Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
4, N
o
. 4
,
A
ugu
st
2014
, pp
. 56
1
~
57
2
I
S
SN
: 208
8-8
7
0
8
5
61
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
A Novel Retin
a
l Blood Vess
el
Segmentation Algorithm using
Fuzzy s
e
gment
a
ti
on
Raz
i
eh Akhavan*,
Karim
F
a
ez
**
* Departm
e
nt
of
Com
puter Engin
eering
,
Qa
zvin
B
r
anch,
Is
lam
i
c
A
zad Univ
ers
i
t
y
,
Qazvin,
Iran
** Electr
i
cal
En
gineer
ing Dep
a
rtment, Amirkabi
r
University
of Technolog
y
,
Tehr
an,
I
r
an
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
May 12, 2014
Rev
i
sed
Jun
23,
201
4
Accepte
d J
u
l
8, 2014
Assessment of blood vessels in r
e
tin
al im
ages is
an important f
a
ctor for man
y
m
e
dical
dis
o
rder
s
.
The
chang
e
s
i
n
the r
e
tin
al
ves
s
els
due to
the
p
a
thologi
es
can be
eas
i
l
y
id
entifi
e
d b
y
s
e
g
m
enting the r
e
t
i
n
al ves
s
e
ls
.
S
e
g
m
entation of
retin
al vessels is done to identif
y
th
e early
d
i
agnosis of the
disease lik
e
glaucoma, d
i
abetic retinop
ath
y
, macular
degener
a
tion
,
h
y
per
t
ensiv
e
retinop
ath
y
and
arter
i
osclerosis
. In this
paper, we propose an auto
matic bloo
d
vessel segmentation
method.
The pr
oposed
algorithm starts
with
the
extraction o
f
b
l
ood vessel
centerlin
e pix
e
ls. The f
i
nal s
e
g
m
entation
is
obtain
e
d using an iter
a
tive reg
i
on grow
ing method that merges
the binar
y
im
ages resulting
from centerl
ine
dete
cti
on par
t
with the im
age resulting from
fuzzy
vessel s
e
g
m
entation p
a
rt. I
n
this proposed
algorithm,
the b
l
ood vessel
is enhanced usin
g modified morphologi
cal operations and the s
a
lt
and pepp
er
noises are removed from retinal imag
es using Adaptive Fuzzy Switching
Median f
ilter.
This method is ap
plied on
two pub
licly
av
ailable d
a
tab
a
ses, th
e
DRIVE and the STARE and the
experiment
al res
u
lts obtain
e
d b
y
using green
channe
l im
ages have been prese
n
ted
and com
p
ared with recen
tl
y
publishe
d
methods. The results demons
trate that our
algor
ithm
is ver
y
eff
ective metho
d
to detect r
e
tin
al
blood vessels.
Keyword:
Ad
ap
tiv
e fu
zzy
switch
i
n
g
Med
i
an
filter
B
l
oo
d vessel
s
e
gm
ent
a
t
i
on
Fuzzy
vessel
s
e
gm
ent
a
t
i
on
M
odi
fi
e
d
m
o
rp
hol
ogi
cal
ope
rat
i
o
ns
Vessel cen
t
erlin
e
d
e
tectio
n
Copyright ©
201
4 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
R
azi
eh A
k
hava
n,
Depa
rt
m
e
nt
of
C
o
m
put
er E
ngi
neeri
n
g
,
Islamic
Azad Uni
v
ersity, Ga
zvin branc
h
,
Gazvin, Iran
Em
a
il: R.ak
h
a
v
a
n.a@g
m
ail.c
o
m
1.
INTRODUCTION
The di
a
g
nosi
s
of t
h
e
f
u
n
d
u
s i
m
age i
s
wi
del
y
used i
n
m
a
ny
m
e
di
cal
di
agnos
es. Im
age s
e
gm
ent
a
t
i
o
n
[1
] i
n
th
e fund
u
s
im
ag
e is th
e im
p
o
r
tan
t
facto
r
fo
r id
en
tifyin
g
t
h
e
retin
al
p
a
tho
l
og
y. Th
e an
alysis o
f
the
hum
an ret
i
n
a hel
p
s t
h
e
op
ht
hal
m
ol
ogi
st
s to i
d
ent
i
f
y
t
h
e
retinal disease. The disease s
u
ch as the
diabetes,
hype
rtension a
n
d arteri
osclerosis a
ffect t
h
e
retina a
n
d ca
us
e t
h
e c
h
a
nges
i
n
t
h
e
ret
i
n
al
bl
oo
d
vessel
s
[
2
]
.
T
h
e
chan
ges i
n
t
h
e
bl
o
od
vessel
a
nd t
h
e ret
i
n
al
p
a
t
hol
o
g
y
ca
n
be id
en
tified b
y
first seg
m
en
ting
th
e
retin
al v
e
ssels
an
d
b
y
pr
op
er an
alysis
of
th
e r
e
tin
al
b
l
o
o
d
vessels.
Au
t
o
m
a
t
i
c seg
m
en
tatio
n
of
retin
al v
e
ssels
is i
m
p
o
r
tant
for ea
rly diagnosis of eye
diseases like
di
abet
i
c
ret
i
n
o
p
at
hy
[
3
]
.
