Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
3
,
Ma
rch
201
9
, p
p.
11
99~
1207
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
3
.pp
1199
-
1207
1199
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Retin
al b
lood v
essel
s
eg
mentati
on
from
r
etinal
i
m
age usin
g
B
-
COSF
IRE and
a
dapti
ve
t
h
re
sh
oldin
g
Az
iah A
li
1
, W
an
Mimi
Diya
na
Wan Z
ak
i
2
,
Aini Hu
ssa
i
n
3
1
,2,3
Cent
er
for
In
t
egr
ated
S
y
s
te
m
s
Engi
ne
eri
ng
and
Advanc
ed
T
ec
h
nologi
es
(INT
E
GRA
),
Facul
t
y
of Engin
ee
r
ing
&
Buil
t
E
nvironment,
Uni
ver
siti
Keba
ngsa
an
Mal
a
y
s
ia,
Ma
lay
s
ia
1
Facul
t
y
of
Com
puti
ng
&
Inform
at
i
cs,
Mult
imedi
a
Univer
si
t
y
,
Ma
lay
s
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Sep
1
5
, 201
8
Re
vised N
ov
2
1,
2018
Accepte
d Dec
2,
2018
Segm
ent
at
io
n
o
f
blood
vessels
(BVs
)
from
ret
in
al
image
is
one
of
the
important
steps
in
deve
lop
ing
a
computer
-
assisted
ret
in
al
di
agn
osis
sy
stem
and
has
bee
n
widely
rese
ar
che
d
espe
cially
for
i
m
ple
m
ent
ing
au
tomati
c
BV
segm
ent
at
ion
m
et
hods.
Thi
s
pa
per
proposes
an
improv
ement
t
o
an
exi
stin
g
ret
in
al
BV
(RB
V)
segm
ent
at
io
n
m
et
hod
b
y
c
om
bini
ng
the
tr
ai
nab
le
B
-
COS
FIRE
fil
te
r
with
ada
pti
v
e
th
resholdi
ng
m
et
hods.
The
proposed
m
et
hod
ca
n
aut
om
at
i
call
y
conf
igur
e
i
ts
sele
c
ti
vi
t
y
giv
en
a
pro
toty
pe
patter
n
to
be
det
e
ct
ed
.
Its
se
gm
ent
a
ti
on
per
f
orm
anc
e
is
compara
bl
e
to
m
any
publi
sh
ed
m
et
hods
with
the
adva
n
ta
g
e
of
robustness
aga
inst
noise
on
ret
ina
l
bac
kground.
Inst
ea
d
of
using
gr
i
d
sea
rch
to
find
the
opt
imal
thres
hold
val
ue
for
a
whole
d
ataset
,
ada
p
ti
ve
t
hre
sholding
(A
T)
is
used
to
d
et
er
m
ine
the
thre
shold
for
each
re
ti
na
l
imag
e.
Two
AT
m
et
ho
ds
inve
stigated
i
n
thi
s
st
u
d
y
were
ISO
DA
TA
and
Otsu’s
m
et
hod.
Th
e
propo
sed
m
et
hod
was
val
id
at
e
d
using
40
images
from
two
benc
hm
ark
dat
ase
ts
f
or
ret
in
al
BV
se
gm
ent
at
io
n
val
id
at
ion
,
name
l
y
DRIV
E
and
S
TAR
E.
Th
e
val
i
dat
ion
resul
ts
indi
cate
d
th
at
the
segm
ent
atio
n
per
form
anc
e
of
the
proposed
unsupervise
d
m
et
hod
is
compara
ble
to
the
origi
n
al
B
-
COS
FIRE
me
thod
and
oth
er
publi
she
d
m
et
hods,
without
req
uiri
ng
th
e
ava
ilabilit
y
of
ground
trut
h
da
ta
for
new
dat
ase
t.
T
h
e
Se
nsiti
vity
and
Speci
f
ic
i
t
y
val
u
es
ac
hi
eve
d
for
DRIV
E
and
STARE
are 0.
78
18,
0
.
9688,
0.
79
57
and
0
.
9648,
r
espe
ctively
.
Ke
yw
or
d
s
:
Ad
a
ptive
t
hr
es
ho
l
ding
B
-
COSF
IRE
f
i
lt
er
Re
ti
nal
b
lo
od
v
essel
Segm
entat
ion
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
:
Aini
Hu
s
sai
n,
Ce
nter fo
r In
te
gr
at
e
d
Syst
em
s
Enginee
rin
g
a
nd Ad
van
ce
d Tec
hnol
og
ie
s
(INT
EGR
A),
Faculty
of E
ngineerin
g & B
uilt
En
vir
onm
ent,
Un
i
ver
sit
i Ke
ba
ngsaan
Mal
ay
sia
, 43600 Ba
ngi, Ma
la
ysi
a
.
Em
a
il
:
dr
ai
ni@ukm
.ed
u.
m
y
1.
INTROD
U
CTION
It
is
pr
oject
ed
that
diabetes
cases
world
wide
will
see
a
st
eady
rise
in
the
near
fu
t
ur
e
wh
ic
h
in
tur
n
will
increase
the
cases
of
diabeti
c
reti
nopathy
(D
R)
.
DR
is
an
ocu
la
r
dis
ease
aff
ect
in
g
diabeti
c
patie
nt
s
wh
ic
h
cou
l
d
le
ad
to
total
blindness
if
the
patie
nts
are
no
t
giv
e
n
tim
el
y
treatm
ent.
T
o
dec
r
ease
the
risk
of
total
blindness
,
m
any
ho
s
pital
s
world
wide
are
no
w
m
on
it
or
in
g
diabeti
c
patie
nt
s
for
any
DR
sy
m
pto
m
s
via
routin
e
reti
nal
screeni
ng.
This
re
su
lt
s
in
a
la
rg
e
nu
m
ber
of
reti
na
l
i
m
ages
to
be
diagnose
d
by
ophth
al
m
olo
gi
sts.
A
com
pu
te
r
-
assis
te
d
reti
nal
diagnosis
syst
em
can
be
de
vel
oped
to
assist
the
op
h
t
ha
l
m
ol
og
ist
s
in
pe
rfo
rm
ing
a
m
or
e eff
ic
ie
nt
and accu
rate
re
ti
nal d
ia
gn
os
is
wh
e
n deali
ng
with lar
ge n
umber
of
reti
nal im
ages.
On
e
of
t
he
cr
ucial
ste
ps
in
a
com
pu
te
r
-
as
sist
ed
reti
nal
diag
nosis
syst
e
m
is
accurate
detect
ion
of
R
BVs
from
a
r
et
inal
i
m
age,
or
ref
e
rr
e
d
to
as
R
BV
segm
ent
at
ion
.
By
anal
ysi
ng
the
RB
V
structu
res
pro
du
c
e
d
by
segm
entat
i
on
al
gorithm
,
ophth
al
m
olo
gi
sts
can
detect
and
dia
gnos
e
a
nu
m
ber
of
oc
ular
diseases
su
c
h
as
Diabeti
c
Re
ti
no
pat
hy,
m
acul
ar
de
gen
e
rati
on
an
d
gla
uco
m
a.
Tech
niques
for
autom
at
ic
s
egm
entat
ion
of
R
BVs
hav
e
bee
n
a
m
ajor
researc
h
t
op
ic
i
n
reti
nal
i
m
a
ge
proces
sing
fiel
d.
Re
vi
ews
of
c
urre
nt
te
chn
iq
ues
f
or
this
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
En
g
&
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1199
–
1207
1200
pur
po
se
ha
ve
been
c
om
pr
ehe
ns
ively
su
m
m
a
rised
by
Fr
az
[
1]
and
rece
ntly
by
Alm
otiri
[2]
.
The
se
te
ch
niques
can
be
ge
ner
al
l
y
div
ide
d
int
o
two
m
ai
n
cat
egories,
nam
el
y
super
vised
a
nd
uns
uper
vised
m
et
ho
ds.
S
uperv
ise
d
m
et
ho
ds
re
qu
i
re
m
anu
al
ly
segm
ented
R
BV
s
i
m
age
to
be
avail
able
f
or
t
rainin
g
m
od
el
s
to
be
us
e
d
la
te
r
to
c
la
ssify
pix
el
s.
Un
s
uper
vised
m
et
ho
ds
on
t
he
ot
her
ha
nd
do
es
no
t
nee
d
to
ha
ve
a
pri
or
trai
ning
s
ession,
bu
t
us
e
ru
le
-
ba
sed
m
et
ho
ds
to
identify
vessel
pix
el
s
against
bac
kgr
ound
pi
xels.
