
racy.
3.
MA
TE
RIA
L
S
&
METHO
D
S
Fo
r
e
ff
ect
ive
de
te
ct
ion
of
re
d
le
sion
s,
the
c
on
t
rast
betwee
n
red
le
sio
ns
a
nd
the
reti
na
l
backg
rou
nd
sh
oul
d
be
hi
gh
an
d
co
ntra
st
be
tween
t
he
reti
nal
bac
kgr
ound
a
nd
bri
ght
le
sion
s
sho
uld
be
low.
T
his
res
ults
i
n
eff
ic
ie
ntly
reducin
g
the
fals
e
po
sit
ives
duri
ng
se
gm
entat
ion
of
can
di
date
red
le
si
on
s
.
L
ocal
en
tro
py
thres
ho
l
ding
(
LET)
te
c
h
ni
que,
ta
kes
the
s
pa
ti
al
distribu
ti
on
of
gray
le
ve
ls
into
co
ns
id
erati
on
a
nd
ef
f
ic
ie
ntly
disti
nguish
e
s
e
nh
a
nce
d
dark
l
esi
on
s
an
d
the
backg
rou
nd
as
it
can
preser
ve
the
st
ru
ct
ur
al
detai
ls
of
a
n
im
age.
Com
bin
ing
the
best
feat
ur
es
of
L
ocal
e
ntropy
thres
holdi
ng
an
d
m
or
ph
olo
gy
(f
aci
li
ta
te
s
detect
ion
of
s
m
al
le
r
vessels)
,
Hy
br
i
d
m
et
ho
d
is propose
d
f
or
the early
detect
ion
of
D
R
as
s
ho
wn
i
n
Fi
gure
2.
The
C
AD
syst
e
m
fo
r
the
early
diag
no
sis
of
DR
i
s
di
vid
e
d
int
o
five
phases:
Acquisi
ti
on
of
the
Im
ages,
Pr
e
-
proces
sin
g,
P
o
st
-
processi
ng
wh
i
ch
inclu
des
re
m
ov
al
of
featu
res
li
ke
Op
ti
c
Disk
,
Bl
oo
d
V
essel
s,
Ex
ud
at
es,
Feat
ur
e
E
xt
racti
on
of MA a
nd
HA, Cla
ssific
at
ion an
d
st
or
a
ge of
p
at
ie
nt
data in
the
database
.
The
reti
nal
im
ages
are
ta
ke
n
from
m
any
retin
al
fun
dus
dat
abases
li
ke
D
I
A
RET
DB0
,
DIARET
DB1
,
and
hos
pital
da
ta
bases
of
Pr
a
bha
Ey
e
Cl
inic
and
Na
rayana
Neth
ralay
a,
Be
ng
al
uru.
The
n,
gr
ee
n
c
om
po
ne
nt
of
the
i
m
age
is
tak
en
si
nce
gree
n
col
or
plane
s
hows
the
best
vessel
co
ntrast
.
Re
d
Lesi
ons
app
ea
r
bri
ghte
r
in
the
gr
ee
n
plane
.
Op
ti
c
Disk
ha
s
featu
res
li
ke
co
ns
ta
nt
siz
e,
high
i
ntensit
y
and
ci
rc
ular
sh
a
pe.
I
niti
ally,
the
m
axi
m
u
m
value
of
the
gr
ee
n
com
po
ne
nt
of
the
i
m
age
and
the
m
edian
is
cal
culat
ed.
Th
en,
a
ci
rcu
la
r
m
ask
is
const
ru
ct
e
d
at
t
he
ce
nter
c
o
-
or
din
at
es
us
in
g (
1).
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
Develo
pm
e
nt
of
a
hybri
d
fr
amew
or
k to
char
ac
te
rize red
le
si
on
s
for
earl
y de
te
ct
ion
… (
De
epashree
D
ev
araj
)
965
Figure
2.
Bl
oc
k Diag
ram
o
f
the
Hyb
rid
m
eth
od
(a−p)
2
+ (
b−
q)
2
=
2
(1)
Wh
e
re,
r
is
the
rad
i
us
an
d
(p,
q)
a
re
the
ce
nter
co
-
or
din
at
es
.
The
ci
rc
ular
m
ask
is
then
overlai
d
on
the
reti
nal
i
m
age
to
get
the
O
D
m
asked
im
age.
Thu
s
,
O
ptic
Disk
i
s
rem
ov
ed
f
r
om
the
i
m
age.
Ba
ckgrou
nd
re
m
ov
al
m
ai
nly
aim
s
at
rem
ov
ing
the
backg
rou
nd
var
ia
ti
ons
i
n
the
im
age.
The
fore
groun
d
fe
at
ur
e
s
are
m
or
e
prom
inent.
