Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
1
,
Febr
uar
y2
020,
pp. 7
46
~
756
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
1
.
pp746
-
756
746
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Mic
roar
ray spot p
ar
titi
on
ing by a
utonomou
sly
organisi
ng m
aps
th
r
o
u
gh cont
our
model
Ka
r
th
ik
S
.
A
.
1
, Manjun
ath
S
.
S
.
2
1
Depa
rtment of I
nform
at
ion
Sci
e
nce
and Engi
ne
e
ring,
Da
y
a
nand
a
Sagar
A
ca
dem
y
of
T
ec
hno
log
y
a
nd
Mana
gemen
t
,
India
2
Depa
rtment of
Com
pute
r
S
ci
en
ce
and Engi
ne
ering,
ATME
,
Indi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
y
12
, 201
9
Re
vised
Sep
2
9
,
2019
Accepte
d
Oct
5,
2
0
19
In
cDNA
m
ic
roa
rra
y
image
ana
l
ysis,
cl
assifi
ca
t
io
n
of
pixe
ls
as
fo
ref
ront
ar
e
a
and
the
area
c
over
ed
b
y
ba
ck
groun
d
is
ver
y
cha
llenging.
In
m
ic
roa
rra
y
expe
riment
at
ion
,
ide
ntif
y
ing
for
efr
ont
ar
ea
of
d
esire
d
spots
is
nothi
ng
but
computat
ion
of
fore
front
pix
els
conc
ent
r
at
ion
,
are
a
cove
r
ed
b
y
spot
and
shape
of
the
spots.
In
thi
s
piece
of
writi
ng
,
a
n
innova
ti
v
e
wa
y
for
spo
t
par
ti
t
ioni
ng
of
m
ic
roa
rra
y
ima
ges
using
aut
o
nom
ously
org
an
iz
ing
m
aps
(AO
M)
m
et
hod
through
C
-
V
mode
l
has
bee
n
p
roposed.
Concept
of
neur
al
net
works
has
be
en
inc
o
rpa
t
ed
to
tra
in and to
t
est m
ic
roa
rra
y
spots
.
In
a
traine
d
AO
M
the
comprehe
nsive
informati
on
ari
sing
fro
m
the
prototy
p
e
s
of
cre
a
te
d
neur
ons
are
clea
rl
y
integra
t
ed
to
dec
id
e
wheth
er
to
get
sm
aller
or
get
bi
gge
r
of
cont
our
.
Duri
ng
the
pro
ce
ss
of
opti
m
iz
a
ti
on
,
thi
s
is
done
in
an
itera
t
ive
m
anne
r.
Next
using
C
-
V
m
odel
,
i
nside
cur
v
e
ar
ea
of
tra
in
ed
spot
i
s
compare
d
with
te
st
spot
fi
nal
l
y
cur
ve
fi
tt
i
ng
is
done.
Th
e
pre
sente
d
m
odel
ca
n
hand
l
e
spots
with
var
i
at
ions
in
te
rm
s
of
shape
and
qual
i
t
y
of
th
e
spots
and
m
ea
nwhile
i
t
is
robust
to
the
no
i
se.
From
the
rev
ie
w
of
exp
eri
m
e
nta
l
work,
pre
sente
d
appr
o
ac
h
is
acc
ura
te
over
th
e
appr
o
a
che
s
l
ike
C
-
m
e
a
ns
b
y
fu
z
z
y
,
Morpholog
y
sec
t
iona
liza
ti
on.
Ke
yw
or
d
s
:
AOM
C
-
V
m
od
el
Level
set
Mi
cro
ar
ray
Spot
pa
rtit
ion
Copyright
©
202
0
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Kar
t
hik
S
.
A
.
,
Dep
a
rtm
ent o
f Info
rm
at
ion
Sc
ie
nce and En
gi
neer
i
ng,
Dayana
nd
a
Sa
gar Aca
dem
y of
Tec
hnology a
nd Mana
gem
e
nt,
Kan
a
kpura
Roa
d,
U
dayp
ur, Ba
ng
al
or
e
-
560082
,
Ka
rn
at
a
ka,
I
nd
ia
.
Em
a
il
:
kar
thiks
a199
0@gm
ai
l.
com
1.
INTROD
U
CTION
In
t
he
dom
ai
n
of
bio
i
nfor
m
at
i
cs
cDNA
m
i
cro
a
rr
ay
is
fa
s
te
st
dev
el
opin
g
te
ch
nolo
gy
wh
ic
h
he
lps
resear
c
hers
to
qu
a
ntify
the
be
hav
i
or
of
genes
(in
te
rm
s
of
thousa
nds)at
t
he
sam
e
tim
e
.
M
ic
ro
ar
ray
s
ubstrat
e
consi
sts
of
ti
ny
sp
ots
w
hich
expresses
s
ome
un
i
qu
e
i
nform
at
ion
about
the
ge
nes
[
1]
.
S
ub
st
rate
prepa
r
at
ion
i
s
done
by
colle
ct
ing
m
RNA
sam
ples
the
pu
r
pose
of
c
ollec
tin
g
m
RNA
is
to
ide
ntify
w
hich
spe
ci
fic
ge
nes
a
r
e
act
ively
involv
ed
in
giv
in
g
ge
ne
e
xpressi
on.
Ne
xt,
colle
ct
ed
m
RNA
sa
m
ple
will
un
de
rgo
a
pr
ocess
cal
le
d
hybri
dizat
ion
wh
e
re
sam
pl
e
get
tra
ns
cri
pt
i
nto
cD
NA.
Further
,
sam
ples
pr
e
par
e
d
by
t
he
m
entioned
proces
s
wer
e
pl
otted
on
m
ic
ro
ar
ray
su
bs
t
rate.
At
la
st,
entire
su
bs
t
r
at
e
is
scann
ed
at
a
sp
eci
fic
laser
wa
velen
gth
wh
ic
h
resu
lt
s i
n
a m
icr
oa
rr
ay
im
age.
By
m
at
ching
norm
al
sp
ots
with
dise
ased
s
pot
on
e
can
judg
e
wh
ic
h
ge
ne
is
aff
ect
ed
an
d
can
de
velo
p
a
drug
f
or
c
uri
ng
disease
d
gen
e
[
2
]
.
E
nti
re
m
ic
ro
arr
ay
analy
sis
invo
lves
three
m
ajor
ste
ps
li
ke
arr
ay
local
iz
at
ion
,
se
par
at
in
g
spot
f
or
e
gro
und
f
rom
back
gro
und
,
qu
al
it
y
assesm
ent
[
3
]
.
Arra
y
loca
li
zat
ion
i
nvolv
e
s
in
ide
ntifyi
ng
e
xact
locat
io
n
of
ge
ne
s
pot
in
s
ub
gri
d.
O
nce
l
ocati
on
of
the
sp
ot
is
i
den
ti
fi
ed
in
a
sub
gr
i
d
ne
xt
ta
sk
is
to
sep
arati
ng
a
s
po
t
fr
om
it
s
back
gro
und
su
c
h
s
epar
at
io
n
is
popula
rly
know
n
as
spot
sepa
rati
on.
Last
step is t
o co
m
pu
ti
ng
uni
qu
e
s
po
t i
ntens
it
y values
of
a
su
bst
rate.
Im
age
segm
entat
ion
is
def
i
ne
d
as
the
pro
cess
of
div
i
din
g
an
im
age
into
co
ns
ti
tuent
re
gions
or
obj
ect
s. In
t
he c
on
te
xt of DN
A
m
ic
ro
arr
ay
im
ages,
segm
entat
ion
’s go
al
is
to
div
i
de
eac
h
gr
i
d
cel
l i
nto
re
gio
ns
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Mi
croarray
s
pot p
ar
ti
ti
on
i
ng
by au
t
onomo
usl
y o
r
ganisin
g mapsth
rou
gh c
on
t
our mo
del
(
Karthik
S
A
)
747
corres
pondin
g
to
the
sp
ot
(
foregr
ound)
a
nd
backg
rou
nd.
Se
gm
entat
i
on
is
ty
pical
ly
per
f
or
m
ed
after
the
m
ic
ro
arr
ay
im
age
has
unde
r
gone
a
gr
i
dd
i
ng
ste
p
an
d
it
has
been
div
i
ded
i
nto
cel
ls
(also
known
as
s
ub
-
gr
i
ds
)
.
Af
te
r
se
gm
ent
at
ion
,
ge
ne
e
xpressi
on
le
vels
are
e
stim
at
ed
from
the
f
or
e
gro
und
a
rea.
A
nu
m
ber
of
factors
m
ake
the
segm
entat
ion
of
m
icr
oa
rr
ay
im
ages
a
chall
eng
i
ng
ta
sk
.
T
o
nam
e
a
few
,
t
he
bac
kgr
ound
is
ty
pical
ly
con
ta
m
inate
d
by
noise
.
S
pots’
sha
pes
a
nd
siz
es
m
igh
t
di
ff
er
within
a
n
i
m
age
from
on
e
spot
to
a
no
t
her.
The
i
ntensity
of
a
s
pot
is
not
necessa
rily
unif
or
m
.
Also
s
ince
the
hy
br
i
di
zat
ion
process
is
not
hom
og
eneous
the s
po
t
reg
i
ons could
b
e
brok
en.
M
or
e
over
, t
he
qual
it
y of
DNA
m
ic
ro
ar
r
ay
i
m
ages mi
gh
t va
ry.
The
backg
rou
nd
noise
a
nd
var
ia
ti
on
in
t
he
sh
a
pe
of
t
he
spots
can
be
cl
early
seen
i
n
the
pict
ur
e
.