S
om
e
of
the
a
vaila
ble
te
chn
iq
ues
incl
ud
e
ke
r
nel
-
bas
ed
m
et
ho
d
[3
]
-
[
5]
,
vessel
trac
king
m
et
h
od
[6
]
,
[
7]
,
m
at
hem
at
ic
al
m
or
phology
-
base
d
m
et
ho
d
[8
]
,
[
9]
,
m
ulti
-
scal
e
m
e
tho
d
[
10
]
-
[
12]
,
m
achine
le
ar
ning
m
et
hod
[13
]
-
[
16]
and
m
od
el
-
ba
sed
m
et
ho
d
[
17
]
,
[
18]
.
B
-
COSF
IRE,
sta
nd
s
f
or
Ba
r
-
sel
ect
ive
Co
m
bin
at
ion
of
Sh
ifte
d
Fil
te
r
Re
sp
onses
,
is
an
ef
fici
ent
m
et
ho
d
to
seg
m
ent
R
BV
fro
m
dig
it
al
fu
nd
us
im
ages
[
19]
.
T
he
res
ulti
ng
segm
ented
ve
ssel
tree
(VT)
with
B
-
COSF
IRE
f
or
bo
t
h
DRI
VE
a
nd
ST
ARE
dataset
s
dem
on
str
at
ed
the
m
et
hod’
s
capa
bili
ty
t
o
acc
ur
at
el
y
de
te
ct
m
ai
n
VT
on
f
undus
im
ages.
The
m
et
ho
d
i
s
al
so
r
obus
t
a
gainst
t
he
no
is
e
from
non
-
ve
ssel
struct
ur
e
s
on
the
reti
nal
i
m
age
ba
ckgr
ound
co
m
par
ed
to
ot
he
r
existi
ng
se
gm
entat
ion
m
eth
ods.
O
f
al
l
the
publishe
d
m
e
thods
s
o
far
,
B
-
C
OSFI
RE
m
et
ho
d
de
m
on
strat
ed
it
s
su
pe
rio
rity
in
the
eff
ic
ie
ncy
of
se
gm
enting
R
BVs
from
a
reti
nal
i
m
age
com
par
ed
to
ot
her
publishe
d
te
chn
i
qu
e
s
[19]
.
T
he
m
et
ho
d
only
ta
kes
ap
pro
xim
at
ely
six
seconds
on
aver
a
ge
t
o
pro
du
ce
the
fi
nal
segm
ented
ve
s
sel
outp
ut
im
a
ge
with
im
ages
f
ro
m
DRI
V
E
an
d
ST
ARE
.
T
his
le
nd
s
the
m
et
ho
d
well
to
real
-
tim
e
app
li
cat
i
on
s
of
B
V
seg
m
entat
ion
in
r
et
inal
i
m
age
di
agnosis
syst
em
.
The
m
et
ho
d
c
ould
al
so
be
ta
il
ore
d
to
s
uit
m
ob
il
e
com
pu
ti
ng
in
the
f
ut
ur
e
wh
ic
h
re
quires
low
c
om
pu
ta
ti
on
al
load
a
nd
shor
t
processin
g
t
i
m
e.
Wh
il
e
th
e
m
et
ho
d
has
a
relat
ively
l
arg
e
num
ber
of
par
am
et
ers
to
be
determ
ined
f
or
each
dataset
to
proces
s,
thi
s
can
be
res
ol
ved
by
usi
ng
the
pa
ram
et
er
est
i
m
ation
m
et
hod
pro
po
se
d by
V
os
ta
te
k
i
n
[
20]
.
In
this
st
ud
y
,
t
he
ef
fecti
ven
e
s
s
of
us
in
g
ada
pt
ive
thres
ho
l
din
g
with
B
-
C
O
SFI
RE
filt
er
to
bin
arize
the
gr
ay
scal
e
im
age
was
in
vestiga
te
d
.
T
hr
es
hold
value
was
or
i
gin
al
ly
determ
in
ed
usi
ng
gri
d
s
earch
by
Azz
opar
di
[19]
,
w
hich
re
qu
i
res
avail
abi
li
ty
of
groun
d
truth
im
age.
The
th
reshold
va
lue
was
syst
e
m
at
ic
ally
adj
ust
ed
an
d
op
ti
m
al
value
was
sel
ect
ed
a
s
the
value
tha
t
m
axi
m
iz
ed
a
cho
se
n
segm
entat
ion
pe
r
for
m
ance
m
et
ric
cal
le
d
MC
C fo
r
al
l t
he
i
m
age
s
in a d
at
aset
.
By
u
sing
ad
a
ptive th
re
sh
ol
ding, we can
el
i
m
inate
th
e n
eed to
em
pirical
ly
determ
ine
a
th
reshold
val
ue
f
or
ea
ch
ne
w
da
ta
set
to
proce
ss
w
hich
re
qu
i
res
groun
d
tr
ut
h
data.
T
wo
ad
aptive
thres
ho
l
ding
m
et
hods
wer
e
in
vestigat
ed
i
n
this
stu
dy,
nam
el
y
IS
ODAT
A
and
Otsu’s
m
et
hod.
T
he
f
oll
ow
i
ng
Sect
ion
2
des
c
ribes
the
bac
kgr
ound
of
B
-
C
OS
F
IRE
m
et
ho
d
a
nd
ho
w
it
perform
s
R
BV
detect
ion
on
r
et
ina
l
i
m
age
.
T
he
m
echan
ism
of
the
two
a
da
ptive
thres
holdin
g
m
et
hods
ar
e
al
so
descr
i
be
d
in
S
ect
ion
2
f
ollowe
d
by
t
he
over
view
of
t
he
pro
pos
ed
m
et
ho
d
in
com
bin
ing
a
dap
ti
ve
t
hr
e
sholdin
g
with
B
-
COS
FI
RE
m
et
hod
.
E
xperim
ental
resu
lt
s
of
t
he
pro
po
se
d
m
et
ho
d
validat
e
d
on
DR
IVE
a
nd
STA
RE
datas
et
are
prese
nted
i
n
S
ect
ion
3
a
nd t
he
c
on
cl
us
io
n
i
n
S
ect
io
n
4
.
2.
RESEA
R
CH MET
HO
D
2.1.
Ov
er
view
A
num
ber
of
pre
-
processi
ng
s
te
ps
are
a
ppli
ed
to
t
he
or
igi
na
l
color
reti
nal
i
m
age
befor
e
B
-
COSF
IRE
filt
er
is
app
li
e
d
to
obta
in
t
he
vessel
-
e
nh
a
nc
ed
gray
i
m
age
ou
t
pu
t.
In
orde
r
to
pro
duce
th
e
final
bin
a
ry
im
age
of
t
h
e
VT,
a
da
ptive
th
res
ho
l
ding
is
a
pp
li
e
d
on
B
-
CO
SF
IRE
outp
ut
im
age.
A
n
ove
rvi
ew
of
the
pro
po
s
e
d
m
et
ho
d
is
il
lus
trat
ed
in
the
F
igure
1
.
Figure
1.
O
verview
of
propos
ed
m
et
ho
d
2.2.
Pre
-
pr
ocessin
g
Gr
ee
n
c
hanne
l
i
m
age
displ
ay
s
the
best
co
ntrast
be
tween
reti
nal
vessels
a
nd
the
reti
nal
backg
rou
nd
[1]
com
par
ed
to
ei
ther
re
d
or
bl
ue
c
hannels
in
an
RI
,
th
us
it
is
extracte
d
i
n
the
first
ste
p
of
pre
-
processi
ng.
Se
condly
,
the
co
ntrast
ar
ound
the
bri
ght
reti
na
l
reg
io
n
in
th
e
Fiel
d
of
View
(FO
V)
agai
ns
t
the
dark
bac
kgr
ou
nd
is
sm
oo
the
n
us
i
ng
S
oa
res
’s
m
et
ho
d
[5]
wh
ic
h
in
vo
l
ve
s
detect
ing
the
FO
V
bounda
r
y,
and
then
dilat
ing
it
by
rep
la
ci
ng
e
ver
y
pix
el
valu
e
just
ou
tsi
de
t
he
bor
der
with
the
m
ean
val
ue
of
the
pi
xels
in
it
s
8
-
neig
hbour
ho
od
that
a
re
ins
ide
the
FOV.