T
h
e
r
e are
vari
ou
s
segm
ent
a
t
i
on m
e
t
hods
fo
r se
gm
ent
i
ng t
h
e r
e
t
i
n
al
vessel
s
i
n
t
h
e
fund
u
s
im
ag
e wh
ich
seg
m
en
ts th
e retin
al v
e
ssels u
s
i
n
g
two
d
i
m
e
n
s
ion
a
l
m
a
tch
e
d
filters an
d
p
i
ecewise
t
h
res
hol
d
pr
o
b
i
ng
[4
,
5]
. T
h
ere are
ot
her
segm
ent
a
t
i
on
pr
ocesses
w
h
i
c
h i
n
cl
ude
se
gm
ent
a
t
i
on o
f
ret
i
n
al
v
e
ssels
u
s
ing
th
e Mu
m
f
o
r
d
-
Sh
ah
m
o
d
e
l and Gabor wav
e
le
t filter [6
]. Ex
tractio
n
o
f
retin
al b
l
ood
v
e
ssels is
d
o
n
e
u
s
ing
Wein
er filter and
th
e Morp
ho
l
o
g
i
cal op
era
tion
s
lik
e op
en
an
d
cl
o
s
e op
eratio
n
[7
]. Th
is
p
a
p
e
r
fo
cu
ses
o
n
segmen
tatio
n
o
f
t
h
e retin
al v
e
ssel
s
to
id
en
tify the ch
ang
e
s in
the retin
al v
e
ssel
wh
ich
o
c
cu
rs
d
u
e
to
retin
al p
a
tho
l
og
ies lik
e d
i
ab
etic retin
op
ath
y
[8
]. Vesse
l segmen
tatio
n
is do
n
e
u
s
in
g Ma
x-Tree t
o
re
pres
ent the
i
m
ag
e and
th
e bran
ch
es filtering
ap
pro
ach
to
seg
m
en
t th
e i
m
ag
e [9
]. M
a
th
em
at
ical
mo
rph
o
l
o
g
y
is
m
o
stl
y
use
d
for
analy
ze the s
h
ape
of the im
ag
e. The two
m
a
in
p
r
ocesses wh
ich in
vo
lv
e are
d
ilatio
n
an
d
ero
s
ion
.
Th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
56
1
–
57
2
56
2
algorithm
s
of open and cl
ose a
r
e ba
sed on the
s
e proces
ses.
T
h
ese algorithms are c
o
m
b
in
ed to
d
e
tect th
e ed
g
e
s
and ide
n
tifying the speci
fic sha
p
es
i
n
t
h
e i
m
age and al
so
for t
h
e
bac
k
g
r
o
u
nd rem
oval
[1
0]
. R
e
t
i
n
al
vesse
l
segm
ent
a
t
i
on i
s
d
o
n
e t
o
cl
assi
fy
t
h
e
pi
xel
as
t
h
e v
e
ssel
a
n
d
no
n
-
vess
el
usi
n
g
m
o
rp
hol
ogi
cal
t
h
res
h
ol
di
n
g
.
The
ret
i
n
al
bl
oo
d
v
e
ssel
i
s
ext
r
act
ed
by
fi
r
s
t
sm
oot
hi
n
g
t
h
e
image a
n
d e
nha
nced
by ap
plying the
fuzzy c
-
means
cl
ust
e
ri
n
g
al
go
ri
t
h
m
[11]
.
2.
METHODS
Thi
s
pa
per
pr
o
pos
es a no
vel
al
go
ri
t
h
m
for ret
i
n
al
bl
o
od
v
e
ssel
s
segm
entat
i
on. T
h
e fu
n
dus i
m
age
u
s
ed
in
th
is research
is ob
tain
ed
fro
m
two
p
u
b
licly
avai
l
a
bl
e dat
a
base
s,
t
h
e DR
I
V
E da
t
a
base [1
2]
an
d t
h
e
STAR
E
dat
a
ba
se [
13]
.
The
se
gm
ent
a
t
i
on o
f
t
h
e ret
i
n
al
bl
o
o
d
vessel
s
h
o
u
l
d
be a
u
t
o
m
a
t
i
c
an
d acc
urat
e
f
o
r t
h
e
d
i
agn
o
sis of t
h
e retin
al
d
i
sease. Th
e
p
r
o
p
o
s
ed
al
gorith
m
an
d
resp
ectiv
e details will b
e
exp
l
ain
e
d h
e
re.
2.
1.
Over
view of T
h
e Pr
oposed
Algorithm
The m
e
t
hod h
e
rei
n
p
r
ese
n
t
e
d can
be sche
m
a
t
i
call
y
descri
be
d by
t
h
e f
unct
i
o
nal
bl
oc
k di
a
g
ram
i
n
Fi
gu
re 1.
Fi
gu
re
1.
R
e
t
i
n
al
vessel
se
gm
ent
a
t
i
on
f
unct
i
onal
di
ag
ram
for
t
h
e
pr
o
p
o
s
e
d
al
g
o
r
i
t
h
m
Th
e alg
o
rith
m co
n
s
ists four
m
a
in
p
a
rts:
1
)
prep
ro
cessi
n
g
wh
ich
in
clu
d
e
s no
ise eli
m
in
atio
n,
back
g
r
o
u
nd
n
o
r
m
a
li
zat
i
on an
d t
h
i
n
vessel
e
nha
ncem
ent
us
i
ng
gree
n c
h
an
nel
o
f
t
h
e
ret
i
n
al
col
o
r i
m
ages 2)
pr
ocessi
ng
wh
i
c
h i
n
cl
u
d
es t
w
o
p
h
ases,
ve
ssel
cent
e
rl
i
n
e
det
ect
i
on a
n
d
Fuzzy
ves
s
el
segm
ent
a
t
i
on,
and
3
)
Co
m
b
in
in
g
two
resu
lts
o
b
t
ai
n
e
d fro
m
p
r
ev
io
u
s
step
s to
fin
a
lly ex
tract th
e co
m
p
lete p
i
x
e
ls b
e
l
o
ng
ing to
the
retinal ves
s
els.
2.
2.
Image Pre
p
rocessing
2.
2.
1.
RGB
to Gree
n Channel
Conversi
o
n
The c
o
lor fundus im
age is conve
rted t
o
gree
n c
h
an
n
e
l im
ag
e to
m
a
k
e
th
e seg
m
en
tatio
n
pro
cess m
o
re
easily and t
o
decrease
the
c
o
m
putational t
i
m
e
. The
gr
ee
n c
h
an
nel
i
m
age
pr
o
v
i
d
es
t
h
e m
a
xim
u
m
cont
rast
betwee
n the i
m
age and t
h
e background,
because the
retinal blood vess
el
inform
ation in the gre
e
n
channel
im
age is m
o
re clear [14,
15].