Figure
1
ex
hibi
ts
so
m
e
sa
m
p
le
sp
ots
of
dif
f
eren
t
DNA
m
i
cro
a
rr
ay
im
ages.
As
ca
n
be
s
een
the
re
is
a
wide
range
of
var
ia
ti
on
s
in
te
rm
s
of
sh
a
pe
a
nd
qu
al
it
y
of
th
e
spots.
S
eei
ng
that
,
se
gm
entat
ion
m
et
hodo
l
og
ie
s
ty
pical
ly
necessit
at
e
s
m
anu
al
interfer
e
nce
to
giv
e
t
he
requisi
te
para
m
et
ers
or
to
exact
their
r
esults.
C
onversel
y,
there
is
a
le
ss
s
cop
e
of
aut
oma
ti
on
that
can
extensi
vely
influ
ence
s
bi
olog
ic
al
interpr
et
at
i
on
s
.
In
fact,
it
has
be
en
s
how
n
t
hat
the
c
hoic
e
of
t
he
se
gm
entat
ion
m
et
ho
d
has
s
ign
ific
a
nt
ef
fec
ts
on
the
pr
ov
i
sion
of the
outc
om
e
of a
n
e
xp
e
rim
ent
.
Figure
1.
Ty
pi
cal
m
ic
ro
arr
ay
sp
ots
Hen
ce
,
for
t
he
perfect
cat
eg
ori
zat
ion
of
ge
ne
ex
pr
es
sio
n
a
nd
to
ge
ne
rali
ze
spot’s
em
inence
m
easur
es
th
ere
is
a
gr
eat
dem
a
nd
for
the
r
obos
ti
c
ap
proa
ch
.
F
ro
m
the
afo
rem
entione
d
sta
ges,
one
can
concl
ude
spot
par
ti
ti
on
i
ng
is
on
e
of
t
he
m
os
t chall
en
ging t
ask
.
A
few
ye
ars
ba
ck
,
num
ero
us
m
ercantil
e
pack
age
s
and
in
ve
sti
gative
m
et
ho
ds
hav
e
been
pro
po
se
d
f
o
r
the
s
egm
entat
ion
of
im
ages
.
Existi
ng
se
gm
entat
ion
[
4
-
8
]
(p
arti
ti
on
of
the
f
or
e
fro
nt
area
f
ro
m
the
entire
i
m
age)
m
e
thodo
l
og
ie
s
a
re
cat
egorised
a
s
rin
g
fixe
d
0
segm
entat
ion
,
ada
ptati
on
rin
g
segm
entat
ion
,
m
or
phologica
l
adap
ta
ti
on
se
gm
entat
ion
.
I
n
Gen
e
Pix
versat
il
e
ring
di
vi
sion
m
et
ho
d
is
util
iz
ed
.
As
per
the
ap
proach,
the
ra
diu
s
of
al
l
sp
ot
is
not
re
gu
la
r
but
fine
t
un
e
d
in
de
pend
ently
fo
r
e
ver
y
sp
ot.
Neverth
el
ess,
a
rou
nd
s
po
t
m
ask
yi
el
ds
a
w
or
st
fit
to
unpredict
able
sp
ots
as
the
m
et
ho
dolo
gy
r
est
rict
s
the
s
ha
pe
of
the
re
gions
as
ci
rcu
la
r.
I
n
Sc
anA
ly
ze
0
[
9
-
12
]
,
a
fixe
d
rin
g
se
gm
entat
ion
m
et
ho
d
im
pl
e
m
ented
by
m
eans
of
the h
yp
oth
e
sis t
hat
area
s ar
e
a
ssu
m
ed
to b
e
gl
obular wit
h
a
pr
eset
rad
i
us
. Sp
ot
-
on
als
o
use
s the sam
e p
rincipl
e
of f
ixe
d
ci
rcle
segm
entat
ion
.
The
SP
OT
Sof
tware
inco
r
por
at
es
the
m
e
thods
[
12
]
su
c
h
as
watersh
e
d
a
nd
seede
d
re
gion
gro
wing
Adva
nced
im
a
ge
ha
ndli
ng
t
echn
i
qu
e
s
us
e
d
in
the
do
m
ai
n
of
m
ic
ro
a
rr
ay
co
ns
ist
of
the
m
or
phol
og
ic
al
adap
ta
ti
on
se
gm
entat
ion
.
T
hi
s
ap
proac
h
does
not
dep
e
nd
ent
on
a
ny
po
stulat
ion
a
bout
the
dim
ension
a
nd
the
co
ntour
of
the
spot.
T
he
f
la
w
of
both
wa
te
rsh
e
d
an
d
se
eded
re
gion
gr
ow
i
ng
a
ppr
oac
hes
are
,
they
dem
and
the
preci
se
c
onditi
on
of
t
he
pr
el
im
inary
points
or
the
see
ds
.
The
seede
d
re
gion
gro
w
ing
a
ppr
oac
h
[
13
]
is
furthe
r
fle
xib
l
e
an
d
a
da
ptabl
e
to
div
e
rse
c
on
t
ours
a
nd
di
m
ension
s
of
t
he
s
pots
bu
t
it
is
quit
e
se
ns
i
ti
ve
to
no
ise
.
Param
et
ers
su
c
h
as
sta
rt
ing
val
ue
an
d
porc
h
value
a
r
e
sel
ect
ed
by
con
si
der
e
d
s
po
t.
Approac
h
w
orks
on
low
inten
sit
y
values
of
spot
s,
i
m
pr
oper
s
ha
ped
s
pots
al
so
.
A
bru
pt
pix
e
l
values
do
not
aff
ect
s
execut
ion
of
pro
po
se
d
a
ppr
oach
.
Most
i
m
po
rtan
t
diff
ic
ulty
of
histo
gr
am
app
ro
ac
hes
is
that
qu
a
ntific
at
ion
is
un
ste
ady
after
a
gr
ea
t
intenti
on
m
ask
is
sit
uate
to
va
riat
ion
.
Acti
ve
co
ntour
m
eth
od
ologies
has
bee
n
discu
sse
d
in
[1
4,
15
]
,
wh
e
re
a
prel
i
m
inary
con
t
our
is
t
o
be
f
ound
on
the
s
po
t
-
im
age
and
a
co
ntour
be
nd
is
pro
duced
f
or
the
jo
b
of
outl
ine
the
spot
bounda
ries.
Mi
xture
m
od
el
m
e
tho
ds
[1
6,
17]
are
inco
r
porated
f
or
the
job
of
pa
rtit
ion
in
g
m
ic
ro
ar
ray
i
m
ages.
O
n
th
e
oth
e
r
ha
nd,
these
m
et
ho
ds
are
de
pe
nd
e
nt
on
i
ntensity
and
do
not
thi
nk
ab
out
the
re
li
able
a
m
on
g
a
djo
ini
ng
pix
el
s.
T
he
y
rely
on
assum
ption
of
norm
al
ity
fo
r
the
ap
plica
ti
on
to
the
segm
e
ntati
on
pro
blem
(p
ara
m
et
ric
assu
m
ption
of
norm
al
distribu
ti
on
).
Othe
r
sch
olars
ini
ti
at
ed
m
uc
h
releva
nt
pro
cedures
for
the
proce
ss
;
to
fu
sio
n
sc
ul
pt
analy
sis
[1
7
]
and
defor
m
able
m
o
dels
[18
]
.
In
t
he
m
ixtur
e
m
od
el
ana
ly
sis
,
the
m
os
t
i
m
po
r
ta
nt
ne
gative
a
sp
ect
is
th
e
detect
ion
of
inten
sit
y
near
t
o
it
s
sp
ot
bac
kgr
ound.
A
pp
ro
ac
he
s
suc
h
as
K
-
m
eans
,
Fu
zzy
C
m
eans
,
Exp
ect
at
io
n
-
Ma
xim
iz
ation
[
19
-
22
]
et
c.,
h
ave
been
us
e
d
by
seve
ral
scho
la
rs.
K
-
m
eans
ref
le
ct
on
a
fe
w
narrow
featu
res,
li
ke
noise
.
Als
o,
the
par
ti
ti
on
e
d
fo
re
fro
nt
an
d
su
r
rou
nd
i
ngs
s
ect
ion
need
not
be
l
ink
e
d
in
this,
te
chn
i
que
but
act
ua
ll
y
the
f
or
e
fron
t
an
d
bac
kdr
op
are
associat
e
d
re
gion
s
.
The
m
os
t
i
m
po
rtant
ne
gative
aspect
is,
they
no
t
get
us
e
d
to
fit
to
irregular
base
cl
us
te
r
s
and
fail
s
to
util
iz
e
al
l
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
1
,
Febr
uar
y
2020 :
74
6
-
756
748
the
a
vaila
ble
pr
i
or
kn
ow
le
dge
a
bout
t
he
data.O
ffset
ve
ct
or
fiel
d
a
nd
exp
ect
at
io
n
m
a
xim
iz
at
ion
al
gorithm
descr
i
bes
a
un
i
qu
e
a
ppro
a
ch
to
suppress
un
wan
te
d
pix
el
s
in
ge
nes.
Usi
ng
m
axi
m
iz
at
ion
al
gorithm
sepa
rati
on
of
bac
kgr
ound
done
.
E
xp
e
ri
m
ental
resu
lt
s
rev
eal
s
t
hat
pro
po
se
d
a
ppr
oach
is
at
tract
ive
an
d
yi
el
ds
ver
y
prom
isi
ng
perfo
rm
ance.