This
ste
p
is
re
pe
at
ed
f
or
20
t
i
m
es
and
t
hen,
the
dilat
ed
im
age
is
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
Ret
ina
l
blood v
essel
segme
ntat
ion
fr
om reti
nal i
m
ag
e
usi
ng
B
-
CO
SF
IRE a
nd ad
a
ptive
.
...
(
Aziah Al
i)
1201
con
t
rast
-
a
dju
st
ed
us
i
ng
c
ontrast
-
li
m
it
ed
adap
ti
ve
histo
gr
a
m
equ
al
iz
at
i
on
(CLA
HE
)
al
gorithm
[21]
.
Th
e
resu
lt
in
g
im
age w
il
l be t
he
i
nput t
o
the
next
ste
p,
B
-
COS
FIR
E
filt
er.
2.3.
B
-
COSFI
RE
B
-
COSF
IRE is
a n
ovel
b
io
-
in
sp
ire
d
filt
er th
a
t can b
e
us
ed
t
o
le
arn
li
ne pat
te
rn
s fr
om
an
i
m
age sig
nal
[4
]
.
It
is
a
trai
nab
le
filt
er
in
the
sen
se
that
the
filt
er’
s
sel
ect
ivit
y
can
be
sp
eci
fied
by
prov
i
ding
a
prot
otype
patte
rn
f
or
the
fil
te
r
to
detect
.
For
a
ppli
cat
ion
i
n
RB
V
se
gm
entat
ion
,
a
ba
r
str
uctu
re
is
us
e
d
as
the
pr
oto
ty
pe
patte
rn
t
o
co
nfi
gure
the
filt
er’
s
sel
ect
ivit
y
towards
ves
sel
structu
res
wh
ic
h
are
bar
-
sh
a
pe
d.
T
he
filt
er
config
ur
at
io
n
i
nvolv
e
s
the
co
nvolu
ti
on
of
D
iffer
e
nce
of
G
a
us
sia
n
(
D
oG)
filt
er
with
the
syntheti
c
bar
patte
r
n
in
the
prototy
pe,
f
ollo
wed
by
analy
zi
ng
local
m
axi
m
um
Do
G
res
po
ns
es
al
ong
a
sp
eci
fied
nu
m
ber
of
con
ce
ntric
ci
rc
le
s.
Bl
ur
re
d
a
nd
sh
i
fted
respo
ns
es
of
the
filt
er
are
al
s
o
co
nsi
der
e
d
to
al
lo
w
f
or
s
om
e
toleran
ce
to
prefe
rred
posit
ion
of
t
he
prot
otype
pa
tt
ern
.
A
D
oG
f
un
ct
io
n
ca
n
be
re
pr
ese
nt
ed
by
D
oG(
x,y
)
in
e
qu
at
io
n 1 [
19
]
.
(
,
)
=
1
2
2
(
−
(
2
+
2
)
2
2
)
−
1
2
(
0
.
5
)
2
(
−
(
2
+
2
)
2
(
0
.
5
)
2
)
(1)
wh
e
re
(
,
)
is
the
pi
xel
posit
ion
in
an
im
age
I
,
a
nd
σ
is
t
he
sta
nd
a
r
d
de
viati
on
of
t
he
Ga
us
s
ia
n
functi
on
t
ha
t
determ
ines
the
extent
of
the
s
urrou
nd.
The
inn
e
r
Ga
us
ssia
n
fu
nctio
n
of
th
e
DoG
has
the
value
0.5
σ
use
d
as
it
’s
sta
nd
a
rd
de
via
ti
on
val
ue,
i.e
.
half
of
sta
nd
a
rd
de
viati
on
for
oute
r
Gaussi
an
f
unct
ion.
T
he
res
ponse
of
a
DoG
filt
er
de
fine
d
in
e
quat
ion
1
t
o
a
n
im
age
I
with
i
ntensity
distrib
ution
(
′
,
′
)
de
no
te
d
a
s
(
,
)
is
c
om
pu
te
d
by con
vo
l
ution:
(
,
)
=
|
∗
|
+
(2)
wh
e
re
|
.
|
+
de
note
s h
al
f
-
wa
ve rect
ific
at
ion
op
e
rat
ion
t
hat cha
nge
s all
n
e
gative
va
lues to
zer
oes.
A
ce
nter
-
on
D
oG
filt
er
is
a
ppli
ed
to
the
gi
ven
pr
oto
ty
pe
patte
rn
to
c
onf
igure
the
B
-
C
OS
F
IRE
filt
er.
A
ce
nter
point,
al
so
known
as
the
ce
nter
of
su
pp
or
t
of
the
B
-
COSF
IRE
fi
lt
er,
is
sel
ect
ed
an
d
the
D
oG
filt
er
respo
ns
es
(
,
)
al
ong
se
ver
al
k
c
oncent
ric
ci
rcle
s
ar
ound
the
c
enter
point
ar
e
co
ns
ide
red.
T
he
do
m
inant
intensit
y variat
ion
s a
rou
nd
t
he
p
at
te
rn
of
i
nterest is charact
erized
by sign
i
ficant local
m
a
xim
a p
os
it
ion
s
on
t
he
con
ce
ntric
ci
rc
le
s.
The
se
po
i
nts
i
are
d
escr
ibed
by
a
tu
ple
of
th
ree
para
m
et
ers
(
,
,
)
wh
e
re
σ_i
is
t
he
sta
nd
a
rd
de
viati
on
of
the
D
oG
filt
er
with
the
stronge
st
re
spo
ns
e,
wh
il
e
(
,
)
are
the
pola
r
co
ordi
nates
with
resp
ect
to
the
cente
r
point.
Th
ese
set
of
3
-
tu
ples
a
r
e
denoted
by
=
{
(
,
,
)
|
=
1
,
…
,
}
w
it
h
n
represe
nting
th
e
nu
m
ber
of
D
oG
respo
ns
es
consi
der
e
d.
T
he
DoG
res
ponse
s
at
the
determ
ined
posit
ions
are
us
e
d
to
cal
c
ulate
the
B
-
COS
F
IRE
outp
ut.
T
he
respo
ns
es
a
re
first
bl
urred
to
al
low
f
or
so
m
e
po
sit
ion
tole
r
ance
of
t
he
points.
Bl
ur
ri
ng
is
pe
rfor
m
ed
by
c
om
pu
ti
ng
the
m
axi
m
u
m
value
of
wei
ghte
d
th
resholde
d
DoG
respo
ns
es w
he
reas
wei
ghti
ng
is
achieve
d
by
m
ul
ti
plyi
ng
th
e
respo
ns
es o
f
D
oG f
il
te
r
with
the
c
oe
ff
ic
ie
nts
of a
Gau
s
sia
n
f
unct
ion
w
hose
sta
ndar
d
de
viati
on
σ'
is
a
li
near
functi
on
o
f
t
he
di
sta
nce
ρ_i
f
rom
the
filt
er’
s
su
pp
or
t
center
′
=
0
′
+
(3)
wh
e
re
both
0
′
an
d
α
a
re
co
ns
ta
nts.
Ne
xt,
eac
h
bl
urred
D
oG
res
pons
e
is
sh
ifte
d
by
a
distance
ρ_i
goin
g
towa
rd
s
t
he
op
po
sit
e
directi
on,
m
eet
ing
at
t
he
sup
port
ce
nt
er.
The
s
hift
ve
ct
or
s
are
Δ
=
−
cos
a
nd
Δ
=
−
sin
.
The
it
h bl
urred an
d
s
hifted
r
es
pons
e
of the
DoG fil
te
r
is
de
note
d by
,
,
(
,
)
=
ma
x
′
,
′
{
(
−
Δ
−
′
,
−
Δ
−
′
)
(
′
,
′
)
}
(4)
wh
e
re
−
3
′
<
′
,
′
<
3
′
.
T
he
outp
ut
of
a
B
-
C
OS
F
IRE
filt
er
is
t
he
n
de
fine
d
as
t
he
wei
ghte
d
ge
om
et
ric
m
ean of
al
l t
he
b
lu
rr
e
d
a
nd s
hi
fted Do
G res
ponse
s c
orres
pond
i
ng to
t
h
e set
of tu
ples
S
:
(
,
)
=
|
(
∏
(
,
,
(
,
)
)
|
|
=
1
)
1
∑
|
|
=
1
⁄
|
,
=
2
2
2
,
=
1
3
ma
x
∈
{
1
…
|
|
}
{
}
(5)
wh
e
re
|
.
|
denote
s
thres
holdin
g
the
B
-
CO
SFIR
E
filt
er
res
ponse
at
t
valu
e
0
≤
≤
1
of
t
he
m
a
xim
u
m
respo
ns
e.