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A N
o
vel
Ret
i
n
a
l
Bl
oo
d Ve
ssel
Seg
m
e
n
t
a
t
i
o
n
Al
gori
t
h
m
usi
n
g F
u
zzy
seg
m
e
n
t
a
t
i
o
n (
R
a
z
i
e
h
Akh
a
v
an)
56
3
2.
2.
2.
No
ise Eliminatio
n using
No
i
s
e Ada
p
tiv
e
F
u
zz
y Switching
Medi
an f
ilt
er
(N
A
F
SM)
NAFSM filter is a recursiv
e d
oub
le
-stage
filter, wh
ere i
n
itially i
t
will
d
e
tect th
e sal
t
-and
-p
epp
e
r
n
o
i
se i
n
ten
s
ities b
e
fore id
en
tifyin
g
th
e l
o
catio
n
s
of po
ssi
b
l
e n
o
i
se
p
i
x
e
ls.
Wh
en
a “no
i
se p
i
x
e
l” is detected
, it
is sub
j
ected
to th
e
n
e
x
t
filterin
g
stag
e.
Wh
en
a
p
i
x
e
l
is i
d
en
tified
as “no
i
se-free,
” it
will b
e
retai
n
ed an
d th
e
filterin
g
actio
n is sp
ared
to avo
i
d
ch
an
g
i
n
g
an
y fi
n
e
d
e
tails th
at are con
t
ain
e
d in
t
h
e
o
r
i
g
in
al im
ag
e.
2.
2.
2.
1.
D
e
t
e
ct
io
n St
ag
e
NAFSM
filter will u
s
e th
e no
isy im
ag
e
h
i
stog
ram
to esti
m
a
te th
e two salt-an
d-p
e
pp
er
n
o
i
se
in
ten
s
ities [1
6]. Th
e l
o
cal
max
i
m
u
m
,
which
is th
e first p
eak en
coun
tered
wh
en
t
r
av
ersing
t
h
e
i
m
ag
e
h
i
stog
ram
in
a
p
a
rticu
l
ar d
i
rectio
n
,
is u
s
ed [1
7
]
.
Th
erefore, th
e NAFS
M filter will
s
earch
fo
r two lo
cal
m
a
x
i
mu
ms
,
=2
55
an
d
=0,
des
c
ri
bi
n
g
t
h
e t
w
o
sal
t
and
pe
p
p
e
r
n
o
i
s
e.
Wh
e
n
bot
h l
o
cal
m
a
xim
u
m
s
are fo
und
, th
e search
will be sto
p
e
d
imme
d
i
ately. A b
i
na
ry n
o
i
se m
a
sk
N(i,j)
will b
e
created
to
sho
w
t
h
e
l
o
cat
i
on
o
f
“
n
o
i
se pi
xel
s
”.
T
h
i
s
m
a
sk i
s
sh
ow
n i
n
E
quat
i
o
n
1
.
(1
)
Wh
ere X(i,j
)
i
s
th
e p
i
x
e
l at lo
catio
n
(i,j) wi
th
in
te
n
s
ity X. N(i,j
)
=1
rep
r
esen
t
“n
oi
se-
fre
e pi
xel
s
’
’
t
o
b
e
safed fro
m
th
e
n
o
i
sy im
ag
e wh
ile N(i,j)=0 rep
r
esen
ts
“n
oise p
i
x
e
ls”
subjected
to
t
h
e
n
e
x
t
filtering
stage.
2.
2.
2.
2.
Filtering Stage
After th
e
b
i
n
a
ry no
ise m
a
sk
N(i,j) is created
,
“n
o
i
se
p
i
x
e
ls” will b
e
rep
l
aced
b
y
an
estim
a
t
ed
co
rrectio
n term
[1
8
]
. NAFSM filter
uses a squ
a
re filtering
wi
n
dow
(i,
j
)
w
ith
od
d (
2
s+1
)
(2
s+1)
di
m
e
nt
i
ons,
gi
ven
as E
q
uat
i
o
n
2.
(2
)
The
n
, t
h
e
num
ber
o
f
“
n
oi
se-f
ree
pi
xel
s
”,
,
in th
e
filterin
g
wind
ow
(
i
,j
) i
s
coun
ted
usi
n
g E
q
uat
i
o
n
3.
(3
)
If th
e cu
rren
t
filterin
g
windo
w
(i
,j
)
d
o
es
not
ha
ve
a
m
i
ni
m
u
m
num
ber
of
on
e “
n
oi
se-f
ree
pi
xel
”
(
,
1
, t
h
en t
h
e filtering
wi
n
dow
will
b
e
ex
p
a
nd
ed
b
y
on
e p
i
x
e
l at
each
of its fou
r
si
d
e
s
(i.e.
s+1) [18]. Th
is pro
c
edu
r
e is rep
eated
un
til th
e con
d
ition
of
,
1
is occure
d. For each
d
e
tected
“no
i
se p
i
x
e
l”, th
e si
ze o
f
th
e
filterin
g
windo
w is i
n
itialized
to
3
3, i
.
e., s=
1 [
18]
.
These “
noi
se
-
free
p
i
x
e
ls”
will all b
e
u
s
ed
as cand
id
ates
for selecti
n
g
th
e m
e
d
i
an
p
i
x
e
l, M
(
i,j
)
, g
i
v
e
n b
y
Equatio
n
4
.
(4
)
Thi
s
cri
t
e
ri
o
n
of ch
o
o
si
n
g
o
n
l
y
“noi
se-f
ree pi
xel
s
” i
s
used
t
o
avoi
d sel
e
c
t
i
ng a “noi
se p
i
xel
”
as t
h
e
m
e
di
an pi
xel
.
