I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
4
3
,
No
.
1
,
Ju
ly
2
0
2
6
,
p
p
.
157
~
17
8
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
:
1
0
.
1
1
5
9
1
/ijeecs.v
4
3
.i
1
.
pp
157
-
17
8
157
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
H
a
wkNe
t:
a
n
int
e
llig
ent
bio
-
inspi
re
d optimiza
tion
ba
sed pa
tch
wise a
da
ptive U
-
Net fra
mewo
rk
for ret
ina
l
bloo
d ves
sel
seg
menta
tion in
f
undus
ima
g
es
Sa
ba
Sh
eiba
1
,2
,
M
Nee
la
k
a
nta
pp
a
3
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
B
EST I
n
n
o
v
a
t
i
o
n
U
n
i
v
e
r
si
t
y
,
G
o
r
a
n
t
l
a
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
S
a
n
d
A
I
,
M
u
f
f
a
k
h
a
m Ja
h
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
H
y
d
e
r
a
b
a
d
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
Te
c
h
n
o
l
o
g
y
,
V
a
s
a
v
i
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
H
y
d
e
r
a
b
a
d
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
1
8
,
2
0
2
6
R
ev
is
ed
J
u
n
2
3
,
2
0
2
6
Acc
ep
ted
J
u
n
2
9
,
2
0
2
6
Th
e
a
c
c
u
ra
te
se
g
m
e
n
tatio
n
o
f
re
ti
n
a
l
b
lo
o
d
v
e
ss
e
ls
p
lay
s
a
p
i
v
o
tal
r
o
le
i
n
t
h
e
e
a
rly
d
iag
n
o
sis
a
n
d
m
o
n
it
o
ri
n
g
o
f
h
e
a
lt
h
issu
e
s
li
k
e
d
iab
e
tes
,
h
ig
h
b
l
o
o
d
p
re
ss
u
re
,
a
n
d
g
lau
c
o
m
a
.
Trad
it
i
o
n
a
l
m
a
c
h
i
n
e
lea
rn
i
n
g
tec
h
n
iq
u
e
s
su
c
h
a
s
m
a
tch
e
d
fil
terin
g
,
m
o
r
p
h
o
l
o
g
ica
l
p
r
o
c
e
ss
in
g
,
a
n
d
e
d
g
e
d
e
tec
ti
o
n
h
a
v
e
a
to
u
g
h
ti
m
e
t
o
se
e
t
in
y
b
lo
o
d
v
e
s
se
ls
c
lea
rly
b
e
c
a
u
se
e
y
e
ima
g
e
s
o
ften
h
a
v
e
u
n
e
v
e
n
li
g
h
ti
n
g
,
p
o
o
r
c
o
n
tras
t,
a
n
d
b
l
o
o
d
v
e
ss
e
ls
th
a
t
twist
a
n
d
tu
r
n
i
n
c
o
m
p
lex
wa
y
s.
To
o
v
e
rc
o
m
e
th
e
se
c
h
a
ll
e
n
g
e
s,
th
is
wo
rk
i
n
tr
o
d
u
c
e
s
a
n
imp
ro
v
e
d
P
a
tch
-
wise
U
-
Ne
t
m
e
th
o
d
with
Ha
rris
Ha
wk
o
p
ti
m
iza
ti
o
n
(HH
O).
Th
e
p
a
tch
sy
ste
m
s
p
li
t
fu
n
d
u
s
im
a
g
e
s
in
t
o
o
v
e
rlap
p
in
g
p
a
tch
e
s,
e
n
a
b
li
n
g
t
h
e
n
e
two
rk
t
o
fo
c
u
s
o
n
lo
c
a
li
z
e
d
v
e
ss
e
l
fe
a
tu
re
s
su
c
h
a
s
fin
e
c
a
p
il
larie
s,
b
ifu
rc
a
ti
o
n
s,
a
n
d
v
e
ss
e
l
b
o
u
n
d
a
ries
th
a
t
a
re
o
ften
m
isse
d
in
g
lo
b
a
l
se
g
m
e
n
tatio
n
.