At
th
is
po
int,
the
fil
te
r
is
sel
ect
ive
to
the
sh
a
pe
an
d
or
ie
ntati
on
of
the
bar
s
pecif
ie
d
in
the
prot
otype
patte
rn. T
o
c
onfig
ur
e
for
sel
ec
ti
vity
to
the sa
m
e p
at
te
rn
at
di
ff
ere
nt
or
ie
nta
ti
on
s,
t
he ori
gin
al
orienta
ti
on
of
t
he
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
En
g
&
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1199
–
1207
1202
bar
in
th
e
prot
otype
patte
r
n
can
be
r
otate
d
at
regular
i
nterv
al
s
to
ob
ta
i
n
a
ne
w
set
of
3
t
up
le
s
(
)
with
diff
e
re
nt orient
at
ion
s
ψ
:
(
)
=
{
(
,
,
+
)
|
∀
(
,
,
)
∈
S
}
(
6)
W
it
h
filt
ers’
s
el
ect
ivit
y
no
w
co
nf
i
gure
d
for
m
ulti
ple
or
ie
ntati
on
s
,
the
r
esp
on
ses
of
B
-
COS
FI
R
E
filt
ers
with
dif
fer
e
nt
ori
entat
ion
a
re
m
erg
ed
and
t
he
m
axi
m
u
m
value
at
ever
y
locat
io
n
(
x,y)
are
ta
ken
t
o
pro
du
ce
the
ov
erall
r
es
pons
e:
̂
(
,
)
=
ma
x
∈
Ψ
{
(
)
(
,
)
}
(7)
wh
e
re
Ψ
is
a
set
of
regula
r
inter
val
or
ie
ntati
on
s
de
fine
d
as
Ψ
=
{
|
0
≤
≤
}
.
T
o
det
ect
vessel
end
i
ngs,
a
ne
w
B
-
COS
FI
R
E
filt
er
with
a
prototype
that
dep
ic
ts
the
ve
ssel
end
in
gs,
i.e.
on
ly
the
u
pper
half
portio
n
of
t
he
bar
prot
otype
i
n
the
previ
ous
filt
er,
is
c
onfigure
d.
T
his
ve
ssel
en
ding
filt
er
is
re
ferre
d
to
as
asym
m
e
tric
fil
te
r
and
the
pre
vious
filt
er
as
sy
m
m
e
tric
filter.
The
fi
nal
r
esp
on
se
is
obta
ined
by
ad
di
ng
the
respo
ns
es
of
both
sym
m
et
ri
c
and asy
m
m
et
ri
c filt
ers bef
or
e
it
is rescaled
ba
ck
to
the
ori
gi
nal gray
scal
e l
evels.
In
this
w
ork,
t
he
pu
blishe
d
B
-
COS
FI
RE
para
m
et
ers
fo
r
D
RIVE
a
nd
ST
ARE
dataset
by
the
or
igi
nal
auth
or
s
in
[
19
]
wer
e
us
ed
.
The
ori
gin
al
B
-
COSF
IRE
m
et
ho
d
sp
e
ci
fied
a
thres
ho
l
di
ng
pa
ram
et
er
t
to
be
determ
ined
usi
ng
gri
d
searc
h
for
eac
h
s
pecif
ic
dataset
to
bi
nar
iz
e
t
he
ou
t
put
of
t
he
m
et
ho
d
as
per
E
qu
a
ti
on
5.
In
t
his
stu
dy,
two
AT
m
et
hods
we
re
in
vest
igate
d
to
a
utom
at
ic
ally
determ
ine
the
op
ti
m
al
thresh
ol
d
t
wh
ic
h
will
el
i
m
inate
th
e
nee
ds
f
or
(
i)
gro
und
tr
uth
data
an
d
(ii)
it
erati
ve
searc
h
for
opti
m
a
l
segm
entat
ion
thr
esh
old
for
ne
w
datase
ts.
Figure
2
s
how
sam
ples
of
reti
nal
i
m
age
s
from
DRIV
E
and
ST
ARE
database
with
their
corres
pondin
g pr
e
-
pr
ocesse
d
i
m
ages and
B
-
COSF
IRE f
il
te
r
re
spo
n
ses.
a)
b)
c)
d)
e)
f)
Figure
2
.
(a
),
(
d) Sam
ple r
et
inal i
m
ages w
it
h
thei
r
c
orrespondin
g (
b)
,
(e
) pr
e
-
pr
ocesse
d
i
m
ages and (c)
,
(f) B
-
C
OS
F
IR
E r
es
pons
e
s fo
r
DR
IVE (R
ow
1) a
nd ST
AR
E (
R
ow 2)
2.4.
Adapt
i
ve
Thr
esho
ldi
n
g (A
T
)
We
i
nv
est
igat
ed
tw
o
AT
m
et
hods
t
o
dete
rm
ine
the
opti
m
al
threshold
val
ue
for
bina
rizi
ng
th
e
B
-
COSF
IRE
m
et
ho
d’s
outp
ut.
First
is
the
Otsu’s
m
et
ho
d,
wh
ic
h
is
wide
ly
us
ed
f
or
se
gm
entat
ion
of
r
egio
n
of
interest
s
in
an
i
m
age
[22]
.
The
m
et
ho
d
a
ssu
m
es
that
e
ver
y
im
age
con
sist
s
o
f
only
two
cl
asses
of
pix
el
s
,
nam
ely
the
ba
ckgr
ound
pi
xe
ls
and
the
for
egro
und
pix
el
s
.
To
pe
rfo
rm
i
m
age
segm
entat
ion
into
the
two
cl
asses,
the
m
et
hod
ch
ooses
a
threshold
va
lue
base
d
on
the
var
ia
nce
of
the
tw
o
cl
asses.
The
oth
e
r
AT
al
gorithm
con
sidere
d
is
IS
O
DA
T
A
m
et
ho
d
[23]
.
This
m
et
ho
d
reli
es
on
the
ass
um
ption
that
the
op
tim
al
thres
ho
l
d
f
or
im
age
bin
a
rizat
ion
is
t
he
ave
r
age
of
t
he
m
ea
n
of
of
t
he
tw
o
cl
asses.
ISO
DA
T
A
is
a
n
it
erati
ve
m
et
ho
d
that
st
arts
with
us
in
g
the
m
ean
i
m
age
for
i
niti
al
i
zat
ion
t
o
cal
cu
la
te
the
init
ia
l
thres
hold,
sin
ce
th
e
m
ean
of
t
he
tw
o
cl
asses
are
not
init
ia
ll
y
kn
own
.
The
i
niti
al
thres
ho
l
d
will
then
be
us
e
d
t
o
dete
rm
ine
the
nex
t
thres
ho
l
d
valu
e
in
sea
rch
f
or
the
opti
m
a
l
value.
T
he
ste
p
is
rep
e
at
ed
unti
l
there
a
re
no
sign
i
ficant
c
hanges
i
n
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
Ret
ina
l
blood v
essel
segme
ntat
ion
fr
om reti
nal i
m
ag
e
usi
ng
B
-
CO
SF
IRE a
nd ad
a
ptive
.
...
(
Aziah Al
i)
1203
the
th
resh
ol
d
va
lue
or
unti
l
a
certai
n
nu
m
ber
of
it
erati
ons
i
s
com
plete
d.
T
hese
t
wo
m
et
h
od
s
w
ere
a
ppli
ed
t
o
the
gray
i
m
age
ou
t
pu
t
of
B
-
COSF
IRE
filt
ers
̂
(
,
)
us
ed
to
det
erm
ine
the
thr
esh
old
t
in
E
quat
ion
5.
The
final
ou
t
pu
t
is
the
bin
ary
i
m
age
res
ulti
ng
fr
om
threshol
ding
the
B
-
C
OS
F
IRE
filt
ers
res
pons
es
usi
ng
t
he
thres
ho
l
d
t
.
2.5.
Perfo
r
ma
nce
Measurem
en
t
To
quantify
t
he
segm
entat
ion
pe
rfor
m
ance,
total
num
ber
of
tr
ue
posit
ive
(TP),
false
pos
it
ive
(F
P
),
true negati
ve (
TN)
a
nd f
al
se
neg
at
ive
(FN) pixels are calc
ula
te
d
f
or
eac
h bina
ry o
ut
pu
t
i
m
age b
y com
par
i
ng
it
to
the
m
anu
al
ly
segm
ented
i
m
age.