Ho
we
ver
,
t
h
e
num
ber
of “
n
o
i
se-free
pi
x
e
l
s
” fo
r sel
ect
i
n
g
t
h
e m
e
di
an pi
xel
i
s
al
so i
m
port
a
nt
because a large num
ber of “
noise
-free pi
xe
l” sa
m
p
les w
ill cons
um
e higher com
putin
g tim
e
and also yield an
uns
ui
t
a
bl
e m
e
di
an t
e
rm
for
rest
o
r
at
i
on
[1
8
]
. Si
nce t
h
e de
t
ect
i
on o
f
“n
oi
se pi
xel
s
” i
s
b
a
sed o
n
t
h
e
de
t
ect
ed
salt-an
d-p
e
pp
er n
o
i
se
in
ten
s
i
ties
and
,
n
o
i
se-free
p
i
x
e
ls
may b
e
falsel
y id
en
tified as “no
i
se
p
i
x
e
ls” at im
a
g
e un
ifo
r
m
reg
i
on
s h
a
v
i
n
g
sam
e
in
ten
s
ities as in
ten
s
ities
or
.
T
h
er
ef
or
e
,
th
e
filtering wi
ndow will be expa
nde
d contin
uously and the se
lected
m
e
dian pi
xel m
a
y be i
n
accurate to be used
as a co
rrection
term
. Con
s
id
eri
n
g th
is
possib
ility, th
e
search fo
r “no
i
se-free
p
i
x
e
ls” is stop
ed
when
th
e
filterin
g
w
i
ndow
h
a
s reach
e
d
a size o
f
7
7 (or s=3) a
n
d no
“noise
-free
pixel” is detected, i.e.,
,
=0. In
th
is case, th
e
first fou
r
p
i
x
e
ls
in
th
e 3
3
filterin
g
wi
n
dow
d
e
fin
e
d
b
y
Equ
a
tio
n 5 will
b
e
u
s
ed
to co
m
p
u
t
e th
e
m
e
di
an pi
xel
M
(
i
,
j
)
i
n
Eq
uat
i
on
6:
(5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
56
1
–
57
2
56
4
(6
)
Fig
u
re
2
.
Fu
zzy set ad
ap
ted
by th
e NAFSM
filter
The fi
rst
f
o
u
r
p
i
xel
s
ch
ose
n
,
w
h
i
c
h m
a
de
up the upper-left
diag
on
al of
th
e
3
3
filtering
wi
n
dow, can
b
e
ju
stified
b
y
th
e recu
rsi
v
e natu
re
o
f
t
h
e NAFSM
filte
r.
As a resu
lt from th
e recu
rsive b
e
h
a
v
i
o
r
, any “n
o
i
se
pi
xel
s
” i
n
t
h
e
up
pe
r-l
eft
di
ag
onal
o
f
t
h
e
3
3 filtering
windo
w wou
l
d
h
a
ve b
e
en
restored
an
d upd
ated with
“noi
se
-f
ree pi
x
e
l
s
” duri
ng ea
r
l
i
e
r proce
ssi
n
g
.
There
f
ore,
us
i
ng t
h
e m
a
xim
u
m
four “n
oi
s
e
-f
ree pi
xel
s
” i
n
t
h
e
uppe
r-left dia
g
onal will yie
l
d a
m
o
re accurate m
e
dian
pixel instead
of consi
d
ering all
eight connected
nei
g
hb
o
r
i
n
g pi
xel
s
.
After t
h
e m
e
d
i
an
p
i
x
e
l M(i,j) is
fo
und
, th
e l
o
cal i
n
fo
rm
ati
o
n i
n
a
3
3
window is e
x
tract
ed by
first
com
put
i
ng t
h
e
abs
o
l
u
t
e
l
u
m
i
nance
di
ffe
re
nce
d
(
i
,
j
)
as
gi
ven
by
E
quat
i
o
n
7.
(7
)
Th
en
, th
e lo
cal in
form
atio
n
is d
e
fin
e
d
as the m
a
x
i
m
u
m
a
b
so
l
u
te lu
m
i
n
a
n
ce
d
i
fference in
th
e
3
3
filterin
g
wi
n
dow b
y
Eq
u
a
tion
8
.
(8
)
As p
a
rt o
f
t
h
e filterin
g
m
ech
an
ism
in
th
e NAFSM f
ilter,
fu
zzy reaso
n
i
n
g
is ap
p
lied
to
th
e ex
tracted
l
o
cal
i
n
fo
rm
ation
D(i
,
j) [
1
8]
. The f
u
zzy
set
ado
p
t
e
d i
s
sh
o
w
n i
n
Fi
g
.
2 a
n
d defi
ned
by
t
h
e fuzzy
m
e
mbers
h
i
p
fun
c
tion
F(i,j) i
n
Equ
a
tio
n 9.
(9
)
wh
ere t
h
e lo
cal in
fo
rm
atio
n
D(i,j) is
u
s
ed as
th
e fu
zzy
in
pu
t
v
a
riab
le, and th
e t
w
o
p
r
ed
efi
n
ed
t
h
res
hol
ds
and
ar
e set to
10
an
d 30
, r
e
sp
ectiv
ely, fo
r op
timal p
e
rf
or
m
a
n
ce [1
7,
1
8
]
.
Fin
a
lly, th
e co
rrectio
n
term
to
resto
r
e a d
e
tected
“n
o
i
se
p
i
x
e
l” is a
lin
ear co
m
b
in
atio
n
b
e
tween
th
e
p
r
o
c
essing
pi
xel
X(i
,
j)
a
n
d
m
e
di
an pi
xel
M
(
i
,
j
)
. T
h
e
rest
orat
i
o
n t
e
rm
Y
(
i
,
j
)
i
s
gi
ve
n
he
re as
Eq
uat
i
o
n
10
.
(1
0)
whe
r
e the fuzz
y
m
e
m
b
ership
v
a
lu
e F(i,j) lead
s a weigh
t
o
n
wh
eth
e
r
m
o
re o
f
p
i
x
e
l X(i,j
)
o
r
M(i,j) is
to
b
e
used.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A N
o
vel
Ret
i
n
a
l
Bl
oo
d Ve
ssel
Seg
m
e
n
t
a
t
i
o
n
Al
gori
t
h
m
usi
n
g F
u
zzy
seg
m
e
n
t
a
t
i
o
n (
R
a
z
i
e
h
Akh
a
v
an)
56
5
2.