M
e
a
n
wh
il
e
,
HH
O
is
e
m
p
l
o
y
e
d
t
o
fi
n
e
-
tu
n
e
t
h
e
U
-
Ne
t’s
h
y
p
e
r
p
a
ra
m
e
ters
a
n
d
lea
rn
in
g
we
ig
h
ts,
a
c
h
iev
in
g
fa
ste
r
c
o
n
v
e
rg
e
n
c
e
a
n
d
e
n
h
a
n
c
e
d
se
g
m
e
n
tati
o
n
a
c
c
u
ra
c
y
wit
h
o
u
t
m
a
n
u
a
l
t
u
n
i
n
g
.
T
o
e
sta
b
li
sh
g
e
n
e
ra
li
z
a
b
il
it
y
a
n
d
r
e
li
a
b
il
i
ty
,
t
h
e
p
ro
p
o
se
d
fra
m
e
wo
rk
is
rig
o
ro
u
sly
tes
ted
o
n
two
sta
n
d
a
rd
re
ti
n
a
l
ima
g
e
re
p
o
sito
r
ies
DRIV
E,
a
n
d
S
TARE.
Ex
p
e
rime
n
tal
e
v
a
lu
a
ti
o
n
d
e
m
o
n
stra
tes
sig
n
i
fica
n
t
im
p
ro
v
e
m
e
n
t
i
n
k
e
y
p
e
rfo
rm
a
n
c
e
m
e
tri
c
s,
in
c
l
u
d
i
n
g
a
c
c
u
ra
c
y
o
f
9
7
.
8
%
,
p
re
c
isio
n
9
5
.
8
%
,
re
c
a
ll
9
6
.
5
%
,
f
1
-
sc
o
re
9
7
.
3
%
,
I
o
U
o
f
9
6
.
2
%
,
a
n
d
d
ice
c
o
e
fficie
n
t
(D
C)
9
4
.
3
%
,
h
ig
h
li
g
h
ti
n
g
th
e
m
o
d
e
l’s
c
a
p
a
b
il
it
y
t
o
c
a
p
t
u
re
t
h
i
n
a
n
d
t
h
i
c
k
v
e
ss
e
ls
stru
c
tu
re
s
wh
il
e
m
in
imiz
in
g
f
a
lse
d
e
tec
ti
o
n
s.
A
d
d
i
ti
o
n
a
ll
y
,
q
u
a
li
tati
v
e
se
g
m
e
n
tatio
n
fu
rt
h
e
r
c
o
n
firm
t
h
a
t
t
h
e
p
r
o
p
o
se
d
fra
m
e
wo
rk
e
ffe
c
ti
v
e
ly
p
re
se
rv
e
s
th
e
c
o
n
ti
n
u
it
y
a
n
d
m
o
r
p
h
o
lo
g
y
o
f
t
h
i
n
a
n
d
th
ick
v
e
ss
e
ls,
e
v
e
n
i
n
re
g
io
n
s
with
lo
w
c
o
n
tras
t
o
r
u
n
e
v
e
n
il
lu
m
i
n
a
ti
o
n
.
O
v
e
ra
ll
,
t
h
e
p
r
o
p
o
se
d
m
e
th
o
d
v
isu
a
l
i
n
sp
e
c
ti
o
n
h
a
s
re
v
e
a
led
th
a
t
t
h
e
su
g
g
e
ste
d
m
e
th
o
d
c
a
n
se
g
m
e
n
t
th
in
a
n
d
t
h
ick
v
e
ss
e
ls
with
g
re
a
ter
a
c
c
u
ra
c
y
th
a
n
p
re
v
io
u
s
m
e
th
o
d
s
.