G
ra
ph
base
d
m
et
h
od
s
[
23
]
,
desc
ribes
a
way
f
or
cl
assifi
cat
io
n
of
pix
el
s
.
Au
t
ho
r
conveys
that
s
po
t
se
gm
entat
i
on
of
m
ic
ro
arra
y
is
a
chall
eng
i
ng
ta
s
k
a
nd
pr
opos
e
d
a
ppr
oac
h
is
aut
om
at
ed.
Descr
i
bed ap
pr
oach w
orks
on i
rr
eg
ular
spot s
hap
e
and
siz
e
wh
ic
h resu
lt
s i
n hig
h
acc
ur
ac
y
of ap
proac
h.
S
el
f
-
m
otivate
d
curvatu
re
ap
proac
h
[
24
]
pro
j
ect
s
a
fe
w
ide
as
to
par
ti
t
on
an
im
age
by
i
gnori
ng
ed
ge
picture
val
ues
.
Im
age
bo
undar
ie
sa
re
undoubte
dly
recti
fied
with
sli
ght
cha
nges.
Level
set
a
nd
C
-
V
appr
oach
[
25,
26
]
has
il
lustr
at
ed
an
inno
va
ti
ve
way
to
separ
at
e
im
age
us
in
g
le
vel
set
scaff
ol
d.
Bo
unda
ry
values
can
hel
p
to
fi
nd
out
preci
se
locat
io
n
of
inte
ntion
s
hap
e
an
d
res
ol
ves
dilem
m
a
of
bord
e
r
li
n
e
le
akag
e
Descr
i
ptive
in
sta
nces
of
the
work
ca
rr
ie
d
ou
t
sig
nifies
the
help
fu
l
nes
s
of
t
h
e
m
e
tho
d.
A
ne
w
aut
om
at
ed
appr
oach
[27]
fo
r
se
gm
enting
m
ic
ro
arr
ay
i
m
ages
descri
bes
a
way
f
or
ac
hieving
op
ti
m
iz
ation
thr
ough
gen
et
ic
al
ly
based
al
gorithm
.
Pr
op
os
e
d
ap
proach
is
ad
va
ntageous
ov
e
r
noise
an
d
res
ults
are
excell
ent
whe
n
they
are
te
ste
d
du
rin
g
w
or
s
t
ca
se
.
Ill
us
trat
es
[
28
]
innov
at
i
ve
s
ect
ion
s
uppor
t
m
et
ho
d
f
or
i
m
age
sect
ion
al
isa
ti
on
that
is
ca
pabl
e
to
re
so
l
ve
c
on
ce
ntrati
on
i
n
hom
og
eneit
y.
Pr
im
aril
y
confine
d
inte
ns
it
y
base
d
gro
up
i
ng
c
rite
r
ion
has
bee
n
f
un
ct
io
nal
the
n
inco
rpor
at
e
d
w
it
h
com
pr
ehe
nsi
ve
crit
er
io
n
f
or
im
age.
Ulti
m
at
ely
reducin
g
of
en
erg
y
is
con
ce
de
d
out
by
L
-
S
m
od
el
.
And
a
m
od
ifie
d
ver
si
on
[
29
]
of
cl
ust
ering
al
go
rithm
to
su
pp
ress
unw
anted
pi
xel
inf
or
m
at
ion
from
i
m
ages.
In
it
ia
ll
y,
qu
al
ity
of
the
i
m
a
ge
i
m
pr
ove
d
us
in
g
enh
a
ncem
ent
appr
oach
t
hen
us
in
g
va
riance
betwee
n
cl
ass
of
s
pots
m
e
tho
d
is
a
pp
li
e
d
to
ad
dress
the
sp
ot
.
Finall
y,
cl
us
te
r
ing
al
gorithm
i
s
app
li
ed
to
s
pot
s
.
A
no
vel
w
ay
[
30
]
to
seg
m
ent
the
i
m
ag
e
by
reducin
g
l
evel
of
no
ise
is
ac
hieved
.I
niti
al
ly
t
he
pro
pose
d
gro
w
-
c
ut
al
gori
thm
app
li
ed
to
ev
e
ry
spot
to
cl
assify
fro
m
its
backg
rou
nd.
T
o
asses
s
the
novelty
,
the
pr
opose
d
a
ppr
oac
h
is
im
ple
m
ented
th
rou
gh
m
ulti
thred
e
d
CP
U
a
nd
GPU.
P
rop
os
e
d
ap
proac
h
[31]
is co
m
bin
ed
s
om
e o
f
the popular tra
diti
on
al
algorit
hm
l
ike cann
y, m
orpholo
gy,
FCM
.
By
experim
enta
l
resu
lt
s
the
pr
op
ose
d
al
gorithm
i
s
fast
en
ough
to
segm
en
t
t
he
spots.
T
o
cl
assify
the
s
po
t
a
s
for
efron
t
pa
rt
a
nd
bac
kgr
ound
obta
ined
im
age
is
pr
e
processe
d
fo
ll
owe
d
by
sepa
rati
on
of
sp
ot
from
f
or
ef
r
on
t
par
t.
Ne
xt, un
i
qu
e
expr
essio
n i
s co
m
pu
te
d
[
32
]
a
nd d
at
a is
norm
al
iz
ed.
Finall
y ob
ta
ine
d resu
lt
s
are
op
ti
m
iz
ed
.
In
m
ic
ro
im
ag
e
analy
sis,
it
is
necessa
ry
to
do
an
im
age
pa
r
ti
ti
on
as
a
f
or
e
fron
t
a
rea
an
d
backg
rou
nd
reg
i
on.
A
m
ajo
r
prob
le
m
that
aff
ect
s
m
ic
ro
arr
ay
im
age
par
ti
ti
on
in
g
is
an
irre
gu
la
r
dist
rib
ution
of
int
ensity
values
.
S
om
et
i
m
es
these
inte
ns
it
y
value
s
a
ppear
to
reside
i
n
the
backgro
und
or
it
m
ay
be
in
t
he
forefro
nt
area
of
the
s
pot.
T
his
le
ads
to
c
onf
us
io
n
w
het
her
a
pi
xel
be
longs
to
the
f
or
e
fro
nt
area
of
s
po
t
or
bac
kgr
ound
reg
i
on.
This
pa
per
disc
us
ses
a
no
vel
m
et
h
od
to
s
olv
e
th
e
m
isc
la
ssific
a
ti
on
of
pix
el
s
as
fo
re
fro
nt
pix
el
s
or
backg
rou
nd
pi
xels.
T
he
w
ho
l
e
edi
to
rial
com
po
sit
io
n
fr
am
ed
as
:
di
vision
2
descr
i
bes
i
nnov
at
ive
way
f
or
s
po
t
par
ti
ti
on
i
ng
ba
sed
on
AO
M
-
CV
m
od
el
.
Divisio
n
3
descri
bes
the
ex
pe
r
i
m
e
ntal
values
of
s
om
e
te
st
sp
ots
.
Pr
ese
nted w
ork
is e
ve
ntu
al
ly
co
ncl
ud
e
d
at
th
e last
.
2.
AU
TO
N
OUM
OUS
L
Y
O
RGANI
SI
NG
M
APS
THRO
UGH
CHA
N
VESE(
AOM
-
CV)
METHO
DOL
OGY
In
t
he
prese
nt
wor
k,
a
nove
l
un
s
uper
vised
auto
nom
ou
sly
orga
nizing
m
ap
ba
sed
on
Chan
Vese
(AOM
-
C
V)
is
p
r
opos
e
d.
T
his
m
od
el
is
base
d
on
a
group
of
trai
ne
d
a
uton
om
ou
sly
orga
ni
zi
ng
neur
on
s
to
sto
r
e
intensit
y
value
s
of
the
im
age.
AO
M
-
C
V
m
o
del
e
m
plo
ys
A
OM
as
a
too
l
to
ex
plore
the
intensit
y
distrib
ution
of
a
spot
an
d
joins
t
he
protot
ypes
of
le
ar
ne
d
A
OMs
t
o
co
ntr
o
l
the
grow
t
h
of
c
on
t
our.
These
prototyp
es
ar
e
e
m
plo
ye
d
to
est
i
m
at
e
intensit
y
distribu
ti
on
an
d
to
gu
i
de
the
gro
wth
of
c
on
t
our.
T
he
ge
neral
idea
of
the
m
et
ho
dolo
gy
is
sh
own
in
Fig
ur
e
2.
Spots
are
captu
re
d
from
t
ypic
al
databases
su
c
h
as
GEO
-
DB
[3
3]
.
This
wo
r
k
c
on
t
rib
utes
.
Fo
rm
ulati
on
of unsupe
rv
ise
d ac
ti
ve
co
ntour
m
od
el
Gu
i
de fo
r
c
on
t
our gro
wth
Pr
ese
nts a
well
experim
ental
w
ork
in
term
s o
f
accu
racy a
nd
rob
us
tness
Figure
2.
O
verv
ie
w
of
AOM
-
C
V
a
ppr
oach
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Mi
croarray
s
pot p
ar
ti
ti
on
i
ng
by au
t
onomo
usl
y o
r
ganisin
g mapsth
rou
gh c
on
t
our mo
del
(
Karthik
S
A
)
749
2
.
1.
Formul
at
i
on
of
act
ivec
ont
our mo
dels
The
C
-
V
m
o
de
l
and
M
-
S
m
od
el
[34
,
35
]
are
the
well
known
m
a
the
m
at
ic
a
l
m
od
el
s
fo
r
se
gm
entat
i
on
base
d
on
re
gion.