2.
3.
Back
ground Enhanc
emen
t
2.
2.
3.
1.
Mo
dified T
o
p
-
ha
t Tr
ans
f
o
r
m
The classical top-hat tra
n
sform
is defined as
the diffe
rence bet
w
een an im
age and its ope
ne
d
v
e
rsi
o
n. A prob
lem asso
ciated
with
th
is classical i
m
p
l
e
m
en
tatio
n
is th
e sen
s
itiv
ity to
n
o
i
se, as a co
n
s
equ
e
n
ce
of the fact that
pixel values i
n
an
ope
n
ed i
m
age are always less than or
eq
u
a
l to
th
e orig
in
al on
es; in
su
ch
co
nd
itio
ns, th
e d
i
fferen
t
im
a
g
e retain
s all sm
a
ll
in
ten
s
ity flu
c
tu
ation
s
th
at can
b
e
fou
n
d
in
th
e
d
a
ta. To
ove
rc
om
e t
h
i
s
pr
o
b
l
e
m
,
a
m
o
di
fi
cat
i
on
was adapt
e
d fr
om
[
19]
, by
co
nsi
d
eri
n
g t
w
o ne
w
st
eps i
n
t
h
e t
op-
ha
t
defi
nition: a c
l
osing precede
s
the
openi
ng result which
is followed
by
a com
p
arison, using a m
i
nim
u
m
ope
rat
o
r, t
o
ge
t
an im
age eq
ual
t
o
t
h
e
ori
g
i
n
al
one e
v
e
r
y
w
he
re e
x
cept
f
o
r
pea
k
s an
d
r
i
dges
.
Eq
uat
i
o
n (
1
1
)
represen
ts th
is
m
o
d
i
fied
t
o
p-h
a
t tran
sfo
r
m
,
wh
ere I is
the
im
age to be processe
d,
while
Sc and So
stand
for
the structuring
ele
m
ents for cl
osing(•
)
a
n
d opening
(
) operators,
res
p
ectively[20].
To
p
H
at= I
–
m
i
n (I•
Sc
)
S
o
; I)
(1
1)
The cl
osi
n
g o
p
erat
i
o
n i
s
co
nsi
d
e
r
ed t
o
ge
nerat
e
a sm
oot
h ve
rsi
o
n o
f
t
h
e o
r
i
g
i
n
al
dat
a
, whe
r
e t
h
e
details sm
a
ller than t
h
e structuring elem
ent are re
pl
aced by hi
ghe
r
ne
arby inte
nsities. The
ope
n
ed i
m
age
essen
tially
m
a
i
n
tain
s th
e
p
i
x
e
l v
a
lu
es,
wh
ile eli
m
in
atin
g
mo
re i
n
ten
s
e imag
e reg
i
on
s
with
sizes sm
al
le
r than
th
e stru
cturing ele
m
en
t size.
Th
e
fin
a
l
resu
lt o
f
th
e sub
t
ractio
n
is an
enhan
ced im
ag
e th
at m
o
stly retain
s th
e
ori
g
inal im
age regions with s
i
ze s
m
al
ler th
an
th
e stru
ct
u
r
i
n
g
elem
en
t which
sho
w
sign
i
f
ican
t lo
cal in
ten
s
ity
vari
at
i
o
ns.
Here
we p
r
op
ose t
h
e m
odi
fi
ed To
p
-
hat
t
r
a
n
sf
orm
t
o
pr
o
duce t
h
e bac
k
gr
o
u
n
d
n
o
r
m
a
li
zed im
age.
Fi
gu
re
3(
b
)
p
r
esent
s
t
h
e
bac
k
g
r
ou
n
d
n
o
r
m
a
l
i
zed i
m
age obt
ai
ned
wi
t
h
t
h
e m
odi
fi
ed
t
o
p-
hat
ope
rat
o
r.
Si
nc
e
t
h
i
n
ve
ssel
s
ar
e very
sm
al
l
struct
ures a
n
d u
s
ual
l
y
have
l
o
w lo
cal con
t
rast, th
eir seg
m
e
n
tatio
n
is a
h
a
rd
task.
There
f
ore, it is necessary to
deepen t
h
e cont
rast of t
h
ese images to
p
r
ov
id
e a b
e
tter transform
rep
r
esentatio
n
for s
u
bseque
nt
im
age analysis steps.
C
L
A
H
E
t
echni
que
i
s
a
d
o
p
t
e
d
t
o
per
f
o
rm
t
h
e cont
ra
st
en
hancem
ent
.
Thi
s
technique
enhances t
h
e c
o
ntrast ada
p
tivel
y acros
s th
e
i
m
ag
e b
y
limitin
g
th
e m
a
x
i
m
u
m
slo
p
e
i
n
th
e
t
r
ans
f
o
r
m
a
ti
on
f
unct
i
o
n.
Fi
g
u
re
3
(
c) pres
ents the
vess
el enhance
d
i
m
ag
e ob
tain
ed
with
t
h
e C
L
AHE
technique.
a
b
c
Fi
gu
re
3.
a)
O
r
i
g
i
n
al
g
r
ee
n c
h
annel
b
)
B
ack
g
r
ound norm
aliz
ed im
age c)
Ve
ssel enhance
d
i
m
age
2.
3.
Image Pr
ocess
i
ng
2.
3.
1.
Detectio
n of
Vessel Centerline
Wh
en
a
first-ord
e
r d
e
riv
a
tiv
e filter is ap
p
l
i
e
d
orthog
on
ally to
th
e m
a
in
o
r
ien
t
atio
n
o
f
th
e v
e
ssel
,
d
e
ri
v
a
tiv
e v
a
lues with
o
ppo
site sig
n
s
are created
o
n
th
e
two
v
e
ssel h
illsides. Th
is id
ea is sh
own
in
Figure 4
(
a)
for a
n
ideal
ves
s
el cross
profil
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
56
1
–
57
2
56
6
Fig
u
re
4
.