To
p
an
d
Bott
om
Hat
Tran
s
f
or
m
is
app
li
ed
to
the
O
D
m
aske
d
i
m
age
(f
od
).
To
p
-
hat
op
e
rat
ion
is
su
bt
racti
ng
the
resu
lt
of
pe
r
f
or
m
ing
a
m
or
phologica
l
op
e
ning
on
the
O
D
m
asked
i
m
a
ge
f
od
from
th
e
O
D
m
asked
im
age (
f
od
)
u
si
ng str
uc
turing elem
ent (
SE
) wh
ic
h
is
of ty
pe
Ba
ll
as
shown i
n (2).
T
hat
(f
od
)
=
f
od
−
(f
od
◦
S
E)
(2)
Wh
e
re,
S
E is Bal
l st
ru
ct
ur
in
g
el
em
ent o
f
si
ze 5
pix
el
s a
nd ◦
is the o
pe
ni
n
g
operati
on a
nd T
hat
(f
od
)
is t
he
i
m
age
after
perform
i
ng
To
p
Hat
operati
on.
It
is
then
fo
ll
owe
d
by
Bott
om
Hat
op
e
rati
on.
B
ottom
Hat
op
e
rati
on
involves
s
ubtr
act
ing
the
res
ul
t
of
T
op
hat
operati
on
from
t
he
cl
os
i
ng
ope
rati
on
pe
rfor
m
ed
on
it
usi
ng
a
Ba
l
l
struc
tu
rin
g
el
e
m
ent (
SE
)
as
s
how
n
in
(3).
B
hat
(x
)
= (T
hat
(f
od
)
• S
E)
-
T
hat
(f
od
))
(3)
Wh
e
re,
B
hat
(
x)
is
the
resu
lt
s
of
the
Bott
om
H
at
op
erati
on,
S
E
is
Ba
ll
struct
ur
i
ng
el
em
ent
of
siz
e
5
pix
el
s
and
•
is t
he
Cl
os
i
ng
op
e
rati
on. It
is
then
c
ontrast
e
nh
a
nce
d usi
ng
(4).
I
CE
= f
od
+ T
hat
-
B
hat
(4)
Wh
e
re,
I
CE
is
the
Co
ntrast
En
han
ce
d
im
age,
T
hat
is
the
Top
Hat
Re
su
lt
an
d
B
hat
is
the
Bott
om
hat
resu
lt
.
The
con
t
rast
en
ha
nc
ed
im
age
is
t
hen
m
edian
fil
te
red
(
I
m
ed
)
an
d
is
the
n
subtr
act
ed
from
con
trast
en
ha
nce
d
im
age
for
c
om
plete
b
ackgr
ound r
em
ov
al
proces
s
usi
ng
(
5)
.
I
BN =
I
m
ed
-
I
CE
(5)
Wh
e
re
I
BN
is
t
he
bac
kgrou
nd
rem
ov
ed
im
age.
T
he
im
age
is
the
n
c
ontrast
stret
ched
to
ob
ta
in
the
f
ull
dy
nam
ic
range
of
the
re
ti
nal
i
m
age
(I
cs
).
It
is
the
n
m
e
dian
filt
ere
d
(
I
md
)
again
us
in
g
a
structu
rin
g
el
e
m
ent.
H
-
m
a
xim
a
trans
form
is
th
en
ap
plied
t
o
t
he
m
edian
filt
ered
im
age
(I
md
)
to
sup
pr
es
s
al
l
the
m
axi
m
a
in
the
inten
sit
y
i
m
age
I
md
that
is
le
ss
than
the
th
res
ho
l
d
h,
t
o
obta
in
the
im
age
f
h.
It
is
t
he
n
t
hr
es
holde
d
with
the
value
of
0.0
5
because
f
eat
ures are
not visi
bl
e if the
value
s
are a
bove 0.
05.
f
thx =
Thr
es
hold
(f
h
,
0.0
5)
(6)
Wh
e
re,
f
thx
is
the
thr
esh
old
e
d
i
m
age
and
f
h
is
the
h
-
m
axi
m
a
Tran
s
f
or
m
.
Thu
s
,
the
resu
lt
ing
im
age
con
s
ist
s
of
Bl
ood
ve
ssels,
MA
an
d
H
A.
Bl
oo
d
ve
s
sel
s
are
detect
ed
us
i
ng
a
m
od
ifie
d
m
eth
od
of
Local
entr
op
y
Thr
e
sholdi
ng
a
nd
Mo
rph
ologica
l
m
et
ho
d.
Fi
rstly
,
local
e
ntr
op
y
Th
res
ho
l
din
g
is
us
ed
f
or
the
detect
ion
of
the
blood ve
ssels,
wh
e
re t
he Op
ti
m
al
Th
res
ho
l
d i
s calc
ulate
d
as
shown i
n
t
he
f
ollow
i
ng equat
ion
.