Using
t
hese
val
ues,
a
nu
m
ber
of
m
etr
ic
s
norm
al
ly
us
e
d
f
or
se
gme
ntati
on
perform
ance
m
easur
em
ent
are
the
n
cal
c
ulate
d.
T
o
c
om
par
e
the
pe
rfor
m
ance
of
t
he
propose
d
m
et
hod
to
the
o
ri
gin
al
B
-
CO
SFI
RE
m
et
ho
d,
the
Accura
cy
(A
cc)
,
Se
ns
it
ivit
y
(S
n),
Sp
eci
fici
ty
(S
p),
an
d
Ma
tt
hew’s
Correl
at
ion
C
oe
ff
ic
ie
nt
(MCC
)
as
de
fine
d below are
calc
ulate
d:
=
+
,
=
+
,
=
+
,
=
(
/
)
−
×
√
×
×
(
1
−
)
×
(
1
−
)
(8)
w
he
re
=
+
+
+
,
=
+
an
d
=
+
.
Sens
it
ivit
y
m
e
asur
e
s
the
abil
it
y
of
the
m
et
ho
d
t
o
s
ucc
essfu
ll
y
detect
vessel
pix
el
s
w
hile
sp
eci
fici
ty
m
easur
es
the
abili
ty
to
detec
t
non
-
vess
el
pi
xels.
MC
C
is
a
m
et
ric
to
norm
al
ly
m
easur
e p
er
for
m
ance
of
bi
nary
cl
assifi
ers
in the
cases wher
e
the
tw
o
cl
ass
es
are
no
t
bala
nced,
l
ike
in
this
ca
se
with
RB
V
seg
m
entat
ion
wh
e
re
the
nu
m
ber
of
ves
sel
pix
el
s
is
fa
r
le
ss
tha
n
t
he
nu
m
ber
of
no
n
-
vess
el
pi
xels.
In
a
ddit
ion
to
t
hese
4
m
et
rics,
an
oth
e
r
m
et
ric
cal
le
d
G
-
m
ea
n
propose
d
i
n
[
24
]
is
cal
culat
ed
a
nd
def
i
ned as:
−
=
√
×
(9)
G
-
m
ean
val
ue
of
1
in
dicat
es
perfect
se
gm
entat
ion
wh
il
e
0
ind
ic
at
es
t
otall
y
wro
ng
pre
dicti
on
.
It
is
a
good c
ho
ic
e
of
m
easur
e that i
ncor
porates
both Se
ns
it
ivit
y and S
pecifici
ty
into
a
sin
gle m
e
asur
e
.
3.
RESU
LT
S
A
ND AN
ALYSIS
Fo
r
validat
io
n,
the
pr
opos
e
d
m
et
ho
d
was
use
d
to seg
m
ent
reti
nal
im
ages
from
two
publi
cl
y
avail
able
reti
nal
i
m
age
databases
,
na
m
el
y
DRIV
E
[25]
and
ST
A
RE
[26]
.
20
i
m
ages
from
t
he
te
st
set
fo
r
DRIV
E
database
a
nd
al
l
20
i
m
ages
in
STA
RE
database
wer
e
us
ed
.
Table
1
su
m
m
arise
s
the
segm
entat
ion
perform
ance o
f
the
pro
po
se
d m
et
ho
d
c
om
par
ed
to ori
gin
al
B
-
COSF
IRE
publis
hed in
[19
]
.
Table
1.
Per
for
m
ance r
es
ults
of the
pro
pose
d
B
-
C
OS
F
IRE
with
AT
c
om
par
ed
to ori
gin
al
B
-
COS
FI
RE
Databas
e
Metho
d
Acc
Sn
Sp
MCC
G
-
m
e
an
DRIVE
Origin
al B
-
COSF
I
RE
0
.94
4
2
0
.76
5
5
0
.97
0
4
0
.74
7
5
0
.86
1
9
B
-
COSFIR
E
&
Ot
su
0
.94
5
0
0
.67
4
3
0
.98
5
1
0
.73
6
2
0
.81
5
0
B
-
COSFIR
E
&
I
S
ODAT
A
0
.94
5
6
0
.69
3
9
0
.98
2
9
0
.74
1
3
0
.82
5
9
STARE
Origin
al B
-
COSF
I
RE
0
.94
6
7
0
.77
1
6
0
.97
0
1
0
.73
3
5
0
.86
5
2
B
-
COSFIR
E
&
Ot
su
0
.95
1
1
0
.70
9
3
0
.97
9
1
0
.72
6
2
0
.83
3
4
B
-
COSFIR
E
&
I
S
ODAT
A
0
.95
0
7
0
.72
4
5
0
.97
6
9
0
.72
8
4
0
.84
1
3
Com
par
ing
O
tsu
an
d
I
SOD
ATA
ac
ross
the
m
et
rics
in
Table
1,
it
seem
ed
that
I
SODA
T
A
ou
t
perform
ed
Otsu
e
xcep
t
f
or
the
S
pecifici
ty
value.
T
hu
s
,
it
is
fair
to
concl
ud
e
that
I
SODAT
A
works
bette
r
than
Otsu
in
de
te
rm
ining
t
he
thres
ho
l
d
valu
e
for
bin
a
rizi
ng
the
B
-
CO
SF
IRE
filt
e
r
ou
t
put.
It
ca
n
al
s
o
be
see
n
from
the
ta
ble
that
us
in
g
AT
res
ulted
i
n
i
nc
rease
of
A
cc
ur
acy
a
nd
S
pe
ci
fici
ty
values
for
bo
t
h
DRI
V
E
an
d
STA
RE
datab
ase.
H
ow
e
ve
r,
for
both
data
bases,
Sens
it
iv
it
y
value
wh
ic
h
re
pr
ese
nt
th
e
m
et
ho
d’s
ab
il
ity
to
detect
vessel
pi
xel
s
decr
eases
,
wh
e
n
us
i
ng
Otsu
or
ISO
D
ATA
for
thre
s
ho
l
ding.
T
his
is
becau
se
bot
h
AT
m
et
ho
ds
pro
vi
de
th
res
ho
l
d
va
lues
that
a
re
s
li
gh
tl
y
higher
than
the
opti
m
i
zed
th
res
ho
l
d
determ
ined
us
i
ng
gri
d
search
m
et
ho
d
in
or
igi
nal
B
-
COS
FI
RE.
T
his
tra
ns
la
te
s
to
le
sser
tr
ue
po
sit
ives
se
gme
nted
by
I
SODAT
A
thres
ho
l
ding
f
or
m
os
t
im
age
s,
as
il
lustrate
d
in
Fig
ure
3(c),
(g)
a
nd
(
h)
w
her
e
the
outpu
t
from
AT
m
et
ho
d
us
in
g
ISO
DAT
A
ha
ve
detect
e
d
vessels
that
are
narr
ow
e
r
in
widt
h
an
d
wi
th
le
sser
sm
all
vessels
com
pared
to
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
En
g
&
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1199
–
1207
1204
the
ou
t
put
o
bta
ined
wit
h
opti
m
iz
ed
threshol
d
in
Fig
ur
e
3(
a
),
(e
)
an
d
(
f)
,
r
especti
vely
.
H
ow
e
ve
r,
the
re
are
al
so
cases
where
I
S
ODAT
A
thre
s
ho
l
ding,
as
s
hown
in
Fig
ur
e
3(d),
detect
s
m
or
e
tr
ue
posit
iv
e
pix
el
s
com
par
ed
to
op
ti
m
iz
ed
thre
sh
ol
d
s
how
n
i
n Fi
gure
3(b).
Or
i
gi
nal B
-
C
O
SFI
RE
B
-
COSF
IRE
WI
T
H IS
ODA
TA
D
R
I
V
E
(a)
(b)
(c)
(d)
S
T
A
R
E
(e)
(f)
(g)
(h)
Figure
3
.
Sam
ple v
essel
ou
t
put im
ages (
a)
, (b
),
(e), (f)
usi
ng
or
i
gin
al
B
-
C
O
SFI
RE
an
d (c)
,
(d)
,
(g),
(h) us
ing
B
-
COSF
IRE
w
it
h
ISOD
AT
A fr
om
the D
RI
V
E d
at
a
base
(to
p row)
and
ST
ARE
database
(bottom
r
ow)
To
inc
rease
the
sensiti
vity
value
f
or
I
S
ODA
TA
th
res
ho
l
ding
m
et
ho
d,
the
t
hr
es
hold v
al
ue
determ
ined
by
the
AT
m
e
thod
was
re
du
ced
by
20%,
a
value
deter
m
ined
em
piric
al
ly
.