Id
eal
v
e
ssel
p
r
o
f
ile
with
ex
p
ected
d
e
ri
v
a
tiv
e si
g
n
s on
o
ppo
site hillsid
es. (b
) C
o
m
b
in
atio
n
s
of
deri
vative si
gns and a
v
era
g
e
deri
vative
values that calcu
lat
e
the
occurrenc
e
of a
cand
id
at
e cen
terlin
e poin
t: +
means
a positi
ve deri
vative value;—m
eans
a
ne
gative
value; 0 m
eans a ze
ro val
u
e;
is a do
not ca
re
con
d
i
t
i
on;
A
D
V
i
s
t
h
e m
ean val
u
e
o
f
t
h
e
de
ri
vat
i
v
e m
a
gni
t
ude
s
obt
ai
ne
d
f
o
r
t
h
e sam
e
set
o
f
f
o
ur
pi
xel
s
were
the s
p
ecific c
o
m
b
ination of si
gns
occurre
d.
As retin
al
v
e
ssels ex
ist in
an
y
d
i
rection
,
we
n
eed
t
o
select
a set o
f
d
i
recti
o
n
a
l
filters who
s
e respon
ses
can be c
o
m
b
i
n
ed t
o
c
ove
r t
h
e
wh
ol
e ra
nge
o
f
p
o
ssi
bl
e
ori
e
nt
at
i
ons
. The
p
a
rt
i
c
ul
ar ke
r
n
e
l
s used i
n
t
h
i
s
wo
rk
are first
-
ord
e
r d
e
rivativ
e filters, w
ith
commo
n
respon
ses to
ho
rizo
n
t
al (0
), v
e
rtical (9
0), and
d
i
ag
on
al
(45
,
13
5) d
i
rect
io
n
s
. Th
ese
filters are u
s
ed
for th
e co
m
p
utatio
n
o
f
t
h
e lo
cal i
m
ag
e g
r
ad
ien
t
in
a specific
di
rect
i
o
n.
Herei
n
, th
e prop
o
s
ed
filters used
for d
e
tecting
cen
t
erlin
e can
d
i
d
a
te p
i
x
e
ls
an
d th
e
resu
lt
o
f
app
l
yin
g
th
ese filters
are sho
w
n
i
n
Fi
gure
5
an
d 6 respectiv
ely.
Fig
u
re 5
.
Th
e p
r
op
o
s
ed
filters
Fig
u
re
6
.
Th
e can
d
i
d
a
te cen
t
erlin
e p
i
x
e
ls b
y
ap
p
l
ying
th
e prop
o
s
ed filters:
a) vert
i
cal
b
)
h
o
ri
z
ont
al
c
)
13
5 d)
4
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A N
o
vel
Ret
i
n
a
l
Bl
oo
d Ve
ssel
Seg
m
e
n
t
a
t
i
o
n
Al
gori
t
h
m
usi
n
g F
u
zzy
seg
m
e
n
t
a
t
i
o
n (
R
a
z
i
e
h
Akh
a
v
an)
56
7
Each
on
e
o
f
the fou
r
d
i
rection
a
l i
m
ag
es resu
ltin
g
fro
m
th
e p
r
o
p
o
s
ed
filters is search
ed for sp
ecifi
c
com
b
i
n
at
i
ons
of
si
g
n
s
o
n
t
h
e ex
pect
ed
di
r
ect
i
on
of
t
h
e
v
e
ssel
cr
oss
sec
t
i
on
[2
0]
;
Si
nc
e real
vessel
s
do
not
have
t
h
e i
d
eal
pr
ofi
l
e
prese
n
t
e
d i
n
Fi
g
u
r
e
4(
a), a
set
o
f
fo
u
r
c
o
m
b
i
n
at
i
ons
re
prese
n
t
e
d
i
n
Fi
g
u
re
4
(
b
)
t
h
at
can
find
a v
e
ssel
are u
s
ed
.
In
t
h
is figu
re, p
l
us an
d
m
i
n
u
s
sig
n
s
sh
ow th
e p
o
s
itiv
e an
d
n
e
g
a
tiv
e
d
e
ri
vativ
e
responses, res
p
ectively,
0 m
eans a
n
u
l
l
ou
t
put
,
an
d
sh
ows a do
no
t care co
nd
itio
n
for the sign of the
deri
vat
i
v
e. M
o
reo
v
er
, i
n
c
o
n
d
i
t
i
ons
2 an
d
3, t
h
e a
v
era
g
e
val
u
e
of t
h
e
d
e
ri
vat
i
v
e m
a
gn
i
t
udes (
A
DV
)
fo
r t
h
e
in
ten
s
ity profile is calcu
lated. It m
u
st b
e
po
sitiv
e
for co
nd
itio
n
2, an
d
n
e
g
a
tiv
e for co
nd
itio
n
3
.
The fin
a
l
cen
terlin
e cand
id
ate is si
m
p
ly th
e p
i
x
e
l with
th
e m
a
x
i
m
u
m in
ten
s
ity in
th
e b
ackgroun
d
n
o
rm
alized
i
m
ag
e
[20
]
. Th
is
fact is o
b
t
ain
e
d
in a n
e
w im
ag
e b
y
app
l
yin
g
the su
m
o
f
th
e h
i
gh
est po
sitive respon
se wit
h
th
e
ab
so
lu
te v
a
lu
e o
f
th
e
m
o
st
n
e
g
a
tiv
e respon
se.
Upt
o
here
, w
e
coul
d ac
hi
eve
vessel
ce
nt
erl
i
n
es i
n
retin
al i
m
ag
es. Th
is meth
od ca
n
fully extract t
h
e
vessel
ce
nt
ers,
but
n
o
t
t
h
e
act
ual
wi
dt
h
o
f
t
h
e vessel
s
.