T
E
= [
=
0
−
1
H
T
(T
h
)]
(7)
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.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
9
6
2
–
9
7
3
966
Wh
e
re,
H
T
(T
h
)
is
the
Second
Or
de
r
Ent
rop
y
and
T
E
is
the
op
ti
m
a
l
threshold
base
d
on
wh
ic
h
the
vess
el
s
are
thres
ho
l
ded
from
the
i
m
age.
Detect
ion
of
blood
vessels
(f
bv1
)
is
done
us
in
g
local
en
tro
py
Thr
es
holdin
g
m
et
ho
d.
Sec
ondly,
m
or
phol
ogic
al
detect
io
n
of
the
bloo
d
vessels
is
done
to
e
nh
a
nce
t
he
detect
ion
of
m
ino
r
vessels.
Mo
rpho
l
og
ic
al
Bl
oo
d
Vessel
detec
ti
on
i
nvolv
es
t
he
fo
ll
owin
g
s
te
ps
.
T
he
gree
n
c
om
po
ne
nt
of
the
i
m
age is ex
trac
te
d
(
f
g
). It is t
he
n
in
ve
rted
t
o ob
ta
in
b
et
te
r
e
nh
a
ncem
ent
.
f
i
= 25
5
-
f
g
(8)
Wh
e
re, f
i
is t
he
inv
e
rted
im
age an
d f
g
is t
he gree
n
c
om
po
ne
nt i
m
age.
A
fter
inv
e
rsion of th
e i
m
age,
can
ny ed
ge
detect
ion i
s p
e
r
form
ed,
which
is a m
ulti
sta
ge
al
gorithm
to
de
te
ct
v
ario
us
ed
ges.
f
c
= can
ny(
f
i
)
(9)
Wh
e
re,
f
c
in
dicat
es
Ca
nn
y
det
ect
ed
im
age
and
f
i
is
the
i
nv
e
r
te
d
im
age.
The
i
m
age
is
dilat
ed
as
well
as
er
od
e
d
.
The
er
oded
im
age
(
f
e
)
is
subt
racted
f
ro
m
the
dilat
ed
i
m
age
(f
d
)
to
detect
the
bor
der
us
ing
t
he
dis
k
-
s
ha
pe
d
structu
rin
g
el
e
m
ent o
f
r
a
diu
s
10 p
i
xels.
f
b
= f
d
-
f
e
(10)
Wh
e
re,
f
b
is
t
he
Borde
r
detec
te
d
im
age,
f
d
i
s
the
dilat
ed
i
m
age
an
d
f
e
is
the
er
oded
im
a
ge.
C
ontrast
Li
m
it
ed
Ad
a
ptive
Hist
ogram
Equ
al
iz
at
ion
(CLA
H
E)
is
then
ap
plied
on
the
gr
ee
n
com
po
ne
nt
i
m
age
(f
cl)
)
.
A
m
or
phologica
l op
e
ning
op
e
rat
ion
was
pe
rfo
r
m
ed
on
t
he
im
age u
si
ng struc
turing elem
ent
(S
E
)
of b
al
l sh
ape as
giv
e
n by (
11).
f
o
= f
cl
◦
SE
(11)
Wh
e
re,
f
o
is
the
resu
lt
of
op
e
ning
ope
rati
on,
f
cl
is
the
con
c
at
enated
CL
A
HE
im
age
and
SE
is
the
str
uc
turing
el
e
m
ent
of
dis
k
sh
a
pe.
T
he
op
e
ne
d
im
age
(f
o
)
the
n
s
ub
t
r
act
ed
from
the
CLAHE
(f
cl
)
perform
ed
i
m
a
ge
an
d
then
th
reshold
ed
(f
th
).
It
is
then
m
edian
filt
ered
for
pr
e
serv
i
ng
the
e
dges.
T
he
obta
ined
im
age
is
the
n
su
bt
racted
fro
m
the
Bord
e
r
detect
ed
im
age
(f
b
)
.
T
he
res
ultant
im
age
c
on
ta
in
s
the
de
te
ct
ed
Bl
ood
Vesse
ls
(f
bv2
).
T
he
n,
t
he
resu
lt
s
of
both
the
bl
ood
ve
ssel
detect
ion
m
et
ho
ds
a
re
c
om
bin
ed
to
obta
in
enh
a
nce
d
blo
od
vessel
detect
ion
as g
i
ven b
y
(12
).
f
f
inalbv
=
f
bv1
+f
bv
2
(12)
Wh
e
re,
f
f
inalbv
is
the
final
detect
ed
blood
vessel
,
f
bv1
is
the
blo
od
ve
ssel
detect
ed
us
in
g
local
en
tro
py
thres
ho
l
ding
and
f
bv2
is
the
bl
ood
vessel
de
te
ct
ed
us
in
g
m
orp
ho
l
og
y.