C
-
V
m
od
el
is
based
on
r
edu
ci
ng
a
ppr
oxim
a
ti
on
funct
ion
,
de
velo
ped
in
su
c
h
a
wa
y
that
it
’s
m
ini
m
u
m
reache
d
by
cl
ose
ap
pro
xim
ati
on
of
e
xact
e
dges
of
di
ff
e
rent
reg
i
on
s
.
T
he
ene
rg
y
f
un
ct
i
on
is
giv
e
n by
:
E
CV
(
C
r
)
=
μ
∙
le
n
(
C
r
)
+
ν
∙
Are
a
(
C
r
)
+
λ
+
∫
(
Sp
(
x
)
−
c
+
(
C
r
)
)
2
dx
2
C
r
in
+
λ
−
∫
(
Sp
(
x
)
−
c
−
(
C
r
)
)
2
2
C
r
ou
t
dx
(1)
Fig
ur
e
3
s
ho
ws
re
presentat
ion
c
onto
ur
,
w
her
e
,
C
r
is
the
con
t
our
a
nd
Sp
(
x)
is
a
m
icr
oa
rr
ay
s
pot
su
ch
t
hat
Sp
(
x)
∊ℝ
.
S
p(x)i
s
the
gr
ey
le
ve
l
value
of
t
he
spot
de
no
te
d
by
locat
io
n
of
the
pictu
re
el
e
m
ent
in
m
ic
ro
ar
ray
sp
ot
de
note
d
by
Ω.
Param
e
ter
of
re
gula
rizat
ion
µ≥
0
hel
ps
in
con
t
our
sm
oo
t
hn
e
ss
i
ns
id
e
and
outsi
de
area
of
the
m
ic
ro
ar
ray
sp
ot
is
de
note
d
by
C
rin
an
d
C
rout
of
C
r
.
ν
≥
0
an
oth
e
r
p
enali
zi
ng
par
a
m
et
er
fo
r
huge
area
offoref
r
on
t
regi
on
of the
spot.
O
ne
ca
n wr
it
e:
c
+
(
C
r
)
=
Expecta
t
ion
(
Sp
(
x
)
|
x
∈
C
rin
)
c
−
(
C
r
)
=
Expecta
t
ion
(
Sp
(
x
)
|
x
∈
C
rin
)
(2)
Figure
3. Re
pr
esentat
ion o
f
c
on
t
our
In
(
2
)
re
pr
e
se
nts
e
xp
ect
e
d
va
lues
of
inten
s
it
ie
s
of
a
rea
i
nsi
de
a
nd
outsi
de
t
he
c
urve
.
λ
+
an
d
λ
-
≥
0
wh
ic
h
c
on
t
ro
l
t
he
in
flue
nce
of
the tw
o
im
age
ene
rg
y t
erm
s
∫
(
(
Sp
(
x
)
−
c
+
(
C
r
)
)
2
)
dx
4
C
r
in
A
nd
∫
(
(
Sp
(
x
)
−
c
+
(
C
r
)
)
2
)
dx
4
C
r
ou
t
(3)
In
a
f
or
m
ulatio
n
of
le
vel
s
et
C
r
is
exp
re
ssed
as
n
ull
set
fu
nctio
n
Ø:
Ω→
ℝ
:
C
r
={x
∊
Ω:
Ø(x)
=
0}
.
Ar
ea
inside
and
ou
tsi
de
the
s
po
t a
rea
of co
nt
our
al
so ex
pre
ssed
as
C
rin
=
{
x
∈
Ω
:
Ø
(
x
)
>
0
}
C
rout
=
{
x
∈
Ω
:
Ø
(
x
)
<
0
}
(4)
Dista
nce
functi
on to
e
xpres
s t
he
c
on
t
our Ø(
x) as
Eucli
dia
n dist
ance
is
gi
ve
n by:
(5)
In
le
v
el
s
et
f
or
m
ulati
on
of
C
-
V
m
od
el
,
the
reli
ance
of
C
+
(C
r
)
a
nd
C
-
(C
r
)
on
Ø
de
pe
nds
on
m
ini
m
iz
at
ion
of e
q(1)
w
hich
is d
one
by appl
y
ing
gr
a
dient
de
c
ent lea
di
ng to follo
wing
P
DE
.
∂
∅
∂
t
=
δ
(
Ø
)
[
μ
∇
∙
(
∇
∅
‖
∇
∅
‖
⁄
)
−
ν
−
λ
+
(
(
Sp
−
c
+
(
Ø
)
)
2
)
+
λ
−
(
(
Sp
−
c
+
(
Ø
)
)
2
)
]
(6)
|
|.|
|
is
no
rm
of
eucli
dian
a
nd
δ
(.
)
is
the
Dir
ac
gen
e
rali
zed
f
unct
ion.
To
kee
p
the
le
vel
set
functi
on
sm
oo
th,
µ
is
us
e
d.
To
c
on
tr
ol the g
r
ow
i
ng
rate of
the evol
ving conto
ur
term
ν
is u
sed.
λ
+
and
λ
−
can b
e g
ene
rali
zed as
inne
r
and outer
for
ce
s that f
orce t
h
e
conto
ur to
ward the
d
e
finite
s
po
t
’s bor
der
li
ne.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
1
,
Febr
uar
y
2020 :
74
6
-
756
750
2.2.
Spot se
gmen
t
at
i
on
u
sing
A
OM
and
C
h
an
-
V
e
se
m
od
el
(AOM
-
C
V
mo
del
)
I
n
A
OM
-
C
V
m
od
el
inv
ol
ve
s
tw
o
sta
ge
s
nam
el
y,
trai
nin
g
a
nd
te
sti
ng
sta
ge.
In
the
tr
ai
nin
g
sta
ge
,
a
set
of
points
are
e
xtracte
d
from
an
ed
g
e
detect
ion
proc
edure
from
the
give
n
im
age
are
us
e
d
as
th
e
in
put
vecto
r.
T
he
w
ei
gh
ti
ng
vect
or
is
assigne
d
init
ia
ll
y
and
then
update
d.
The
trai
ni
ng
s
ta
ge
is
fo
ll
ow
ed
by
the
te
sti
ng
sta
ge
wh
e
re
the
ed
ge
points
are
c
on
t
ro
ll
ed
acc
ordin
g
to
a
PD
E.
Finall
y,
the
m
od
el
is
i
m
ple
m
ented
by an al
gorith
m
.
2.2.1.
Tr
ainin
g
sessi
on
Durin
g
a
tr
ai
ni
ng
sessio
n
by
sel
ect
ing
a
a
ppropr
ia
te
nu
m
ber
of
neur
on
s
[
35,
36]
a
nd
t
opol
og
y
[37]
for
A
OM
the
intensit
y
SP
in_tr
(x
t
)
of
ar
bitrari
ly
cho
sen
pix
e
l
x
t
of
trai
ning
i
m
age
ta
ken
as
AO
M
in
pu
t
at
tim
e
t=
0,
1,
…
….t
max_
tr.
Let
t
m
ax_
tr
be
the
it
erati
ng
val
ue
for
trai
ning
of
AO
M.
S
ubseq
uen
tl
y,
neur
ons
a
re
auto
no
m
ou
sly
arr
a
ng
e
d
t
o
m
ai
ntain
it
s
to
polog
ic
al
st
ru
ct
ure.
E
ver
y
ne
uro
n
n
is
a
sso
ci
at
e
d
to
the
wei
gh
vecto
r
+
of D dim
ensio
n on
ce
init
ia
li
zat
ion
is c
om
plete
d
wei
gh
t
+
r
ul
e can be
w
ritt
en
as
:
+
(
+
1
)
=
(
)
+
(
)
ℎ
(
)
[
(
)
−
(
)
]
(7)
Ra
te
of
le
ar
ning
is
sym
bo
li
cally
rep
rese
nted
by
η(t),
t
he
fea
s
ible
m
a
tc
hin
g
un
it
(
FMU
)
of
nu
m
ber
of
neur
on
b
a
r
ound
ke
r
nel
nei
ghbo
urh
ood
de
note
d
a
s
h
bn
(t).
To
optim
iz
e
the
weig
hts
+
(
)
,
the
te
rm
η(t)
and
h
bn
(t)
m
us
t
be
dev
el
op
e
d
as
f
un
ct
io
n
of
ti
m
e
decr
easi
ng.
On
ce
trai
ning
session
of
ne
uro
n
is
com
pleted
,
on
e
can a
ppr
ox
im
a
te
ly
m
ap
pix
el
intensit
ie
s d
ist
r
ibu
ti
on to
wei
ghts
of
FMU
.
T
her
e
fore c
hoic
e of lear
ning
ra
te
is
(
)
=
0
(
)
exp
(
−
)
(8
)
0
is
the
init
ia
l
ra
te
of
le
a
rn
i
ng
0
≥
0
an
d
>
0
is
a
tim
e
const
ant
w
he
re
as
h
bn
(t
)
c
hoose
s
a
f
unct
ion
of
Gau
s
sia
n
ce
ntr
ed on ne
uro
ns
assum
e the form
ℎ
(
)
=
(
−
‖
−
2
2
(
)
‖
2
)
(9
)
Wh
e
re
r
b
a
nd
r
n
∊ℝ
are
ad
dre
ss
of
vect
or
s
i
n
the
ne
ur
al
m
ap
of
ou
t
pu
t
of
ne
uro
n
b
an
d
n.
ti
m
e
decr
easi
ng
neig
hbour
hood
of r
a
dius r
(t)>
0
is
giv
e
n by
r
o
i
niti
al
n
ei
ghbo
urh
ooda
nd
(
)
=
0
(
−
)
(10)
Th
us
, val
ue
s
of w
ei
ghts
+
(
+
1
)
are
obta
ined
in
the
tr
ai
nin
g sessi
on.