2.
3.
2.
Fuz
z
y
Segmen
ta
tion
W
i
t
h
th
e aim
o
f
ach
i
ev
ing
a co
m
p
lete seg
m
en
tatio
n
,
retin
al v
e
ssels n
e
ed
to
b
e
filled
startin
g
fro
m
the
detected ce
nterlines. For t
h
is
pu
r
p
o
s
e,
we pr
opo
se a fu
zzy app
r
o
ach
which uses
fuzz
y C-m
eans clustering
t
echni
q
u
e t
o
g
e
nerat
e
re
p
r
es
ent
a
t
i
on
of t
h
e
ret
i
n
al
vasc
ul
ar net
w
o
r
k
wi
t
h
act
ual
wi
dt
h
of t
h
e
vessel
s
.
W
e
apply the
fuzz
y C-m
eans clustering to t
h
e
vessel en
hance
d
im
age o
b
t
a
i
n
e
d
fr
om
prep
roc
e
ssi
ng
st
ep
.
2.
3.
2.
1.
F
u
zzy
C
-
m
e
an
s
(
F
CM
)
In
pattern
recognition a cl
ust
e
ring m
e
thod
known
as F
u
z
z
y C-Means
(FCM) is
widel
y
used. FCM
base
d se
gm
entat
i
on i
s
fuzzy
pi
xel
cl
assi
fi
ca
t
i
on [
2
2]
. FC
M
al
l
o
ws
dat
a
poi
nt
s o
r
pi
xel
s
t
o
bel
o
ng
t
o
m
u
lt
i
p
l
e
cl
asses wi
t
h
v
a
ry
i
ng
deg
r
ee
of m
e
m
b
ershi
p
fu
nct
i
on
bet
w
een 0 t
o
1. FC
M
possesse
s p
r
eci
o
u
s ad
vant
age o
f
g
r
ad
ing
lingu
istic v
a
riab
les
to
fit for ap
prop
riate an
alysis
in
d
i
screte
do
m
a
in
o
n
pro-rata b
a
sis. FCM
com
put
es cl
ust
e
r cent
e
rs by
m
i
nim
i
zi
ng t
h
e di
ssim
i
l
a
ri
ty
fu
nct
i
o
n usi
n
g
an i
t
e
rat
i
v
e app
r
oach
. B
y
updat
i
n
g
the cluster ce
nters and the me
m
b
ershi
p
gra
d
es for eac
h
un
iqu
e
p
i
x
e
l, FCM sh
ifts th
e clu
s
ter cen
ters to
th
e
"tru
e" lo
cation with
in
set o
f
p
i
x
e
ls. To
acco
m
m
o
d
a
te
th
e
in
trodu
ctio
n
o
f
fu
zzy p
a
rtitio
n
i
ng
, th
e m
e
mb
ersh
i
p
matrix (U) =
[
uij
]
is ra
ndom
ly initia
lized according t
o
Equa
tion
12,
whe
r
e
ui
j
being t
h
e
degree
of m
e
m
b
ershi
p
fun
c
tion
o
f
t
h
e d
a
ta
po
in
t of
cl
ust
e
r
xi
.
(1
2)
The
perform
ance inde
x
for m
e
m
b
ershi
p
m
a
trix
U a
n
d
’s
us
ed i
n
FC
M
i
s
g
i
ven
by
E
q
uat
i
o
n
1
3
.
(1
3)
ui
j
i
s
bet
w
ee
n
0 a
n
d
1.
ci
i
s
t
h
e ce
nt
er
o
f
cl
u
s
ter i.
d
ij is t
h
e Eu
clid
ian
distan
ce
b
e
tween
cente
r
(
c
i)
and
d
a
ta po
in
t. m
є
[1
,
∞
] is a
weighting
exponent. T
o
reach a
m
i
nim
u
m
of
dissimilarity function t
h
e
r
e
are t
w
o c
o
ndi
t
i
ons
[
2
2]
. T
h
es
e are
gi
ve
n i
n
Eq
uat
i
o
n
1
4
a
n
d E
q
uat
i
o
n
1
5
.
(1
4)
(1
5)
Al
g
o
ri
t
h
m
of
F
C
M
i
s
desc
ri
be
d as
bel
o
w:
Step
1
.
Th
e
m
e
m
b
ersh
ip
m
a
tri
x
(U)
th
at h
a
s co
nstrain
t
s
i
n
Eqn
1
2
is
rando
m
l
y
in
itialize
d
.
St
ep
2. C
e
nt
ers
(ci
)
a
r
e cal
c
u
l
a
t
e
d by
usi
n
g
Eq
n
14
.
Step
3
.
Dissimilarity b
e
tween cen
ters an
d
d
a
ta po
in
ts
u
s
ing
Eqn
1
3
is co
mp
u
t
ed
. Stop
i
f
i
t
s i
m
p
r
ov
em
en
t ov
er
pre
v
i
o
us i
t
e
rat
i
on
i
s
bel
o
w a t
h
res
h
ol
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
56
1
–
57
2
56
8
St
ep 4. A ne
w U usi
n
g
E
q
n 1
5
i
s
c
o
m
put
ed.
Go
t
o
St
ep 2 [
2
2]
.
Th
e
resu
lt o
f
ap
p
l
ying
th
e FC
M alg
o
rith
m
to
th
e
v
e
ssel en
han
ced im
ag
e is sho
w
n
i
n
Fi
gure
7
.
Fi
gu
re
7.
Im
age res
u
l
t
e
d
fr
om
f
u
zzy
segm
ent
a
t
i
on st
e
p
No
tic th
at
th
e
th
in
v
e
ssels i
n
th
e
fu
zzy
segmen
te
d
im
age
are not detected. The
r
efore
,
we
c
o
m
b
ine
t
h
e res
u
l
t
e
d i
m
ages
fr
om
t
w
o
pr
ocessi
ng
p
h
a
s
es,
vessel
ce
nt
erl
i
n
e
det
ect
i
on a
n
d
F
u
zzy
ve
ssel
segm
ent
a
t
i
on
t
o
ach
iev
e
a co
mp
lete v
e
ssel
seg
m
en
tatio
n
.