T
he
resu
lt
ant
blood
vessels
are
again
dilat
ed
usi
ng a
disk SE.
An
d f
inall
y i
t i
s su
bt
racted
from
f
thx.
f
f
inal
=
f
thx
-
f
f
inalb
v
(13)
Wh
e
re,
f
f
inalbv
is
the
detect
ed
blood
vessel
,
f
t
hx
is
the
i
m
age
con
ta
ini
ng
M
A
,
HA,
bloo
d
ve
ssels
and
f
f
inal
i
s
the
final
MA,
H
A
detect
ed
im
age.
Th
us
,
t
he
im
age
co
ntaini
ng
MA
an
d
H
A
are
detect
ed
i
n
the
i
m
age.
Th
ey
are
then
se
pa
rated
base
d
on
t
he
nu
m
ber
of
p
ix
el
s
into
MA
a
nd
H
A.
He
nce
,
after
t
he
dete
ct
ion
of
M
A
a
nd
H
A,
featur
e
s ar
e
e
xtracted
for
cl
ass
ific
at
ion
.
3.1. Fe
at
ure
Extr
act
i
on
Feat
ur
e
e
xtract
ion
helps
i
n
ef
fici
ently
rep
re
sentin
g
the
int
eresti
ng
portio
ns
of
a
n
im
age
as
featu
r
e
vecto
r
.
Feat
ures
li
ke
A
rea,
Perim
et
er,
Eccentric
it
y
(E
cc
),
Ma
jor
A
xis
Len
gth
(MAL
),
Mi
nor
A
xis
Len
gt
h
(MIL
),
E
nergy
(E),
Co
ntra
st
(C)
a
nd
H
om
og
eneit
y
(
H)
ar
e
extracte
d.
A
rea
is
the
num
ber
of
w
hite
pi
xels
existi
ng
i
n
the
reg
i
on
of
inter
est
in
the
bin
ar
y
i
m
age.
Perim
et
er
is
the
nu
m
ber
of
wh
it
e
pix
el
s
pr
ese
nt
in
th
e
bounda
ry
of
t
he
r
e
gion
of
i
nte
rest in t
he bina
ry im
age.
Eccentric
it
y i
s the m
easur
e of
dev
ia
ti
on
of th
e co
nic sect
io
n from
b
ei
ng cir
cular.
I
t i
s
giv
e
n by,
E
cc
=
√
(
1
−
m
2
n
2
)
(14)
Wh
e
re,
m
-
sem
i
-
m
ajo
r
a
xis
and
n
-
sem
i
-
m
i
nor
a
xis a
nd e i
s the ecce
ntrici
ty
.
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
Develo
pm
e
nt
of
a
hybri
d
fr
amew
or
k to
char
ac
te
rize red
le
si
on
s
for
earl
y de
te
ct
ion
… (
De
epashree
D
ev
araj
)
967
Con
tra
st i
s the
separ
at
io
n bet
ween t
he b
rig
ht
est
an
d t
he da
r
kest area
s
of
t
he
i
m
age.
It is
c
al
culat
ed
as,
C=
∑
∑
D
2
|i
, j
|
(15)
Wh
e
re,
D
|i
,
j
| i
s the
Norm
al
iz
e
d GLCM
.
Energy is t
he
s
um
o
f
the
elem
ents that a
re sq
uar
e
d
i
n
the
m
at
rix.
It is cal
c
ulate
d
as
,
E
=
∑
∑
(
−
)
2
D|i, j
|
(16)
Ho
m
og
e
neity
m
easur
es t
he p
roxim
i
ty
o
f
the
ele
m
ent’s
dist
rib
ution i
n
t
he m
at
rix.
I
t i
s
gi
ven b
y,
H
=
∑
∑
D
|
i
,
j
|
1
+
|
i
−
j
|
j
i
(17)
Along wit
h t
he
se f
eat
ures,
tot
al
n
um
ber
of
MAs a
nd HAs
are
cal
culat
ed
.
3.2.
Clas
si
ficat
i
on
Neural
Netw
ork
is
us
e
d
f
or
Cl
assifi
cat
ion
after
t
he
e
xtra
ct
ion
of
the
fe
at
ur
es.
Q
uic
k
pro
pag
at
i
on
Neural
Netw
ork
is
us
ed.
T
he
nu
m
ber
s
of
hidden
Lay
ers
us
e
d
are
two
a
nd
act
ivati
on
f
un
c
ti
on
us
e
d
is
Sigm
oid
Functi
on.