2.2.2.
Te
s
tin
g
session
Af
te
r
su
cce
ssf
ul
com
pleti
on
of
t
rainin
g
ses
sion
of
A
OM
-
CV
m
od
el
,
trai
ned
net
work
is
de
plo
ye
d
t
o
te
sti
ng
ph
ase
with
ty
pical
m
i
cro
a
rr
ay
s
pots
for
the
cu
r
ve
pro
gr
e
ssio
n
of
C
r
.
At
the
ti
m
e
of
ev
ol
ution
of
c
urve
forefro
nt
im
age
intensit
y
an
d
bac
kgrou
nd
i
nt
ensity
say
E(Sp(
x)
|
x
∊
C
r_in
)
and
E(Sp
(x)
|
x
∊
C
r_out
)
res
pe
ct
ive
l
y
are
passe
d
as i
nput to
the t
rained net
wor
k.
F
or eve
ry n
e
uro
n
the
qua
ntit
ie
s ar
e
def
i
ned as
+
(
)
=
|
−
(
(
)
|
∈
)
|
−
(
)
=
|
−
(
(
)
|
∈
)
|
(11)
+
(
)
and
−
(
)
a
re
t
he
di
sta
nces
of
sub
sidiary
prototy
pe
w
n
f
ro
m
m
ean
inte
ns
it
y
of
cu
rr
e
nt
appr
ox
im
at
ion
s
of
f
or
e
fro
nt
of
a
spot
an
d
backg
rou
nd
of
a
spot
an
d
it
erati
vely
com
pu
te
d
durin
g
t
est
ing
session.
T
he
n,
on
e
can
d
e
fine
the tw
o
set
s,
{
+
(
)
}
=
{
:
+
(
)
≤
−
(
)
}
{
−
(
)
}
=
{
:
+
(
)
>
−
(
)
}
(12)
+
(
)
=
|
{
+
(
)
}
|
a
nd
−
(
)
=
|
{
+
(
)
}
|
wh
ic
h
are
associat
ed
pro
toty
pes
of
set
of
neur
on
.
Accor
ding
to
t
heir
e
xp
ect
e
d
value
of
inten
s
it
y
su
ch
protot
ypes
are
c
hose
n
as
forefr
on
t
area
an
d
bac
kgr
ound
area
of
spot.
F
un
ct
io
nal
of AOM C
-
V
m
od
e
l i
s g
ive
n by
:
_
(
)
=
+
∫
+
(
,
)
+
−
∫
−
(
,
)
0
0
(13
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Mi
croarray
s
pot p
ar
ti
ti
on
i
ng
by au
t
onomo
usl
y o
r
ganisin
g mapsth
rou
gh c
on
t
our mo
del
(
Karthik
S
A
)
751
Wh
e
re,
im
age en
er
gy term
s ar
e
+
(
,
)
=
∑
(
(
)
−
+
(
)
)
2
=
1
…
…
+
(
)
(14)
−
(
,
)
=
∑
(
(
)
−
−
(
)
)
2
=
1
…
…
−
(
)
(15)
Wh
e
re,
par
am
et
ers
λ
+
,
λ
-
≥
0,
a
re
the
ass
oc
i
at
ed
weig
hts
of
s
pot
ene
rgy
te
r
m
s.
Fo
r
t
he
le
vel
set
f
unct
ion
rep
la
ce C
r
by
Ø
the
n
_
(
Ø
)
=
+
∫
+
(
,
Ø
)
+
−
∫
−
(
,
Ø
)
0
Ø
<
0
0
Ø
>
0
(16)
The
e
xpli
ci
t reli
ance of e
+
an
d e
-
on le
vel set
functi
on
Ø
is
ba
sed o
n Heavis
ide ste
p f
un
ct
i
on H
:
ℝ
→R
(17)
By
r
ew
riti
ng
e
qu
at
io
n o
f
E
AO
M_C
V(Ø)
_
(
Ø
)
=
+
∫
+
(
,
Ø
)
(
(
∅
(
)
)
)
+
−
∫
−
(
,
Ø
)
(
1
−
(
(
∅
(
)
)
)
0
0
(18)
Finall
y con
t
our
evoluti
on is
gi
ven b
y
∅
=
∅
[
+
+
+
−
−
]
(19)
Fr
om
(19
),
ca
n be
so
l
ved it
era
ti
vely
u
sin
g
th
e sam
e s
m
oo
thing
a
nd
discret
iz
at
ion
tech
niques.
2.3.
Algori
th
m
for
AOM
C
V
im
plem
ent
at
i
on
Algorithm
f
or
AO
M C
V
m
odel
i
m
ple
m
entation
1.
In
ti
al
iz
at
ion
of in
pu
t
par
am
et
er
In
ti
al
iz
at
ion
of input
par
am
eter
s
a.
Train
_spo
t=
“i
m
age.b
m
p”
an
d
Test
_spo
t=
“
i
m
age1
.b
m
p”
b.
No.
Of
Iterati
on a
nd Max
. I
te
rati
on t
m
ax_t
r
and
t
m
ax_evol
c.
Global
se
gm
entat
ion
=σ, lea
rni
ng
rate a
nd r
a
diu
s
of m
ap
η
and r
o
d.
Ti
m
e
con
sta
nts,
weig
hts
of
e
nergy
and
bin
a
ry
approxim
ati
on
of
le
vel
set
fu
nctio
n:
(τ
n
and
τ
r
),
(λ
+
a
nd λ
-
)
and
ρ
2
.
Neur
on pro
t
otype
In
ti
al
iz
at
ion:
Tr
ai
ning
Sess
ion
Do
3
.
To dete
rm
ine FM
U of ne
uro
n
to
inten
sit
y I
int_tr
(x
t
)
c
hoose
a p
ixel
x
t
i
n
a
dom
ai
n
of
s
pot
Ω :
4
.
Mod
el
updatin
g usin
g
e
quat
io
n (7),
(8),(9
)
a
nd
(10)
5
.
Wh
il
e(t
m
ax_tr
=t
m
a
x_evol
)
6
.
Test
ing
Sessio
n
In
ti
al
iz
at
ion
of level
set f
unct
ion i
n
t
he do
m
ai
n
of s
pot Ω
wi
th bo
unda
ry Ω
’
w
ho
se
s
ub
set
is Ω
0
7
.
If(x
∊
Ω
0
/Ω
’
)
th
en x=ρ
8
.
Else
if (x
∊
Ω’
) t
he
x=
0
9
.
Othe
rw
ise
x=
-
ρ
Do
10.
If
E
AOM C
-
V
has
been sel
ect
ed
t
hen
11
.
Ca
lc
ulate
+
an
d
−
f
or
e
ve
ry n
e
uron
12
.
Grow the
cont
our usin
g Ø
13
.
Else
co
m
pu
te
+
a
nd
−
14
.
En
d
if
15
.
Update c
urren
t
Ø
t
o bina
ry u
s
ing
Ø←ρ
(H(Ø)
-
H(
-
Ø
))
16
.
Stop ev
olu
ti
on
if t
m
ax_
tr
reached.
Fr
om
this alg
ori
thm
, o
ne
ca
n im
ple
m
ent this u
si
ng MATL
AB.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
1
,
Febr
uar
y
2020 :
74
6
-
756
752
3.
E
X
PERI
MEN
TAL STU
D
Y AND
OUT
COME
S
In
the
e
xp
e
rim
ental
stud
y
,
th
e
rejoin
der
o
f
the
pu
t
f
orwa
r
ded
a
ppr
oach
i
s
est
i
m
a
te
d
on
a
var
ie
ty
of
m
ic
ro
arr
ay
s
po
ts.
Im
ages
are
captu
red
from
GEO
-
DB
.
Depend
i
ng
on
ty
pe
of
spot
f
our
tup
es
of
im
age
are
analy
zed
.
T
he
four
s
po
ts
are:
ty
pical
ly
cl
ean
sp
ot,
s
po
t
wit
h
arti
facts,
do
nut
sp
ot
a
nd
noi
sy
sp
o
t.
I
n
this
wo
r
k
MATLAB
(R
2014
b)
is
us
e
d
f
or
im
age
analy
sis
as
it
is
a
hi
gh
pe
rfor
m
anc
e
too
l.
To
qua
ntify
the
le
vel
exact
par
ti
ti
on
i
ng
of
fo
re
fro
nt
par
t
fr
om
rear
part
reg
io
n,
s
up
e
r
iority
of
the
s
po
t,
var
i
ou
s
c
on
si
der
at
io
ns
s
uch
a
s
Total
im
age Em
inence in
de
x (TIE
I), c
oeffici
ent of
determ
inati
on
r
2
, c
on
c
orda
nce c
orrelat
ion
p
c
, a
re
us
e
d.
3.1.
Total
im
age e
mi
nence inde
x
(TIEI)
To
m
ake
a
dist
incti
on
t
he
trai
ned
an
d
te
sta
bl
e
spot
this
par
a
m
et
er
gauge
is
util
iz
ed
[
32
]
.