It
a
lso
d
e
m
o
n
stra
tes
it
s p
o
ten
ti
a
l
f
o
r
re
a
l
-
li
fe
c
li
n
ica
l
a
p
p
li
c
a
ti
o
n
.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
Fu
n
d
u
s
im
ag
es
Har
r
is
Haw
k
o
p
tim
izatio
n
Me
d
ical
im
ag
e
an
aly
s
is
Patch
-
W
is
e
U
-
Net
R
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Sab
a
Sh
eib
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
B
E
ST
I
n
n
o
v
atio
n
Un
iv
er
s
ity
Go
r
an
tla
,
An
d
h
r
a
Pra
d
esh
,
I
n
d
ia
E
m
ail:
s
ab
a.
s
h
eib
a@
m
jco
lleg
e.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
h
u
m
an
r
etin
a
s
er
v
es
as
a
n
o
n
-
in
v
asiv
e
g
atew
ay
to
th
e
c
ir
cu
lato
r
y
an
d
n
er
v
o
u
s
s
y
s
tem
s
,
o
f
f
er
i
n
g
v
ital d
iag
n
o
s
tic
s
ig
n
s
f
o
r
th
e
e
ar
ly
d
etec
tio
n
o
f
s
y
s
tem
ic
an
d
o
p
h
th
alm
ic
d
is
ea
s
es su
ch
as d
i
ab
etic
r
etin
o
p
ath
y
,
g
lau
co
m
a,
h
y
p
er
ten
s
io
n
,
an
d
ar
ter
io
s
cler
o
s
is
[
1
]
.
R
etin
al
f
u
n
d
u
s
im
a
g
in
g
p
r
o
v
id
es
d
etai
led
v
is
u
aliza
tio
n
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
4
3
,
No
.
1
,
Ju
ly
20
2
6
:
157
-
17
8
158
th
e
v
ascu
lar
s
tr
u
ctu
r
e,
wh
er
e
s
u
b
tle
v
a
r
iatio
n
s
in
v
ess
el
m
o
r
p
h
o
lo
g
y
,
ca
lib
e
r
,
a
n
d
b
r
a
n
ch
in
g
p
atter
n
s
ca
n
in
d
icate
d
is
ea
s
e
p
r
o
g
r
ess
io
n
o
r
n
e
u
r
o
lo
g
ical
ab
n
o
r
m
alities
.
Au
to
m
ated
r
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
th
er
ef
o
r
e
p
lay
s
a
f
u
n
d
am
en
tal
r
o
le
in
clin
ical
im
ag
e
an
aly
s
is
,
as
ac
cu
r
ate
ex
tr
ac
tio
n
o
f
b
o
th
th
in
an
d
th
ick
v
ess
els
en
ab
les
r
eliab
le
q
u
an
tific
atio
n
o
f
p
ath
o
lo
g
ical
b
io
m
a
r
k
er
s
.
I
n
r
ec
en
t
y
ea
r
s
[
2
]
,
t
h
e
r
ap
id
ev
o
lu
tio
n
o
f
d
ee
p
lear
n
in
g
h
as
s
ig
n
if
ican
tly
tr
an
s
f
o
r
m
ed
m
e
d
ical
im
ag
e
s
eg
m
en
tatio
n
,
p
r
o
v
id
in
g
d
ata
-
d
r
iv
en
m
o
d
els
th
at
ac
h
iev
e
r
em
a
r
k
ab
le
ac
cu
r
ac
y
in
d
elin
ea
tin
g
a
n
ato
m
ical
s
tr
u
ctu
r
es
[
3
]
,
[
4
]
.
Desp
ite
th
ese
ad
v
an
ce
m
e
n
ts
,
p
r
ec
is
e
s
eg
m
en
tatio
n
o
f
r
etin
a
l
v
ess
els
r
em
ain
s
a
ch
allen
g
in
g
p
r
o
b
lem
d
u
e
to
th
e
c
o
m
p
le
x
ch
ar
ac
ter
is
tics
o
f
f
u
n
d
u
s
im
ag
es
[
5
]
.