T
he
m
et
ho
d
for
Li
near
a
nd
Mult
i
cl
ass
cl
as
sific
at
ion
of
ea
rly
Diabeti
c
Re
ti
no
pat
hy
us
i
ng
Alyuda
Neur
o
I
ntell
ige
nce So
ftwar
e
a
re
discuss
e
d be
low.
Firstl
y,
the
featur
es
ext
racted
fr
om
the
Trainin
g
set
data
are
store
d
in
th
e
Excel
Sh
eet
s
in
Com
m
a
separ
at
e
d
valu
e
(CS
V)
For
m
at
.
On
e
Ex
cel
Sh
eet
c
onta
ins
the
t
otal
num
ber
of
Mi
cro
ane
ury
sm
s
and
Haem
or
rh
a
ges
al
ong
with
ty
pe
of
gradi
ng
accor
ding
t
o
T
able
1
as
s
ugge
ste
d
by
D
up
a
s
et
al
.
[
22]
.
A
no
t
he
r
Excel
Sh
eet
co
ntains
the
six
f
eat
ur
es
ext
racted
each
f
or
Mi
cro
a
ne
ur
ysm
s
and
Haem
or
rhages.
Sec
ondly
,
afte
r
the
cr
eat
io
n
of
the
Test
data
set
us
in
g
E
xce
l
sh
eet
,
Alyu
da
Neuro
I
ntell
igence
S
of
t
war
e
is
us
ed
.
T
hird
l
y,
the
te
st
data
file
(Ex
cel
Sh
eet
)
con
ta
ini
ng
t
he
num
ber
of
MAs
and
H
As
al
on
g
with
the
re
spe
ct
ive
gr
a
ding
crit
eria
sp
eci
fied
in
th
e
la
st
colu
m
n,
is
loaded
f
or
Mult
i
cl
ass
cl
a
ssific
at
ion
of
early
DR
i.e.,
N
or
m
al
,
Mild,
Mod
e
rate
and
Seve
re.
T
he
Test
data
f
il
e
con
ta
inin
g
12
featu
res
of
MA
an
d
H
A
is
loaded
f
or
Linear
cl
assif
ic
at
ion
seper
at
el
y
i.e.,
Norm
al
and
A
bnorm
al
Early
DR
Im
ages.F
inall
y,
the
data
is
analy
zed
a
nd
pr
e
-
processe
d.
It
is
then
f
ollow
e
d
by
sel
ect
ing
the
desig
n,
where
the
ty
pe
of
neural
netw
ork
is
chosen
a
s
Qu
ic
k
P
ropa
gation
Neural
Netw
or
k,
a
nd
the
nu
m
ber
of
Hidde
n
Lay
ers
as
tw
o
f
or
trai
ning.
It
is
then
te
ste
d
to
see
the
cl
assifi
er
resu
lt
s.
A
nd
th
e
te
s
t
res
ults
di
sp
la
ye
d
on
the
G
UI
is
fe
d
for
li
near
a
nd
Mu
lt
i
-
cl
ass
cl
assifi
cat
ion
res
pecti
vely
for
the
classi
fi
cat
ion
resu
lt
s t
o be
disp
la
ye
d.
Table
1.
R
ule
Ba
sed Gra
ding
of Ea
rly
D
ia
be
ti
c Ret
ino
pathy
Ear
l
y
Di
ab
etic Ret
in
o
p
ath
y
Stadiu
m
Nu
m
b
e
r
o
f
M
ic
roa
n
eu
r
y
s
m
s
Nu
m
b
e
r
o
f
Hae
m
o
rr
h
ag
es
No
r
m
al
0
0
Mild
1
≤MAs≤5
0
Mod
erate
5
≤MAs≤1
5
0
≤HAs≤5
Seriou
s
MAs≥1
5
HAs>5
3.3.
P
at
ie
n
t Data
base
Managemen
t
S
ystem
Pati
ent
Databa
se
m
anag
em
ent
syst
e
m
is
create
d
us
in
g
My
SQ
L
f
or
sto
rin
g
the
detai
ls
of
patie
nts.
It
has
recor
ds
re
la
te
d
to
Pati
en
t
ID,
Nam
e,
Ag
e
,
Weig
ht,
Gende
r,
a
nd
Date,
Pati
e
nt
History
li
ke
D
ia
betes,
Glauc
om
a,
Ca
t
aract
and
the
Count
of
Mi
cr
oan
e
ury
sm
s
an
d
Haem
or
r
hages
ob
ta
ine
d
by
the
hybr
i
d
m
e
thod.
A
Gr
a
phic
al
User
In
te
r
face
is
create
d
us
i
ng
MATLAB
R2
016a.