I
n
descr
i
bed
m
et
ho
dolo
gy
par
ti
ti
on
e
d
s
pot
is
al
ie
nated
into
co
rr
el
at
ion
l
os
s,
lum
inance
dist
or
ti
on
an
d
m
or
phol
og
ic
a
l
con
t
rast is s
pec
ifie
d by
(
,
)
=
(
,
)
(
)
∙
(
)
(20)
(
,
)
=
2
∙
(
)
∙
(
)
(
(
)
)
2
+
(
(
)
)
2
(21)
(
,
)
=
2
∙
(
,
)
(
(
)
)
+
(
(
)
)
(22)
wh
e
re
ave
rage
outl
ine
of
cu
rv
at
ur
e
a
nd
pa
rti
ti
on
ed
c
urva
ture
,
s
t
(i
)
a
nd
s
t
(j)
a
re
no
rm
a
l
di
vergen
c
e
of
spot
a
nd
pa
rtit
ion
ed
spot,
s
t
(i,j)
is
c
o
-
vari
ance
of
trai
ne
d
c
urvatu
re
a
nd
par
ti
ti
on
e
d
c
urvatu
re.
I
n
(
20)
i
s
a
m
easur
e
w
hich
c
om
pu
te
s
th
e
relat
ion
sh
i
p
betwee
n
in
put
sp
ot
a
nd
se
gme
nted
spot.
Ra
ng
e
of
c
o
r
rloss
(i,
j
)
i
s
[
-
1,
1].
Be
st
va
lue
f
or
this
pa
r
m
et
er
is
1.
L
u(i
,
j)
value
li
es
betwee
n
0
a
nd
1
wh
ic
h
cl
ea
rly
m
easur
es
cl
ose
nes
s
value
of
l
um
in
ance
bet
ween
s
po
t
ta
ke
n
as
i
nput
a
nd
portio
ned
s
pot.
Mo
rpho
l
og
ic
al
co
ntr
ast
s
betwee
n
im
ages
are
m
easur
ed
by
S
t
(i,
j
)
it
s
range
of
value
is
betwee
n
0
and
1.
G
ood
value
for
L
u(
i,
j
)
a
nd
St(i,
j
)
is
1.
Ultim
at
ely To
ta
l Im
age Em
inence in
de
x (TI
EI) co
nvey
ed
a
s
[
38, 39]
.
=
(
,
)
∗
(
,
)
∗
(
,
)
(23)
(
,
)
=
4
∙
∙
∙
(
,
)
(
+
)
2
+
(
(
)
)
+
(
(
)
)
(24)
In
ge
ner
al
ch
oi
c
e
of
this
para
m
et
er
is
[
-
1,
1].
As
s
how
n
i
n
T
a
ble
1
s
ho
ws
sta
ti
sti
cal
assessm
ent
of
T
IEI
f
or
sp
ot
p
a
rtit
ion
.
Wh
e
n TIE
I
is e
qu
al
t
o un
it
y
or al
m
os
t u
nity
sign
i
fy b
est
qual
it
y of
s
po
t,
-
1
s
i
gn
i
fyw
or
st
qu
al
it
y.
Table
1.
Stat
ist
ic
al
assessm
ent
s of T
IEI f
or s
po
t
pa
rtit
ion
[38,
39]
TI
E
I
o
f
a
rang
e of
sp
o
ts
Sp
_
id
[16
]
C
-
m
eans
b
y
f
u
zzy
Morp
h
_
Section
alizatio
n
AOM
-
CVtech
n
iq
u
e
Sp
o
t8
0
.41
8
0
.62
5
0
.95
0
Sp
o
t9
0
.44
0
0
.35
1
0
.48
3
Sp
o
t6
0
.32
0
0
.62
1
0
.79
2
Sp
o
t1
2
0
.46
0
0
.32
6
0
.61
6
Sp
o
t1
3
0
.55
0
0
.71
0
0
.73
3
In
order
to
m
ake
a
rev
ie
w
of
se
gm
entat
i
on
process
th
e
fo
ll
owin
g
m
et
rics
are
cal
culat
ed
f
ro
m
the sim
ulate
d
im
ages:
The
coeffic
ie
nt
of
de
termi
na
ti
on
r
2
in
di
cat
es
the
strong
point
of
the
li
near
in
vo
lve
m
ent
between
si
m
ulate
d
an
d
cal
culat
ed
s
pots,
as
well
as
,
it
giv
e
the
pro
portio
n
of
the
fluctuati
on
of
the
cal
cul
at
ed
data
[40
]
.
2
=
∑
(
(
)
−
)
2
=
1
∑
(
−
̅
)
2
=
1
(25)
Wh
e
re
I
segm
ented
and
I
actual
are
t
he
m
ean
intensit
y
values
of
the
cal
culat
ed
and
sim
ulate
d
sp
ots,
res
pecti
vely
.
Her
e
i
re
fers
to
ind
ivi
du
al
cel
l
i
m
ages
(
i
=
1….3
24),
a
nd
̅
is
the
over
al
l
m
ea
n
of
the
f
oreg
r
ound
of
al
l
sp
ots
in
sim
ula
te
d
im
age.
The
algorit
hm
that set
s
r
2
val
ue
cl
os
est
to
the
unit
y has the
b
est
perform
ance.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Mi
croarray
s
pot p
ar
ti
ti
on
i
ng
by au
t
onomo
usl
y o
r
ganisin
g mapsth
rou
gh c
on
t
our mo
del
(
Karthik
S
A
)
753
The
conc
orda
nce
correl
at
i
on
p
c
m
easur
es
the
ag
reem
ent
betwee
n
sim
ul
at
ed
an
d
cal
c
ul
at
ed
data
a
nd
is
us
e
d
to
ev
al
uate the
reprod
uci
bili
ty
o
f
the
pr
opos
e
d
se
g
m
entat
ion
m
et
ho
d
[4
1
]
.
=
2
∗
∗
2
+
2
+
(
−
)
2
(26)
Wh
e
re
A
a
nd
B
are
tw
o
sam
ples,
an
d
μ
B
are
the
m
ean
value
s,
and
are
the
st
and
a
r
d
de
viati
on
of
the
sam
ples.
p
c
value
de
ci
des
t
he
pe
rfor
m
ance
o
f
al
gorithm
,
hig
he
r
the
p
c
values
te
ll
s
that
bette
r
per
f
or
m
ance
of
a
ppr
oac
h.
Com
par
isi
on
o
of
e
xisti
ng
m
eth
od
with
our
t
echn
i
qu
e
usi
ng
assessm
ent
pa
ram
et
ers
is
sh
own
in
T
able
2.
Table
2.
C
om
par
iso
n of exist
ing m
et
ho
d wit
h our tec
hn
i
que u
si
ng assess
m
ent p
aram
et
e
rs
I
m
ag
e
Id
C
-
m
eans
b
y
f
u
zzy
(
FCM)
Morp
h
_
Section
alizatio
n
(M
o
Seg
)
AOM
-
CV tech
n
iq
u
e
2
2
2
Sp
o
t
0
.86
1
2
0
.74
8
0
0
.86
3
0
0
.91
9
0
0
.98
3
5
0
0
.99
1
7
Sp
o
t8
0
.87
2
1
0
.67
2
7
0
.57
3
2
0
.63
2
8
0
.97
2
1
0
.98
5
7
Sp
o
t9
0
.68
4
0
0
.49
0
0
0
.76
9
3
0
.85
2
5
0
.96
0
7
0
.98
0
1
Sp
o
t6
0
.43
0
6
0
.18
4
8
0
.68
2
9
0
.76
8
2
0
.89
7
2
0
.94
2
2
Sp
o
t1
2
0
.80
5
5
0
.72
6
1
0
.57
0
6
0
.63
5
6
0
.89
8
1
0
.93
7
2
Sp
o
t1
3
0
.81
0
7
0
.63
1
5
0
.73
3
5
0
.80
3
3
0
.99
3
8
0
.99
6
8
Figure
4
sho
w
s
boxplot
com
par
isi
on
of
T
I
EI.
Bo
x
plo
t
of
Fu
zzy
C
-
m
eans
m
et
ho
d
is
c
om
par
at
ively
sh
ort
-
wh
ic
h
s
ugge
sts
that gene
rall
y
sp
ot
’s
to
ta
l
i
m
age
e
m
inence
in
de
x
[
40
-
4
2
]
ha
ve
a
hi
gh
le
vel of
a
gr
e
e
m
ent
with
each
oth
e
r
but
fail
s
to
re
ach
cl
os
est
to
un
it
y.
Mor
pho
-
segm
entat
ion
m
et
ho
d
box
co
m
par
at
ively
tal
l
and
interp
rets
dif
fe
ren
t
im
m
inence
ind
e
x
val
ue
wh
e
n
c
om
par
ed
to
oth
e
r
spot
s.
A
OM
-
C
V
m
et
hod
bo
x
pl
ot
sh
ow
s
cl
os
e
val
ue
o
f
im
m
inence
ind
e
x
to
1
w
hich
is
a
pro
m
isi
ng
res
ult
to
judge
t
he
qu
al
it
y
of
t
he
spot
.
Fr
om
Fig
ur
e
5
and
Fi
gure 6
it
is cle
ar th
at
the
re is close ap
pr
ox
im
at
ion
o
f
A
OM_C
V
m
e
tho
d
to
wards th
e
b
est
perform
ance
of
sp
ot
po
rtion
i
ng
,
com
par
ed
to
FCM
m
et
ho
d
an
d
Moseg
m
et
ho
d
.
Fi
gure
s
5
and
6
dem
on
strat
es
our
e
xperim
ental
w
or
k wit
h diff
e
ren
t t
ypes
of
spot.
Figure
4
.
Box
-
plo
t c
om
par
iso
n of TIE
I
.
Figure
5.
Com
par
is
on of
2
Figure
6.