T
h
e
v
a
r
ia
b
ilit
y
in
v
ess
el
th
ick
n
ess
,
th
e
lo
w
co
n
tr
ast
b
etwe
en
v
ess
el
s
an
d
b
ac
k
g
r
o
u
n
d
tis
s
u
e,
u
n
ev
en
illu
m
in
atio
n
,
an
d
th
e
p
r
esen
ce
o
f
b
r
ig
h
t
lesi
o
n
s
o
r
o
p
tic
d
is
c
r
eg
io
n
s
f
r
eq
u
en
tly
lead
to
m
is
class
if
icatio
n
an
d
d
is
co
n
ti
n
u
ities
in
v
ess
el
m
ap
s
.
T
r
ad
itio
n
al
ap
p
r
o
ac
h
es
b
ased
o
n
m
atch
e
d
f
ilter
in
g
,
m
o
r
p
h
o
l
o
g
ical
tr
an
s
f
o
r
m
atio
n
s
,
an
d
th
r
esh
o
ld
in
g
h
av
e
s
tr
u
g
g
led
to
g
en
er
alize
ac
r
o
s
s
d
if
f
er
en
t
d
atasets
b
ec
au
s
e
o
f
th
eir
s
en
s
itiv
ity
to
n
o
i
s
e
an
d
illu
m
in
atio
n
v
ar
iatio
n
[
6
]
,
[
7
]
.
E
v
en
with
th
e
in
tr
o
d
u
ctio
n
o
f
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs),
ca
p
tu
r
in
g
m
icr
o
-
v
ess
el
d
etails
an
d
p
r
eser
v
i
n
g
v
ess
el
co
n
tin
u
ity
r
em
ain
d
i
f
f
icu
lt,
as
co
n
v
en
tio
n
al
ar
ch
i
tectu
r
es
o
f
ten
l
o
s
e
s
p
atial
in
f
o
r
m
atio
n
d
u
r
in
g
p
o
o
lin
g
o
p
er
atio
n
s
an
d
p
r
o
d
u
ce
in
co
m
p
lete
s
eg
m
en
tatio
n
in
ar
ea
s
with
wea
k
g
r
ad
ien
ts
o
r
p
ath
o
l
o
g
ical
d
is
tu
r
b
an
ce
s
[
8
]
.
T
o
ad
d
r
ess
th
ese
ch
allen
g
es,
r
esear
ch
e
r
s
h
av
e
p
r
o
p
o
s
ed
a
v
a
r
iety
o
f
d
ee
p
lear
n
in
g
m
o
d
els
aim
ed
at
en
h
an
cin
g
s
tr
u
ctu
r
al
f
ea
tu
r
e
r
eten
tio
n
a
n
d
im
p
r
o
v
in
g
g
e
n
er
aliza
tio
n
.
Am
o
n
g
th
em
,
th
e
U
-
Net
ar
ch
itectu
r
e
h
as
em
er
g
ed
as
th
e
p
r
ed
o
m
in
an
t
f
r
am
ew
o
r
k
f
o
r
m
ed
ical
im
ag
e
s
eg
m
en
tatio
n
b
ec
au
s
e
o
f
its
s
y
m
m
etr
ic
en
co
d
er
–
d
ec
o
d
e
r
s
tr
u
ctu
r
e
a
n
d
s
k
ip
c
o
n
n
ec
tio
n
s
th
at
allo
w
ef
f
ec
tiv
e
f
u
s
io
n
o
f
lo
w
-
a
n
d
h
i
g
h
-
lev
el
f
ea
tu
r
es
[
9
]
.
Nu
m
er
o
u
s
U
-
Net
v
ar
ian
ts
,
in
clu
d
in
g
r
esid
u
al
U
-
Net,
atten
tio
n
U
-
Net,
an
d
d
e
n
s
e
U
-
Net
,
h
av
e
b
ee
n
d
esig
n
ed
to
en
h
an
ce
f
ea
tu
r
e
p
r
o
p
ag
atio
n
an
d
v
ess
el
b
o
u
n
d
ar
y
clar
ity
.