2.
4.
Regi
on Grow
i
n
g
Th
e fin
a
l im
a
g
e with
th
e seg
m
en
ted
v
e
ssels is o
b
t
ain
e
d
b
y
iterativ
ely
co
m
b
in
in
g
th
e cen
terlin
e
im
ages with the im
age that resulted
f
r
om
the f
u
zzy
segm
ent
a
t
i
on
part
.
Ve
ssel centerli
n
e pixels are used as
p
r
im
ary p
o
i
n
t
s fo
r a reg
i
on
growing
algo
rith
m
,
wh
ich
fill
th
ese po
in
ts b
y
ag
greg
ating
the p
i
x
e
ls in
th
e fu
zzy
seg
m
en
tatio
n
imag
e. Fin
a
l
resu
lt of th
e v
e
ssel fillin
g
p
a
rt is
illu
strated
in Fi
g
u
re
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A N
o
vel
Ret
i
n
a
l
Bl
oo
d Ve
ssel
Seg
m
e
n
t
a
t
i
o
n
Al
gori
t
h
m
usi
n
g F
u
zzy
seg
m
e
n
t
a
t
i
o
n (
R
a
z
i
e
h
Akh
a
v
an)
56
9
a
b
c
Figure
8. a) C
o
m
b
ination of t
h
e ce
nterline i
m
age w
ith t
h
e
fuzzy se
gm
ented im
age b) T
h
e fuzzy segm
ented
im
age c) T
h
e s
e
gm
ented im
ag
e usi
n
g re
gion
growi
n
g algori
thm
At the end, the final segm
ented im
age is a
c
hieve
d
by
ad
di
n
g
t
h
e fo
u
r
i
m
ages obt
ai
ne
d fr
om
t
h
e
regi
on
gr
o
w
i
n
g st
ep.
Som
e
exam
pl
es of t
h
e segm
ent
e
d
im
ages from
DR
I
V
E an
d S
T
AR
E dat
a
bas
e
s are
sho
w
n i
n
Fi
gu
r
e
9 a
n
d Fi
gu
re
10
res
p
ect
i
v
el
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
56
1
–
57
2
57
0
Fi
gu
re
9.
Tw
o
exam
pl
es of
se
gm
ent
e
d i
m
ages fr
om
DR
IV
E
dat
a
base
Fi
gu
re
1
0
. T
w
o e
x
am
pl
es of
segm
ent
e
d i
m
ages
fr
om
STA
R
E dat
a
ba
se
Th
e ev
alu
a
tion resu
lts are
g
i
v
e
n as the p
i
xel-wise
sen
s
itiv
ity, sp
ecificit
y
,
and
accu
r
acy o
f
all the
seg
m
en
tatio
n
in
co
m
p
arison
with
groun
d
tru
t
h
,
wh
ere
sensitiv
ity
is a n
o
r
m
a
l
i
zed
m
eas
u
r
e
o
f
tru
e
po
sitiv
es,
speci
fi
ci
t
y
m
e
asure
s
t
h
e
pr
o
p
o
r
t
i
o
n o
f
t
r
ue
negat
i
ves, a
n
d acc
uracy
re
prese
n
t
s
t
h
e
p
r
o
p
o
rt
i
o
n
of t
h
e t
o
t
a
l
n
u
m
b
e
r
o
f
correctly classified
p
i
x
e
ls relativ
e to
th
e t
o
tal nu
m
b
er of p
i
x
e
l
s
. In
th
e fo
llowing
eq
u
a
tions, tru
e
p
o
s
itiv
e (TP) i
s
a n
u
m
b
e
r o
f
b
l
oo
d
v
e
ssels co
rrectly d
e
tected
, false po
sitiv
e (FP) is a n
u
m
b
e
r o
f
non
-b
loo
d
vessel
s
w
h
i
c
h
are det
ect
ed
w
r
on
gl
y
as bl
oo
d
vessel
s
,
fal
s
e
negat
i
v
e
(F
N) i
s
a n
u
m
b
er of
bl
o
od
vessel
s
t
h
at
are
not
det
ect
ed a
nd t
r
ue
ne
gat
i
v
e (T
N) i
s
a
n
u
m
ber of
no
n
-
bl
oo
d
vessel
s
w
h
i
c
h a
r
e c
o
r
r
ec
t
l
y
i
d
ent
i
f
i
e
d a
s
n
o
n
-
b
l
oo
d v
e
ssels.
Tabl
e
1 a
n
d
2
com
p
ares e
v
a
l
uat
i
o
n
res
u
l
t
s
of
p
r
o
p
o
sed
meth
od
with
t
h
e lis
te
d
ap
pr
o
a
ch
e
s
on
the
DR
I
V
E an
d S
T
AR
E dat
a
bas
e
s, res
p
ect
i
v
el
y
.
C
o
m
p
ari
s
ons base
d
on
DR
I
V
E dat
a
b
a
se i
ndi
cat
e t
h
at
o
u
r
propose
d
m
e
thod has the
highest accur
acy, sensitivity and specific
ity compare
d
to [24].
In a
ddition, it
has the
highest accura
cy and se
nsitivity com
p
ar
ed to [23], [25], [27]
, [28] and [29]. More
ov
er, it has the
highest
accuracy a
n
d s
p
ecificity in
com
p
arison with [26].
Com
p
aring the
m
e
thods
on t
h
e STARE
database illust
rate that our
propos
ed m
e
thod has
the highes
t
accuracy, sensitivity and spec
ificity co
m
p
ared to [30], [31] and [5]. In
a
ddition, it
has
the hi
ghe
st acc
uracy
an
d sen
s
itiv
ity in
co
m
p
arison
with
[32
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.