Qu
al
it
at
i
ve
com
par
iso
n
of
the
ou
t
p
ut
betwee
n
ori
gi
nal
ISO
DA
T
A
th
res
hold
an
d
the
reduced
ISOD
AT
A
thr
esh
old
is
sh
ow
n
in
Fi
gure
4.
I
t
can
be
see
n
t
ha
t
with
decr
ea
se
d
th
res
ho
l
d
val
ue,
m
or
e
vesse
l
pix
el
s
a
re
det
ect
ed
in
the
f
orm
of
enlar
ge
d
m
ai
n
vessels
an
d
de
te
ct
ion
of
ad
di
ti
on
al
sm
a
ll
er
vesse
ls.
T
he
i
m
pr
ovem
ent
i
s
al
so
evid
ent
in
the
quantit
at
ive
resu
lt
s
su
m
m
arised
in
Table
2.
By
us
ing
the
decr
ease
d
th
re
sh
ol
d
val
ue,
as
pr
e
dicte
d
the
sensiti
vity
valu
es
fo
r
bo
t
h
DR
IVE
a
nd
ST
ARE
dat
abases
i
ncr
eas
ed
al
on
g
with
an
inc
rease
in
G
-
m
ean
value
s
.
Wh
e
n
co
m
par
ed
t
o
the
or
i
gin
al
B
-
COSF
IRE
r
esults,
the
res
ults
ob
ta
in
ed
with
dec
rease
d
thre
shold
va
lue
achieve
d
bette
r
perform
ance
acro
ss
th
ree
pe
rfor
m
ance
m
etr
ic
s
nam
ely
Accu
racy,
Se
ns
i
ti
vity
and
G
-
m
ean.
Howe
ve
r,
the
m
et
ho
d’s
perf
or
m
ance
in
de
te
ct
ing
no
n
-
ve
ssel
pi
xels
dec
rease
sli
ghtl
y
com
par
ed
t
o
ori
gin
al
B
-
COS
FI
RE
wh
ic
h
c
ou
l
d be
due to
sli
ght i
ncr
ease
in
t
he nu
m
ber
of f
al
s
e posit
ive
pix
el
s.
a)
b)
c)
d)
e)
f)
g)
h)
Figure
4
.
Com
par
is
on of
vess
el
o
ut
pu
ts
u
si
ng
(a)
,
(
e
) IS
ODATA t
hres
ho
l
d,
(
b),
(f) r
e
duce
d ISOD
AT
A
thres
ho
l
d for
D
RIVE
(
R
ow 1)
and ST
ARE
(R
ow 2)
database
s.
(c
), (
d), (g
)
a
nd (h) are
the e
nlar
ged re
d boxes
in (
a
),
(
b)
,
(
e
)
a
nd (f)
r
es
pecti
ve
ly
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
Ret
ina
l
blood v
essel
segme
ntat
ion
fr
om reti
nal i
m
ag
e
usi
ng
B
-
CO
SF
IRE a
nd ad
a
ptive
.
...
(
Aziah Al
i)
1205
Table
2
. Per
for
m
ance r
es
ults
of the
pro
pose
d
B
-
C
OS
F
IRE
with Re
du
ce
d Thr
e
shold
(
B
-
CO
SF
IRE
+ISOD
AT
A(
R
))
c
om
par
ed
to ori
gin
al
B
-
C
O
SFI
RE
an
d oth
er
publishe
d
m
et
hods
Metho
d
DRIVE
STARE
Acc
Sn
Sp
G
Acc
Sn
Sp
G
Un
su
p
ervis
ed
Origin
al B
-
COSF
I
RE
[
1
9
]
0
.94
4
2
0
.76
5
5
0
.97
0
4
0
.86
1
9
0
.94
6
7
0
.77
1
6
0
.97
0
1
0
.86
5
2
B
-
COSFIR
E
&
I
S
ODAT
A
0
.94
5
6
0
.69
3
9
0
.98
2
9
0
.82
5
9
0
.95
0
7
0
.72
4
5
0
.97
6
9
0
.84
1
3
B
-
COSFIR
E
&
I
S
ODAT
A (
R
)
0
.94
4
6
0
.78
1
8
0
.96
8
8
0
.87
0
3
0
.94
7
1
0
.79
5
7
0
.96
4
8
0
.87
6
2
Mend
o
n
ca
[
8
]
0
.94
6
3
0
.73
4
4
0
.97
6
4
0
.84
6
8
0
.94
7
9
0
.69
9
6
0
.97
3
0
0
.82
5
1
Fraz
[
9
]
0
.94
3
0
0
.71
5
2
0
.97
6
8
0
.83
5
8
0
.94
4
2
0
.73
1
1
0
.96
8
0
0
.84
1
3
Mar
tin
ez
-
Per
ez
[
1
2
]
0
.93
4
4
0
.72
4
6
0
.96
5
5
0
.83
6
4
0
.94
1
0
0
.75
0
6
0
.95
6
9
0
.84
7
5
Su
p
ervis
ed
Asl
[
1
3
]
0
.74
2
8
0
.97
3
2
0
.85
0
2
0
.74
1
9
0
.97
0
6
0
.84
8
6
Orland
o
[
1
8
]
0
.78
9
7
0
.96
8
4
0
.87
4
5
0
.76
8
0
0
.97
3
8
0
.86
4
8
Mar
in
[
1
4
]
0
.94
5
2
0
.70
6
7
0
.98
0
1
0
.83
2
2
0
.95
2
6
0
.6
944
0
.98
1
9
0
.82
5
7
So
ares
[
5
]
0
.94
6
6
0
.73
3
2
0
.97
8
2
0
.84
6
9
0
.94
8
0
0
.72
0
7
0
.97
4
7
0
.83
8
1
Fraz
[
1
6
]
0
.94
8
0
0
.74
0
6
0
.98
0
7
0
.85
2
2
0
.95
3
4
0
.75
4
8
0
.97
6
3
0
.85
8
4
Strisciu
g
lio
[
1
5
]
0
.94
5
4
0
.77
7
7
0
.97
0
2
0
.86
8
6
0
.95
3
4
0
.80
4
6
0
.97
1
0
0
.88
3
9
Re
su
lt
o
f
the prop
os
ed
B
-
C
O
SFI
RE & I
S
O
DA
T
A
(R
)
m
eth
od (wit
h
Acc
=0
.
9446)
com
par
ed
to o
t
he
r
publishe
d
m
eth
ods
sho
w
c
om
par
able
pe
rfor
m
ance.
It
al
so
has
t
he
high
est
Sens
it
ivit
y
and
G
-
m
ean
va
lues
of
0.781
8
a
nd
0.8
703
for
DRI
V
E,
a
nd
0.7
957
and
0.8
762
f
or
STA
RE
com
pa
red
to
oth
e
r
unsupe
rv
ise
d
m
et
hods
.
Fo
r
s
uper
vised
m
et
ho
ds
,
on
l
y
Or
la
ndo’s
m
et
ho
d
f
or
D
RIVE
a
nd
Stri
sci
ug
li
o’s
m
eth
od
f
or
S
TAR
E
hav
e
sli
gh
tl
y
hig
he
r
Sens
it
ivit
y
and
G
-
m
ean
va
lues
than
t
he
pro
po
se
d
m
eth
od.
Highest
Sp
eci
fici
ty
value
f
or
DRI
VE
was
ac
hieve
d
by
B
-
C
OS
F
IRE+I
SODAT
A
m
et
ho
d
at
0.
9829
w
hile
fo
r
STA
RE
by
Ma
rin’
s
m
eth
od
at
0.981
9.
4.
CONCL
US
I
O
N
In
this
pap
e
r,
an
eff
ic
ie
nt
unsu
pe
r
vised
m
eth
od
to
se
gm
ent
RB
Vs
on
reti
nal
i
m
age
is
pr
op
os
ed
by
com
bin
ing
trai
nab
le
B
-
C
OS
F
IRE
filt
er
with
AT.
Or
i
gin
al
B
-
COSF
IRE
i
m
ple
m
entat
ion
us
ed
gri
d
sea
r
ch
to
determ
ine
the
thres
ho
l
d
val
ue
that
m
axi
m
iz
es
the
MCC
value.