So
m
e
ap
p
r
o
ac
h
es
em
p
lo
y
m
u
lti
-
s
ca
le
co
n
v
o
lu
tio
n
s
o
r
r
ec
u
r
r
e
n
t
la
y
er
s
to
b
etter
ca
p
tu
r
e
s
p
atial
d
ep
en
d
en
cies,
w
h
ile
o
th
er
s
in
teg
r
ate
p
o
s
t
-
p
r
o
ce
s
s
in
g
f
ilter
s
to
r
ef
in
e
v
ess
el
co
n
n
ec
tiv
ity
.
Alth
o
u
g
h
th
ese
m
o
d
if
icatio
n
s
y
ield
p
er
f
o
r
m
an
ce
im
p
r
o
v
em
en
ts
,
m
an
y
U
-
Net
-
b
ased
m
o
d
els
s
till
r
ely
o
n
g
lo
b
ally
p
r
o
ce
s
s
ed
im
ag
es,
wh
ich
ca
n
o
b
s
cu
r
e
f
in
e
-
g
r
ain
e
d
v
ess
el
p
a
tter
n
s
,
p
ar
ticu
lar
ly
in
s
m
all
ca
p
illar
ies
an
d
b
if
u
r
c
atio
n
s
[
1
0
]
.
Ad
d
itio
n
ally
,
th
eir
p
er
f
o
r
m
an
ce
d
ep
en
d
s
h
ea
v
ily
o
n
m
a
n
u
al
h
y
p
e
r
p
ar
am
ete
r
tu
n
in
g
,
s
u
ch
as
lea
r
n
in
g
r
ate,
c
o
n
v
o
lu
ti
o
n
al
k
er
n
el
s
ize,
an
d
n
u
m
b
er
o
f
f
ilter
s
p
er
lay
er
,
wh
ich
n
o
t
o
n
ly
r
e
q
u
ir
es
ex
p
er
t
ex
p
e
r
ien
ce
b
u
t
also
lim
its
s
ca
lab
il
ity
ac
r
o
s
s
d
atasets
.
T
h
e
ab
s
en
ce
o
f
ad
ap
tiv
e
o
p
tim
izatio
n
in
t
h
ese
m
o
d
els
o
f
ten
lea
d
s
to
s
u
b
o
p
t
im
al
co
n
v
er
g
en
ce
,
o
v
er
f
itti
n
g
,
o
r
p
o
o
r
g
e
n
er
aliz
atio
n
wh
en
ap
p
lied
t
o
im
ag
e
s
ca
p
tu
r
ed
u
n
d
er
d
if
f
e
r
en
t
lig
h
tin
g
o
r
r
eso
lu
tio
n
co
n
d
itio
n
s
.
Fig
u
r
e
1
r
ep
r
esen
ts
g
en
e
r
al
p
r
o
ce
s
s
es
in
v
o
lv
ed
in
s
eg
m
e
n
tatio
n
o
f
r
eti
n
al
b
lo
o
d
v
ess
els.
I
n
p
u
r
s
u
it
o
f
h
ig
h
er
p
r
ec
is
io
n
an
d
r
o
b
u
s
tn
ess
,
o
p
tim
izatio
n
-
d
r
iv
en
d
e
ep
lear
n
in
g
h
as
b
ec
o
m
e
a
p
r
o
m
in
e
n
t
r
esear
ch
d
ir
ec
tio
n
.
Me
tah
eu
r
is
tic
alg
o
r
ith
m
s
in
s
p
ir
ed
b
y
n
atu
r
al
a
n
d
b
io
lo
g
ical
p
r
o
ce
s
s
es
s
u
ch
as
p
ar
ticle
s
war
m
o
p
tim
izati
o
n
(
PS
O)
,
g
e
n
etic
alg
o
r
ith
m
s
(
GA)
,
an
d
g
r
ey
wo
lf
o
p
tim
izatio
n
(
GW
O)
h
av
e
d
em
o
n
s
tr
ate
d
ef
f
ec
tiv
en
ess
in
tu
n
in
g
m
o
d
el
p
ar
am
eter
s
f
o
r
co
m
p
lex
,
n
o
n
li
n
ea
r
p
r
o
b
lem
s
[
1
1
]
.