It
is
bu
il
t
us
ing
pus
h
butt
ons,
popu
p
m
enu
s
,
pan
el
s
,
edit
tex
t
boxes,
sta
ti
c
te
xt
bo
xes
a
nd
axes.
It
co
nsi
sts
of
disp
la
y
op
ti
ons
f
or
the
var
io
us
sta
ges
o
f
th
e
al
gorithm
fo
r
the
detect
io
n
of
M
A
a
nd
H
A
al
on
g
with
the
Exec
utio
n
Ti
m
e,
Cl
assifi
cat
ion
us
i
ng
Neural
Netw
ork
a
nd P
at
ie
nt D
at
abase
Mana
gem
ent Syst
e
m
.
4.
RESU
LT
S
AND DI
SCUS
S
ION
S
The
Re
ti
nal
f
undus
im
age
is
ob
ta
ine
d
as
s
how
n
i
n
Fig
ur
e
3(
i)
,
T
he
G
ree
n
c
om
p
on
e
nt
of
the
im
age
is
ta
ken
with
t
he
O
ptic
disk
m
aske
d
im
age
ov
erlai
d
on
it
a
s
sho
wn
i
n
Fi
gure
3(
ii
)
,
T
he
To
p
a
nd
Bott
om
Hat
Transf
or
m
s
ar
e
ap
plied
on
the
im
age
for
backg
rou
nd
re
m
ov
al
.
Firstl
y,
To
p
Hat
is
pe
rfor
m
ed
an
d
then
Bott
om
-
Hat
is
perform
ed
as
sh
ow
n
in
Fig
ure
3(
ii
i).
Ba
ck
gro
und
rem
ov
al
al
on
g
with
co
ntrast
en
han
c
e
m
ent
i
s
done
f
or
el
im
i
nating
uneve
n
var
ia
ti
ons
as
s
how
n
in
Fig
ure
3(
iv
).
T
o
inc
r
ease
the
dyna
m
ic
ran
ge
of
the
reti
nal
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.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
9
6
2
–
9
7
3
968
i
m
age,
co
ntras
t
stret
chin
g
is
done
as
s
how
n
in
Fig
ur
e
3(
v)
an
d
it
is
the
n
m
edian
filt
ere
d,
t
hus
pr
e
ser
vi
ng
al
l
the
edg
es
a
nd
reducin
g
the
s
al
t
and
pepper
no
ise
as
sho
w
n
in
Figure
3(
vi).
h
-
Ma
xim
a
is
then
app
li
ed
on
the
i
m
age
to
s
uppr
ess
the
val
ues
le
ss
tha
n
a
de
finite
intensit
y
a
s
sho
wn
in
Fig
ur
e
3(
vii).
It
is
the
n
th
res
ho
l
de
d
a
s
sh
ow
n
i
n
Fi
gure
3(viii
).
T
he
blood
Ve
ssel
i
s
detect
ed
usi
ng
L
ocal
E
ntr
opy
th
resholdi
ng
a
nd
m
or
phol
og
ic
al
m
et
ho
d
as
sho
wn
in
Fig
ure
3(i
x).
Af
te
r
the
detect
ion
of
bl
ood
ves
sel
s
by
the
hybr
id
m
e
thod,
it
is
su
btracted
from
Figu
re
3(
viii
)
to
obta
in
HA
s
a
nd
MAs
.
N
um
ber
of
pi
xels
pr
ese
nt
in
the
co
nnect
ed
reg
i
on
s
disti
ng
uish
es
HA an
d
M
A.
F
inall
y, HAs a
nd MA
s ar
e
detect
ed
as s
how
n
i
n
Fi
gures 3
(x) a
nd 3(
xi).
Figure
3. (i)
Ret
inal Fu
ndus I
m
age (
ii
) Op
ti
c
D
isk
m
ask
ov
erlai
d on G
ree
n
c
om
po
ne
nt i
m
age
(iii
)
A
fter
T
op
and Bott
om
H
a
t Transf
or
m
(
iv
)
Ba
ck
gro
und r
e
m
ov
al
(v) C
ontrast
S
t
retchin
g (v
i)
Media
n
Fil
te
red
(
vii)
h
-
Ma
xim
a Transform
(
viii
)Thr
esh
old
e
d
(i
x)
Bl
ood V
essel
de
te
ct
ion
based
on H
y
br
i
d
m
eth
od
consi
sti
ng
of l
ocal ent
ropy th
resholdi
ng and
m
or
phologica
l m
e
tho
d (
x) H
a
e
m
or
r
hag
e
s (xi
)
Mi
cr
oan
e
ur
ys
m
s
4.1.