Com
par
is
on of
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
1
,
Febr
uar
y
2020 :
74
6
-
756
754
Figure
7
s
how
s
ex
per
im
ental
stud
y
on
var
i
ous
ty
pes
of
s
pots.
For
t
he
pur
po
s
e
of
e
xperi
m
ental
stud
y
var
i
ou
s
ty
pes
of
spots
a
re
c
on
si
der
e
d.
A
O
M
-
CV
a
ppr
oa
ch
works
well
in
al
l
cases
i
nc
lu
ding
no
ise
,
do
nu
t
sh
a
ped
s
pot.
M
or
e
ov
e
r
it
identifie
s
c
on
t
our
in
s
pots
(
donut
spot)
.
Finall
y
fo
re
fro
nt
area
an
d
bac
kgr
ound
reg
i
on clea
rly
diff
e
re
ntiat
ed.
Spot T
yp
e
Te
st
Sp
ot
Spot P
or
tion
i
ng
by
AOM
-
CV
me
t
ho
d
Fore
ground
Extr
act
i
on
Back
ground
Extr
act
i
on
Ty
pical
Spot
Spot with
Ar
ti
facts
Don
ut S
po
t
No
isy
Spot
Figure
7.
Ex
pe
rim
ental
ou
tc
om
e
4.
CONCL
US
I
O
N
In
this
w
ork,
a
novel
m
et
ho
d
of
s
po
t
porti
on
i
ng
im
ages
i
s
done.
Fir
st,
the
te
sti
ng
of
ne
uro
ns
an
d
the
trai
ni
ng
proces
s
of
ne
uro
ns
is
c
ar
ried
ou
t
us
i
ng
a
uto
m
at
ic
or
ga
nizing
m
aps
(
AO
M
).
T
his
m
et
ho
d
consi
ders
the
aver
a
ge
i
ntensi
ty
values
in
sid
e
an
d
outsi
de
the
c
urvatu
re
of
the
s
pot.
E
vo
luti
on
of
co
ntour
is
done
by
e
dg
e
m
ap.
Fr
om
this
on
e
ca
n
easi
ly
disti
n
gu
is
h
the
f
oref
r
on
t
area
from
the
rear
reg
i
on.
Now,
this reso
l
ves
t
he
c
halle
nges
c
reated b
y t
he
ot
her
m
et
h
od
s
. A
OM
-
CV
ap
proac
h
is
al
so
a
ppli
cable
im
ages
that
con
sist
of
arb
it
ra
ry
dev
ia
te
d
of
pix
el
s
al
so
.
T
hro
ugh
co
nductin
g
t
est
s
and
ob
ta
i
ned
r
esults
total
pro
cedure
is
dynam
ic
,
in
the
prese
nce
of
ar
bitra
ry
de
vi
at
ion
of
pi
xels,
topol
og
ic
al
c
ha
ng
e
s,
a
nd
ir
re
gu
la
rity
of
spo
t
and
m
or
phologica
l changes
.
REFERE
NCE
S
[1]
X.
Xie
and
G.
Beni
,
“
A
Vali
dity
m
ea
sure
for
Fu
zzy
Cluste
r
ing
,”
IEE
E
Tr
ansacti
o
ns
on
Pat
te
rn
Anal
ysis,
machine
Inte
ll,
vo
l.
13(8)
,
pp.
841
-
847,
19
91
.
[2]
“
Introduc
ti
on
to m
ic
roa
rra
y
,
”[Onl
ine
]
,
Ava
il
ab
le:
ww
w.c
eb
it
a
c.
u
nibi
elefe
ld
d
e/
gr
oups/brf/
softwar
e/
emm
a
.
[3]
Schena
M,
Shal
on
D,
Davis
R
W
and
Brown
P
O
,
“
Quanti
ta
t
ive
m
onit
oring
of
gene
expr
ess
ion
pat
t
ern
s
with
a
complemen
ta
r
y
DN
A m
ic
ro
arr
a
y
,”
S
cienc
e
,
vo
l. 270(5235), pp.
467
-
70,
1995
.
[4]
Buhle
r
J,
Ide
k
e
r
T,
&Ha
y
nor
D,
Dapple
,
“
Im
prove
d
te
ch
nique
s
f
or
findi
ng
spots
on
DN
A
m
ic
roa
rr
a
y
s
,”
(UW
CS
E
Te
ch)
,
W
ashingt
on
UW
TR
depa
r
tm
ent
of com
put
er
sci
ence and
E
ng,
Univer
si
t
y
of
W
ashingt
on,
p
p
.
198
-
207
,
2000
.
[5]
Buckl
e
y
M
.
J
,
“
Spot use
r’s gui
d
e
,”
S
y
dn
e
y
,
Aus
t
ral
i
a:
CSI
RO
Ma
the
mati
cal
and
I
nformation
Sc
ience
s
,
2000
.
[6]
W
ang
X,
Gho
sh
S,
Guo
SW
,
“
Quanti
tati
ve
qua
li
t
y
con
trol
in
m
ic
ro
ar
ra
y
image
pro
ce
ss
ing
and
da
ta
ac
quisit
ion,
”
Nuc
le
i
c
A
ci
ds
Re
sea
rch
,
vol
.
29(15
),
pp.
e75
,
2001
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Mi
croarray
s
pot p
ar
ti
ti
on
i
ng
by au
t
onomo
usl
y o
r
ganisin
g mapsth
rou
gh c
on
t
our mo
del
(
Karthik
S
A
)
755
[7]
E.
K.
Lobe
nho
f
er,
P.
R.
Bushe
l,
C.
A.
Afs
har
i,
and
H.
K.
Ham
ade
h,
“
Progress
in
the
appl
i
ca
t
ion
of
DN
A
m
ic
roa
rra
y
s
,”
En
vi
ronm
ent
Hea
lth Pe
rs
pectives
,
v
ol.
1
09
(
9
)
,
pp
.
8
81
–
891,
2001
.
[8]
Yang
YH
,
Buckl
e
y
MI,
Dudoit
S,
Speed
TP
,
“
Com
par
ison
of
m
et
hods
for
image
anal
y
s
is
on
c
DN
A
m
ic
roa
rr
a
y
dat
a
,”
Journal
of
Computati
onal
a
nd
Gr
aphic
al
S
t
ati
stic
s
,
vol
.
11(
1)
,
pp
.
108
–
36
,
2002
.
[9]
Boz
inov
D,
Rah
nenf
ühre
r
J,
“
Uns
uper
vised
te
chn
ique
for
robust
t
arg
et
sep
ara
t
ion
and
ana
l
y
sis
of
DN
A
m
ic
roa
rr
a
y
spots throu
gh
ad
apt
iv
e
pix
el c
lus
te
ring
,”
B
ioi
nfor
matic
s
,
pp
.
747
–
56,
2002
.
[10]
Ahm
ed
AA
,
Vias
M
,
I
y
er
NG
,
Cal
das
C,
Br
enton
JD
,
“
Microa
rra
y
segm
ent
atio
n
m
et
hods
signifi
ca
n
tly
inf
lue
nc
e
dat
a
p
re
ci
sion
,”
Journal
of
Nucle
ic
Ac
ids
Re
searc
h
.,
vol.
32(5)
,
pp
.
e50
,
2004
.
[11]
M.
B.
Ei
sen
,
“
Sc
anAl
y
ze
, “
[Onli
ne]
,
Available:
h
tt
p://r
an
a.
Stanfor
d.
EDU/software
/
for
software
and
documenta
t
ion.
[12]
K.
Ble
kas,
N.
P.
Ga
la
tsanos
,
an
d
I.
Georgi
ou,
“
An
unsupervise
d
art
ifa
ct
cor
r
ection
appr
oa
ch
for
the
ana
l
y
sis
of
DN
A m
ic
roa
rra
y
images
,”
in
Pro
c
.
I
EEE
Int
.
Con
f.
Image
Proc
ess.
(
ICIP
)
,
vol. 2,
pp.
165
–
168
,
20
04.
[13]
Dee
pa
J
,
T
essam
m
a
Thomas
,
“Automati
c
Seg
m
ent
at
ionof
DN
A
Microa
rra
y
I
m
age
s
using
an
Im
prove
d
see
d
ed
reg
ion
growing
m
et
hod
,”
World
Congress
on
Nat
ure
&
Bi
ologi
cally
Inspired
Computing
(
NaBIC)
,
IEE
E
,
2009
.
[14]
T.
Srinark
and
C.
Kam
bhamettu,
“
A
m
ic
roa
rra
y
imag
e
an
aly
si
s
sy
st
em
base
d
on
m
ult
ipl
e
sna
kes
,
”
Journal
Biol.
Syst.
Sp
ec.
Iss
ue
,
vol. 12, pp. 127
–
157
,
2004
.
[15]
J.
Ho
and
W
.
-
L.
Hw
ang,
“
Automa
tic
m
ic
roa
rra
y
spot
segm
ent
at
io
n
using
a
snake
-
fisher
m
odel
,
”
I
E
EE
Tr
ansacti
ons
on
Me
di
cal Imaging
,
vo
l. 27
(
6
)
,
pp.
847
–
857
,
20
08
.
[16]
K.
Ble
k
as,
N.
G
al
a
tsanos
,
A.
Likas,
and
I.
E
.
L
a
gar
is,
“
Mixtur
e
m
odel
anal
y
s
is
of
DN
A
m
ic
roa
r
ra
y
images
,”
IE
E
E
Tr
ans.
Me
d.
Ima
g
.
,
vol. 24
(
7
)
,
pp
.
901
–
909
,
2005
.
[17]
Athana
siad
is
E
.
I
,
e
t
al
.