T
w
o
AT
m
et
ho
ds
wer
e
in
vestig
at
ed
for
bin
a
rizi
ng
B
-
COSF
IRE
filt
er
r
esp
onses,
nam
ely
IS
O
D
ATA
a
nd
Ots
u.
Re
s
ults
in
di
cat
e
that
ISO
DA
T
A
achieve
d
bette
r
p
erfor
m
ance acro
ss m
os
t of
the conside
red
m
et
rics co
m
par
ed
to O
ts
u.
W
hile t
h
e com
bi
ned
B
-
COSF
IRE
a
nd
ISOD
AT
A
m
et
hod
ac
hieve
d
bette
r
Acc
ur
a
cy
and
S
pecifi
ci
ty
values
c
om
par
ed
to
t
he
or
i
gin
al
B
-
COSF
IRE,
Sens
it
ivit
y,
MC
C
and
G
-
m
ea
n
values
a
re
nota
bly
lo
wer.
To
im
pr
ove
th
e
overall
pe
rfo
rm
ance,
the
thre
shold
de
te
rm
ined
by
I
SODAT
A
m
eth
od
was
reduc
ed
by
20%.
T
hi
s
resu
lt
ed
i
n
si
gn
i
ficant
inc
re
ase
in
Sens
it
ivit
y
and
G
-
m
ean
valu
es,
w
hich
a
re
directl
y
relat
ed
to
t
he
abili
ty
of
the
pro
po
sed
m
et
ho
d
to
bette
r
detect
ves
sel
pi
xels.
C
om
par
ison
of
the
res
ults
obta
ine
d
t
o
oth
e
r
publis
hed
m
et
ho
ds
ind
i
cat
ed
c
om
par
able
perform
ance
t
o
both
s
up
e
r
vi
sed
an
d
uns
uper
vised
m
et
ho
ds,
with
a
ve
rag
e
processi
ng
tim
e
of
around
5
seco
nd
s
pe
r
im
age.
The
us
e
of
AT
el
im
in
at
es
the
nee
d
f
or
a
vaila
bili
ty
of
gro
und
tr
ut
h
data
for
fi
nding
t
he
op
ti
m
al
thr
esh
old
value
usi
ng
grid
sea
rc
h.
ACKN
OWLE
DGE
MENTS
This
researc
h
i
s
sup
ported
by
MO
HE
Ma
la
ysi
a
(F
RG
S/1/
2015/TK
04/U
KM/01
/
3)
an
d
UK
M
(
DI
P
-
2015
-
012).
REFERE
NCE
S
[1]
Fraz
M
.
M
.
,
et
al
.
,
“
Blood
vess
el
segm
entati
on
m
et
hodologi
es
in
re
ti
na
l
imag
e
s
–
A
surve
y
,”
Comput
Me
thod
s
Program
s B
i
omed
,
vol
.
108
,
pp
.
4
07
–
33
,
2012
.
[2]
A
lmotiri
J
.
,
et
a
l.
,
“
Re
ti
na
l
Vess
el
s
Segm
ent
atio
n
Te
chn
iq
ues
an
d
Algorit
hm
s:
A
Surve
y
,”
App
l
Sci
.
,
vol
.
8
,
pp
.
155
,
2018
.
[3]
Chaudhuri
S
.
,
e
t
al.
,
“
Detect
ion
of
blood
vessels
in
re
ti
na
l
imag
es
using
two
-
di
m
ensiona
l
m
at
c
hed
filters
,”
IE
E
E
Tr
ans Me
d
Imaging
,
vo
l.
8
,
pp.
2
63
–
9
,
1989
.
[4]
Strisci
ugli
o
N
.
a
nd
Petkov
N.
,
“
Deli
ne
at
ion
of
l
i
ne
pat
t
ern
s
in
i
m
age
s
using
B
-
COS
FIRE
fil
te
rs
,”
2017
Int.
Con
f.
Work.
B
ioi
nspir
ed
Int
el
l
.
,
IE
EE
,
p
p.
1
–
6
,
2017
.
[5]
Soare
s
J
.
V
.
B
.
,
et
al.
,
“
Ret
in
a
l
vessel
segm
ent
a
ti
on
using
the
2
-
D
Gabor
wave
l
et
and
supervise
d
cl
assificat
ion
,”
Me
d
Imaging
,
IE
EE
Tr
ans
,
vol
.
25
,
pp
.
1214
–
22
,
2006
.
[6]
Bekke
rs
E
.
,
et
al
.
,
“
A
m
ult
i
-
or
ie
n
ta
ti
on
an
aly
sis
a
pproa
ch
to
r
et
in
al
v
e
ss
el
tr
ac
k
in
g
,”
J
Ma
th
Imag
ing
Vi
s
.
,
vo
l.
49
,
pp.
583
–
610
,
20
14
.
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
En
g
&
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1199
–
1207
1206
[7]
C
huta
ta
p
e
O
.
,
et
al
.,
“
Re
ti
na
l
bl
ood
vessel
detec
ti
on
and
tra
ck
in
g
b
y
m
atc
hed
G
aussian
and
Kal
m
an
fil
te
rs
,”
Pr
oc
20th
Annu
Int
C
onf
IEEE
Eng
Me
d
Bi
o
l
Soc
V
ol20
Bi
omed
En
g
Tow
ar
Y
ear
2
000
Be
yond
(
Cat
No98CH36286)
,
vol.
6
,
pp
.
44
–
9
,
1998
.
[8]
M
endonç
a
A
.
M
.
and
Campil
ho
A.
,
“
Segm
ent
at
ion
of
re
ti
n
al
blood
v
essel
s
b
y
combining
the
d
et
e
ct
ion
of
ce
nt
erl
in
es
and
m
orphologi
ca
l
r
ec
onstruc
ti
on
,”
I
EE
E
Tr
ans Me
d
Imaging
,
vol
.
25
,
pp
.
1200
–
13
,
2
006
.
[9]
Fraz
M
.
M
.
,
et
al.
,
“
An
appr
oa
ch
to
loc
aliz
e
th
e
ret
in
al
blood
vessels
using
bit
pla
nes
and
ce
nt
erl
in
e
det
e
ct
ion
,”
Comput
Me
thod
s P
rogr
ams
Bi
omed
.
,
vol.
108
,
p
p.
600
–
16
,
2012
.
[10]
Ngu
y
en
U
.
T
.
V
.
,
et
al
.
,
“
An
eff
ec
t
ive
r
et
in
al
bl
ood
vessel
segm
ent
a
ti
on
m
et
hod
u
sing
m
ult
i
-
sca
l
e
li
n
e
detec
ti
on
,”
Pat
te
rn
Recogni
t
.
,
vol.
46
,
pp.
70
3
–
15
,
2013
.
[11]
Strisci
ugli
o
N
.
,
et
al.
,
“
Multi
sca
l
e
Blood
Vess
el
Deli
ne
at
ion
Us
i
ng
B
-
COS
FIRE
Filt
ers,
Springer
,
Cham
,”
p
p.
300
–
12
,
2015
.
[12]
Marti
ne
z
-
Pere
z
M
.
E
.
,
et
al
.
,
“
Segm
ent
ation
of
blood
vessels
from
red
-
fre
e
and
fluore
sce
in
reti
nal
imag
es
,”
Me
d
Image
Anal
,
vo
l.
11
,
pp.
47
–
61
,
2
007
.
[13]
As
l
M
.
E
.
,
et
al.
,
“
Tracki
ng
and
dia
m
eter
e
stim
at
ion
of
retina
l
vessels
usi
ng
Gauss
ia
n
proc
ess
and
Rado
n
tra
nsform
,”
J
M
e
d
Imaging
,
vol.
4
,
pp
.
1
,
2017
.
[14]
Marín
D
.
,
e
t
al.
,
“
A
New
Supervised
Method
fo
r
Blood
Ves
sel
Segm
ent
at
ion
in
Ret
inal
Im
age
s
b
y
Us
ing
Gra
y
-
Le
ve
l
and
Mom
ent
Inv
ariant
s
-
B
ase
d
Fea
ture
s
,”
I
EE
E
Tr
ans Me
d
Imaging
,
vol
.
30
,
pp
.
146
–
58
,
20
11
.
[15]
Strisci
ugli
o
N
.
,
et
al
.
,
“
Supervised
vessel
d
el
in
e
at
ion
in
re
ti
na
l
f
undus
images
with
th
e
au
tomatic
select
ion
of
B
-
COS
FIRE f
il
te
rs
,”
Ma
ch Vi
s
App
l
.,
2016.