Results
of C
l
as
sific
at
i
on usin
g Neur
al
N
etw
ork
4.1.1
Results
of Linear
C
l
as
s
ific
at
io
n
fo
r
Early De
tection
of Diab
etic
R
etino
path
y
Var
i
ou
s
feat
ures
li
ke
Ar
ea,
P
erim
e
te
r,
Ecce
ntrici
ty
,
Ma
jor
A
xis,
Mi
nor
a
xis,
C
on
t
rast
a
re
e
xtracted
from
Mi
cro
ane
ur
ysm
s
and
H
ae
m
or
r
hag
e
s
of
the
te
st
data
us
in
g
Hyb
rid
m
et
ho
d
for
Ea
rly
Detect
ion
of
DR.
The
feat
ur
es
e
xtracted
a
re
en
te
red
in
the
Mi
cro
s
oft
Excel
s
heet
an
d
the
la
st
colum
n
con
t
ai
ns
the
cl
assif
ic
at
ion
of
the
trai
ning
set
as
no
rm
al
or
a
bnor
m
al
DR
fo
r
li
nea
r
cl
assifi
cat
ion
as
sh
ow
n
in
Fig
ure
4.
It
is
the
n
loade
d
into
Alyu
da
N
eur
al
I
ntell
igen
ce
so
ft
war
e
w
her
e
t
he
ne
ural
netw
ork
is
an
al
yse
d,
pre
-
pro
cessed,
desi
gned
a
nd
trai
ned.
The
Qu
ic
k
pro
pa
ga
ti
on
ne
ur
al
network
is
c
ho
se
n,
w
hich
us
es
sigm
oid
al
activati
on
f
unct
io
n.
T
he
nu
m
ber
s
of
hi
dd
e
n
la
ye
rs
use
d
ar
e
tw
o.
It
i
s
then
te
ste
d
f
or
260
im
ages,
w
hich
a
re
cl
as
sifie
d
int
o
Nor
m
al
or
Abn
or
m
al
as sh
ow
n
in
Fig
ure
5
.
Figure
4. Feat
ures E
xtracte
d
f
or Linea
r
Cl
ass
ific
at
ion
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
Develo
pm
e
nt
of
a
hybri
d
fr
amew
or
k to
char
ac
te
rize red
le
si
on
s
for
earl
y de
te
ct
ion
… (
De
epashree
D
ev
araj
)
969
Figure
5. The
Test
I
m
age classi
fied
as
an A
bnorm
al
I
m
age af
te
r
e
nteri
ng the
featur
e
s
4.1.2. Resul
ts of M
ulti Cl
as
s
C
las
sific
ati
on f
or
Earl
y Det
ection
of
Diabeti
c R
e
tinopa
t
hy
Nu
m
ber
of
Mi
cro
a
ne
ur
ysm
s an
d Haem
or
r
ha
ges of
t
he
te
st data o
btaine
d from
the H
ybri
d
m
e
tho
d
a
re
entere
d
in
the
Mi
cro
s
of
t
E
xc
el
Sh
eet
.
They
are
us
e
d
as
th
e
Test
data
as
sh
ow
n
i
n
Fi
gure
6.
It
is
the
n
load
e
d
into
Alyu
da
N
eur
al
I
ntell
igen
ce
so
ftwa
re
w
he
re,
it
is
analy
s
ed,
pre
-
process
ed,
desi
gn
e
d
a
nd
trai
ne
d.
Once
the
syst
e
m
is
trai
ned
,
it
is
te
s
te
d
for
260
Im
ages
an
d
gr
a
de
d
i
nto
norm
al
,
m
il
d
,
m
od
erate
a
nd
seve
re
as
s
hown
i
n
Figure
7.
Figure
6. Feat
ures e
xtracted
for M
ulti
class
cl
assifi
cat
ion
/R
ule b
ase
d g
ra
ding
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.
1
3
, N
o.
3
,
Ma
rc
h
201
9
:
9
6
2
–
9
7
3
970
Figure
7
.
Re
su
l
ts of Mult
i cl
ass cla
ssific
at
ion/
Rule base
d Gra
ding
4.2.
Results
of P
at
ie
n
t Dat
abase
M
ana
ge
ment Sys
tem
Pati
ent
Detai
ls
li
ke
Pati
ent
ID
,
Nam
e,
Ag
e
,
W
ei
ght,
Ge
nder
,
Date,
Pati
ent
History
li
ke
Diabetes,
Glauc
om
a,
Ca
t
aract
an
d
C
ount
of
Mi
cr
oaneur
ysm
s
and
Haem
or
rh
a
ges
obta
ined
by
t
he
Hybr
i
d
m
eth
od
a
r
e
store
d
in
the
p
at
ie
nt
database
create
d
us
in
g
MYSQL
5.5
a
s
sh
ow
n
in
Fig
ur
e
8.