,
“
Com
ple
m
ent
ar
y
DN
A
m
ic
roa
rr
a
y
image
proc
essing
base
d
on
th
e
Fuzz
y
Gauss
ia
n
m
ixt
ur
e
m
odel
,”
In
I
EEE
Tr
ansacti
on
on
I
nformation
Tech
nology
in
B
iome
dic
in
e,
vol. 13(
4
),
2009
.
[18]
Cec
c
are
l
li
M
,
A
ntoni
ol
G,
“
A
def
orm
abl
e
gr
id
-
m
at
chi
ng
appr
o
ac
h
for
m
ic
roa
rr
a
y
images
,”
IE
E
E
Tr
ansacti
on
o
n
Image
Proc
essing
,
vol
.
15(10)
,
p
p.
3178
–
88
,
200
6
.
[19]
Rah
nenf
ühre
r
J,
Bozi
nov
D,
“
H
y
brid
c
luste
rin
g
for
m
ic
roa
rra
y
image
ana
l
y
si
s
combining
int
ensity
and
shap
e
fea
tur
es,
”
BMC Bi
oinf
orm
atics
,
pp.
1
–
11
,
2004
.
[20]
W
engGuir
ong,
Su
Jian,
"M
ic
roa
rra
y
Im
age
s
Proce
ss
ing
U
sing
the
Offs
et
Vec
tor
Fiel
d
and
Expe
ctat
i
on
Maximiza
t
ion
A
lgori
thm
,”
I
EE
E
co
nfe
ren
ce
,
201
0.
[21]
Manjuna
th
S
S
and
La
lithaRangara
j
an,
“
Refinem
ent
of
K
-
me
ans
Cluste
ring
for
Segm
ent
at
i
on
of
Microa
rr
a
y
Im
age
s
,”
Journal
of
Con
ve
rgen
ce
and
Information T
ec
hnology
, v
o
l. 6
(
9
)
,
pp
.
403
-
41
1,
2011
.
[22]
J.
Hari
kira
n
,
e
t
al.
,
“
Fuzz
y
C
-
m
ea
ns
with
bi
dimen
sionae
m
pi
ric
a
l
m
ode
decom
positi
on
for
segm
ent
at
ion
of
m
ic
roa
rra
y
imag
e,
”
Int
ernati
onal
Journal
o
f
Com
pute
r Sc
ie
nc
e
,
v
ol.
9
,
pp
.
189
–
19
8,
Sep
2012.
[23]
N
Kar
imi,
S
Sam
avi
,
S.
Shirani
,
“
Segm
ent
at
ion
ofDN
A
m
ic
roa
rr
a
y
Im
ag
es
using
ada
pt
ive
gra
ph
b
ase
d
m
et
hod
,”
I
E
T
I
mage
proce
ss
in
g
, v
ol
.
4
,
pp
.
19
-
27
,
2008
.
[24]
Shenghua
N,
Pa
n
W
ang
,
“
Spotte
d
Cdna
Mi
cro
arr
a
y
Im
age
Se
gm
ent
at
ion
Us
i
ng
ACW
E
,
”
Ro
manian
Journal
Of
Information
Sc
ience
And
Te
chnology
,
v
ol
.
12
(
2
),
pp.
249
-
263
,
20
09.
[25]
Yan
Zha
ng,
Bog
dan
J.
Matusze
w
ski,
Li
k
-
Kw
an
Shark
,
“
Medical
I
m
age
Segm
ent
ation
Us
ing
New
H
y
brid
L
eve
l
-
Se
t
Method,
”
Fi
f
th
Inte
rnational
Co
nfe
renc
e
Bi
oM
e
dic
al
Vi
sual
izat
ion:
Information
Vi
sualizati
on
in
Me
dic
al
an
d
Bi
omedi
cal
In
formatic
s
,
vo
l. 2
,
p
p.
89
-
97
,
2009
.
[26]
Kaustubha
A.
Mendhurwar,
R
aj
ase
kh
arKa
kum
ani
,
Vi
j
a
y
Dev
ab
h
akt
uni
,
“
Microarra
y
image
seg
m
ent
at
ion
using
cha
n
-
vese
a
ct
iv
e
con
tour
m
odel
and
l
eve
l
set
m
e
thod,
”
Annual
I
nte
rnational
Co
nfe
renc
e
of
the
I
EE
E
Engi
n
ee
rin
g
in
Me
d
ic
in
e
and
Bi
ology Society
,
vol.
5
,
pp
.
256
-
2
64,
2009
.
[27]
El
eni
Zacha
r
i
a
and
Dim
it
ris
Marou
li
s,
“
3
-
D
Spot
Modeli
ng
f
or
Autom
at
ic
S
egmenta
t
ion
ofc
DN
A
Microa
r
r
a
y
Im
age
s
,”
I
EE
E
T
rans
act
ions
on
Nanobiosci
en
ce
,
v
ol
. 9
(
3
)
,
2010
.
[28]
Guifa
ng
Shao,
Ti
ngna
W
an
g,
W
u
peng
Hong,
Zhi
gang
Chen,
“
An
Im
prove
d
SV
M
Method
for
eDNA
Microa
rra
y
Im
age
Segm
en
ta
ti
on
,”
The
8
th
Inte
rnational
Confe
renc
e
on
Computer
Sci
en
ce
&
Educ
ati
on
,
2013
.
[29]
Chunm
ing
Li
,
et
al.
,
“
A
Le
ve
l
Set
Met
hod
for
Im
age
Segm
ent
at
ion
in
the
Prese
nce
of
Inte
nsi
t
y
Inhom
ogene
it
i
es
with
Appli
ca
t
ion
to
MRI
,”
IE
EE
Tr
ansacti
ons
on
Image
Proc
essing
,
vol
.
20
,
pp
.
20
07
-
20
17
,
2011
.
[30]
Stamos
Katsigi
anni
s
,
Eleni
Z
ac
h
a
ri
a,
and
Dim
it
ris
Marouli
s,
“
Grow
-
Cut
Based
Autom
at
ic
cDNA
Microa
rra
y
Im
age
Segm
ent
at
ion
,”
I
EEE
Tr
ansacti
o
ns On
Nanobios
ci
en
ce
,
vol
.
14
(
1
)
,
2015
.
[31]
Guifa
ng
Shao,
Sh
unxia
ng
W
u,
Ti
e
jun
Li,
“
c
DN
A
Mieroa
rra
y
Im
age
Segm
ent
at
ion
wi
th
an
I
m
prove
d
Moving
K
-
m
ea
ns Cluste
r
ing
Method
,”
I
E
EE
ICSC
,
2015
.
[32]
Mac
hm
ud
R
Allaham
di,
Ito
W
asito,
“
Im
prove
d
Micro
arr
a
y
image
s
Canc
er
Cla
ss
if
ic
a
ti
onusing
k
-
n
ea
rest
nei
ghbou
r
using pa
rt
ic
l
e
sw
arm
opti
m
iz
a
ti
o
n,
”
I
EEE
conf
er
enc
e
,
IW
BIS
,
20
17
.
[33]
T.
B
arr
e
tt
,
S.
E
.
W
il
hite,
P.
L
ed
oux,
e
t
al
.
,
“
NCBI
GEO:
ar
ch
ive
for
func
t
ion
al
g
enomics
data
sets
—
upda
te,”
Nucl
eic Ac
ids Research
,
vol. 41, pp.
991
–
995
,
20
13.
[34]
Pasca
l
Ge
tre
u
er
,
“
Chan
-
Vese
Seg
m
ent
at
ion
,
”
IPO
L J
ournal
Image
Proce
ss
ing
On
Line
,
vo
l. 2
,
pp.
109
-
129
, 2
014
.
[35]
Moham
m
ed
M.
Abdelsa
m
ea
,
Gi
orgio
Gnec
co,
Moham
ed
Medha
t
Gab
er
,
“
A
SO
M
-
base
d
Chan
–
Vese
m
odel
fo
r
unsupervise
d
i
m
age
segm
ent
a
ti
on
,”
Springer
-
Ve
rlag
Be
rlin
Heide
lb
erg
201
5
,
Sof
t
Compu
t
(
2017)
,
no.
21
,
pp.
2047
–
2067
,
2015
.
[36]
B
y
eongk
eun
Ka
ng,
Truong
Q
N
,
“
Ran
dom
Fores
t
with
Le
arn
ed
R
epr
ese
nt
at
ions
fo
r
Sem
ant
ic
Segm
ent
at
ion
,”
IEEE
Tr
ansacti
ons on Im
age
Proc
essing:
a
Pub
licati
on
of
th
e
I
EE
E
Sign
al
Proc
essing
So
ci
e
ty
,
pp
.
1
-
1
,
20
19.
[37]
Z.
W
ang,
B.
Zi
neddi
n
,
J.
Liang,
e
t
al.
,
“
A
novel
neur
al
net
work
appr
oac
h
to
cDNA
m
i
cro
arr
a
y
ima
ge
segm
ent
at
ion
,
”
Computer
Me
tho
ds and
Program
s in
B
iomedicine
,
vol
.
11
,
pp
.
189
–
198,
Jul 20
13
.
[38]
N.
Gianna
k
ea
sa
,
et
a
l
.
,
“
Segm
ent
at
ion
of
m
ic
ro
a
rra
y
images
usin
g
pixe
l
c
la
ss
ifi
c
at
ion
-
compari
son
withc
lust
eri
ng
-
base
d
m
et
hods,
”
Computers i
n
Biology
and
Me
d
icine
,
vol
.
43
(
6
)
,
p
p.
705
–
716
,
201
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.