[16]
Fraz
M
.
M
.
,
e
t
a
l
.
,
“
An
ense
m
ble
cl
assifi
cation
-
ba
sed
appr
oa
ch
ap
pli
ed
to
re
ti
na
l
b
lood
vessel
seg
m
ent
at
ion
,”
I
E
E
E
Tr
ans B
iomed
E
ng
.
,
vol.
59
,
pp.
2538
–
48
,
2012
.
[17]
Verm
ee
r
K
.
A
.,
et
al.
,
“
A
m
odel
base
d
m
et
hod
f
or
ret
ina
l
blood
vessel
det
e
ct
ion
,”
Comput
Bi
ol
Me
d
.
,
vol
.
34
,
pp.
209
–
19
,
2004
.
[18]
Orlan
do
J
.
I
.
,
et
al.
,
“
A
Discriminat
ive
l
y
T
rained
Full
y
Conn
ec
t
ed
Condit
ion
al
Random
Fiel
d
Model
for
Blood
Vess
el
Segm
en
t
at
ion
in
Fundus
I
m
age
s
,”
I
EE
E
T
rans
Bi
omed
En
g
.
,
vol
.
64
,
pp.
16
–
27
,
2017
.
[19]
Azz
opar
di
G
.
,
e
t
al.
,
“
Tra
in
able
COSFIRE
fil
te
r
s
for
ves
sel
del
i
nea
t
ion
wit
h
appl
icati
on
to
reti
nal
images
,”
Me
d
Image
Anal
.
,
vol
.
19
,
pp
.
46
–
57
,
2015
.
[20]
Vos
ta
te
k
P
.
,
e
t
al.
,
“
Perform
ance
compari
son
of
publi
cly
av
ai
l
a
ble
retina
l
bl
oo
d
vessel
segm
ent
at
ion
m
et
hods
,”
Comput
Me
d
Imaging
Gr
aph
,
20
16.
[21]
Pize
r
S
.
M
.
,
et
a
l
.
,
“
Adapti
ve
his
togra
m
equalizat
ion
and
it
s
va
riations
”
Comp
V
ision,
Gr
aph
Ima
ge
Proc
ess
,
vol
.
39
,
pp
.
355
–
68
,
1987
.
[22]
Otsu
N.
,
“
A
Thre
shold
Selecti
on
Mthd
from
Gra
y
-
L
evel
His
togr
ams
,”
IEE
E
Tr
a
ns
Syst
Man
Cybe
rn
,
vol.
C
,
pp.
62
–
6
,
1979
.
[23]
Bal
l
G
.
H
.
and
Hall
D
.
J.
,
“
ISODATA,
a
nove
l
m
et
hod
of
da
ta
a
nal
y
s
is
and
p
atte
rn
class
ifi
cation
,”
Analysis
,
pp
.
1
–
79
,
1965
.
[24]
Orlan
do
J
.
I
.
,
et
al.
,
“
Ret
in
al
b
l
ood
vessel
seg
m
ent
at
ion
in
hi
gh
reso
lut
ion
fu
ndus
photogra
p
hs
using
aut
om
a
t
ed
fea
tur
e
p
ara
m
ete
r
esti
m
ation
”
in
Brie
va
J
.
,
e
t
al
.
,
13th
Int
.
Con
f.
Me
d.
Inf. P
roc
ess.
Anal
.
,
vol
.
105
72,
pp
.
8
,
2017
.
[25]
Staa
l
J
.
,
e
t
al
.
,
“
Ridge
-
base
d
v
essel
segm
ent
ation
in
col
or
images
of
the
r
et
in
a
,”
IE
EE
Tr
ans
M
ed
Imaging
,
vol.
23
,
pp.
501
–
9
,
2004
.
[26]
Hoover
A
.
and
Goldbaum M.
,
“
Loc
a
ti
ng
th
e
opt
ic
ner
v
e
in
a
r
eti
nal
imag
e
using
the
fuz
z
y
c
onver
genc
e
of
th
e
blo
od
vessels
,”
IEEE
T
rans
Me
d
Imagi
ng
,
vol
.
22
,
pp
.
9
51
–
8
,
2003
.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Azia
h
Al
i
is
a
PhD
ca
ndidate
at
Fa
cul
t
y
of
E
ngine
er
ing
and
Buil
t
Envi
ronm
ent
,
Univer
sit
i
Keba
ngsaa
n
Ma
lay
s
ia
and
a
le
c
ture
r
at
th
e
Fac
ulty
of
Com
puting
&
Inform
at
i
c
s
,
Multi
m
edi
a
Univer
sit
y
,
Mal
a
y
si
a.
Her
rese
a
rch
intere
sts
in
c
lude
imag
e
and
signal
pro
ce
ss
i
ng
foc
using
on
biomedic
a
l
field
.
More
spec
i
fically
,
her
work
exa
m
ine
s
how
image
and
sign
al
proc
essing
te
chn
ique
s
cou
l
d
cont
ribu
te
to
wards
bet
t
er
a
nd
m
ore
eff
i
cient
wa
y
s
of
a
naly
z
ing
and
under
standi
ng
m
edi
c
al
images
a
nd
signal
s.
Apart
from
tha
t,
her
rese
arc
h
in
te
r
est
al
s
o
inc
lude
s
assistive
te
chno
l
ogie
s for
p
eople w
it
h
disab
il
i
ti
es.
W
an
Mim
i
Di
y
a
na
W
an
Za
k
i
ob
ta
in
ed
her
B
ac
h
el
or
degr
ee
(
El
e
ct
roni
cs
Engi
ne
e
ring)
in
2000,
Master
degr
ee
(
Engr.
Sc
.
)
in
20
05
and
PhD
de
gre
e
in
2012,
all
from
Multi
m
edi
a
Univ
ersi
t
y
(MM
U),
C
y
ber
j
a
y
a,
Malay
si
a.
She
is
cu
rre
nt
l
y
a
rese
ar
che
r
and
senior
l
ec
tur
er
a
t
the
Centre
of
Inte
gra
te
d
S
y
st
e
m
s
Engi
nee
ring
and
Advanc
ed
T
ec
hnolog
y
(INT
EGRA),
Univer
siti
Keba
ngsa
an
Malay
s
ia
(UK
M),
which
she
joi
ned
in
200
8.
Her
rese
a
rch
spec
ialisation
i
s
in
biome
dic
al
engi
ne
eri
ng,
an
d
her
rese
arc
h
i
nte
rests
include
int
el
l
ige
n
t
s
y
st
ems
,
image
proc
essing
and
IoT
rel
a
te
d
h
ealthcar
e
t
ec
hnolog
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
Ret
ina
l
blood v
essel
segme
ntat
ion
fr
om reti
nal i
m
ag
e
usi
ng
B
-
CO
SF
IRE a
nd ad
a
ptive
.
...
(
Aziah Al
i)
1207
A
ini
Hus
sain
o
bta
in
ed
her
B.
Sc.
in
Elec
tr
ical
Engi
ne
eri
ng
fro
m
Loui
siana
Stat
e
Univer
si
t
y
(LSU),
Bat
on
R
ouge,
US
A;
M.
Sc.
i
n
S
y
st
ems
and
Contro
l
fro
m
the
Unive
rsit
y
o
f
Manc
h
este
r
Instit
ute
of
Science
and
Techno
log
y
(UM
IST),
Manc
heste
r
,
U.
K.,
and
Ph.D.
in
El
ec
t
rical
and
El
e
ct
roni
c
Eng
i
nee
ring
from
th
e
Nati
on
al
Uni
ver
sit
y
of
Ma
lay
sia
in
1985
,
1
991
and
1997
,
respe
ctively
.
Sh
e
is
a
Profess
or
and
cur
r
entl
y
,
th
e
Cha
ir
of
the
I
NTEGRA
rese
ar
ch
c
ente
r
a
lso
known
as
“
Cen
tre
for
Inte
gr
ated
S
y
stems
Engi
nee
r
ing
and
Advanc
ed
T
ec
h
nologi
es”
Her
rese
arc
h
,
for
w
hic
h
she
has
r
ec
e
ive
d
funding
,
foc
uses
on
In
te
lligent
S
y
st
ems
and
Im
age
Proce
ss
ing.
H
er
cur
ren
t
r
ese
arc
h
int
er
ests
are
in
m
ac
hine
learni
n
g,
pat
t
ern
rec
og
nit
ion
and
vid
eo
&
image
proc
essing.
Evaluation Warning : The document was created with Spire.PDF for Python.