Data
bas
e
create
d
in
My
SQ
L
was nam
ed
as ‘pat
ie
nt’
a
nd ta
ble create
d wit
h
al
l t
he rec
ord
s w
as
n
am
ed
a
s ‘Deta
il
s’.
Figure
8. Early
D
R Pat
ie
nt D
a
ta
base cr
e
at
ed usin
g
My
S
QL
4.3.
Results
of Gr
aphi
cal
U
ser Inte
rf
ace c
reat
e
d for E
ar
ly D
e
tecti
on
of
Di
ab
e
tic
Re
t
inop
at
h
y
A
G
raphical
User
In
te
rf
ace
is
create
d
usi
ng
M
ATL
AB
R201
6a.
It
c
onsist
s
of
dis
pl
ay
op
ti
ons
f
or
three
al
gorith
m
s
fo
r
the
dete
ct
ion
of
MA
and
HA
al
on
g
with
the
Exec
ut
ion
Tim
e,
C
lassificat
ion
us
i
ng
Ne
ur
a
l
Netw
ork
an
d
Pati
ent
Databa
se
Ma
nag
em
e
nt
Sy
stem
.
The
featur
es
ext
r
act
ed
for
the
hybri
d
al
gorithm
are
disp
la
ye
d
i
n
th
e
GUI.
T
he
pat
ie
nt
detai
ls
can
be
entere
d
in
t
he
G
UI,
w
hich
is
stored
i
n
th
e
database
.
Fig
ur
e
9
sh
ows
the
GUI
o
f
the
Early
DR
f
or a
reti
nal
fun
du
s
I
m
age w
hic
h
is
gr
a
de
d
as m
od
e
rate.
I
t i
s obse
rv
e
d
t
hat fo
r
the r
et
inal i
m
a
ge
loa
de
d,
t
he nu
m
ber
of MA
’s
a
re
7
a
nd the
num
ber
of
HA’s
a
re
3.
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
Develo
pm
e
nt
of
a
hybri
d
fr
amew
or
k to
char
ac
te
rize red
le
si
on
s
for
earl
y de
te
ct
ion
… (
De
epashree
D
ev
araj
)
971
Figure
9. G
raphical
U
se
r In
te
rf
ace
of the
Re
ti
nal
Fundus
Im
age grad
e
d
a
s m
od
erate
Table
2
s
how
s
the
perf
or
m
ance
analy
sis
of
the
hybri
d
appro
ac
h
a
nd
the
res
ults
for
dif
fer
e
nt
databases
usi
ng th
ree
dif
fer
e
nt alg
or
it
hm
s ar
e
pr
ese
nted
in Fi
gure
10.
Figure
10. Per
f
or
m
ance Cha
rts of the
th
ree
di
ff
ere
nt alg
ori
thm
s f
or d
if
fer
e
nt d
at
a
bases
The
res
ults
of
the
work
were
com
par
ed
w
it
h
the
groun
d
-
trut
hs
avail
ab
le
fo
r
eac
h
i
m
age
of
th
e
DIARET
DB
and
ME
SSID
O
R
databases
a
nd
validat
io
n
w
as
carried
out
for
the
hos
pital
i
m
ages
by
a
Vitreo
-
re
ti
nal
sp
eci
al
ist
.
The
outc
om
es
sh
ow
s
tha
t
the
values
of
accuracy
an
d
pr
eci
sio
n
ha
ve
increased
us
i
ng
th
e
hybri
d
a
ppr
oa
ch
a
nd
f
or
eac
h
fun
du
s
im
age
the
e
xec
utio
n
ti
m
e
is
on
ly
7
-
9s
w
hich
is
substanti
al
.
T
able
3
il
lustrate
s
the
com
par
ison
of
the
3
m
et
ho
dolo
gie
s
f
or
th
e
Detect
ion
of
red
le
sio
ns
.
In
the
Mor
phologica
l
m
et
ho
d,
s
om
e
of
the
uniq
ue
featu
res
i
den
t
ifie
d
a
re
the
di
vision
of
a
n
i
m
age
into
f
our
qu
a
dr
a
nts
be
for
e
perform
ing
the
pr
e
-
processin
g
al
on
g
with
the
m
asking
of
the
opti
c
dis
k
us
in
g
t
he
cen
troid
m
et
ho
d.
In
t
he
Entr
op
y
base
d
m
e
tho
d,
s
om
e
of
the
uniq
ue
featur
e
s
ide
ntif
ie
d
are
local
e
ntr
op
y
t
hr
es
ho
lding
pe
rfo
rm
e
d
for
detect
ing
an
d
el
i
m
inati
ng
t
he
bloo
d
vessel
s,
al
on
g
with
the
act
ive
c
on
tour
m
et
ho
d
f
or
the
detect
io
n
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.