T
h
ese
alg
o
r
ith
m
s
ar
e
ca
p
ab
le
o
f
b
alan
ci
n
g
e
x
p
lo
r
atio
n
an
d
ex
p
lo
itatio
n
with
in
th
e
s
ea
r
ch
s
p
ac
e,
allo
win
g
n
eu
r
al
n
etw
o
r
k
s
to
r
ea
ch
n
ea
r
-
o
p
tim
al
co
n
f
ig
u
r
atio
n
s
with
f
ewe
r
m
an
u
al
in
ter
v
e
n
tio
n
s
.
H
o
wev
er
,
m
an
y
o
f
th
ese
o
p
tim
izatio
n
tech
n
iq
u
es
s
till
f
ac
e
ch
allen
g
es
s
u
ch
as
p
r
em
atu
r
e
c
o
n
v
e
r
g
en
ce
,
lo
ca
l
m
in
im
a
en
tr
a
p
m
en
t,
o
r
h
ig
h
c
o
m
p
u
tatio
n
al
co
s
ts
wh
en
s
ca
led
to
d
e
ep
ar
c
h
itectu
r
es.
T
h
u
s
,
th
e
r
e
is
an
in
s
is
ten
t
n
ee
d
f
o
r
an
in
tellig
en
t
an
d
ef
f
icien
t
o
p
tim
izatio
n
m
ec
h
an
is
m
th
at
ca
n
a
u
to
m
atic
ally
f
in
e
-
tu
n
e
d
ee
p
s
eg
m
e
n
tatio
n
m
o
d
els
to
ad
ap
t
to
th
e
in
tr
in
s
ic
v
ar
iab
ilit
y
o
f
r
etin
al
f
u
n
d
u
s
im
ag
es.
Fig
u
r
e
1
.
Gen
e
r
al
p
r
o
ce
s
s
es in
v
o
lv
ed
i
n
s
eg
m
en
tatio
n
o
f
r
eti
n
al
b
lo
o
d
v
ess
els f
r
o
m
a
f
u
n
d
u
s
im
ag
e
Alth
o
u
g
h
p
r
e
v
io
u
s
s
tu
d
ies
h
a
v
e
ex
p
lo
r
ed
p
atc
h
-
b
ased
U
-
N
et
ar
ch
itectu
r
es
a
n
d
o
p
tim
izati
o
n
-
ass
is
ted
s
eg
m
en
tatio
n
s
ep
ar
ately
,
th
e
p
r
o
p
o
s
ed
Haw
k
Net
f
r
am
ewo
r
k
d
if
f
er
s
b
y
estab
lis
h
in
g
a
s
y
n
er
g
is
tic
in
ter
ac
tio
n
b
etwe
en
lo
ca
lized
p
atch
-
wis
e
lear
n
in
g
an
d
Har
r
is
Ha
wk
o
p
tim
izatio
n
(
HHO)
.
U
n
lik
e
co
n
v
en
tio
n
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Ha
w
kNet:
A
n
in
tellig
en
t b
io
-
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
b
a
s
ed
p
a
tch
w
is
e
a
d
a
p
tive
U
-
N
et
…
(
S
a
b
a
S
h
eib
a
)
159
ap
p
r
o
ac
h
es,
HHO
co
n
tin
u
o
u
s
ly
ad
ap
ts
lea
r
n
in
g
p
a
r
am
eter
s
to
im
p
r
o
v
e
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
c
o
n
v
er
g
en
ce
,
wh
ile
p
atch
-
wis
e
lear
n
in
g
p
r
eser
v
es
th
in
v
ess
el
s
,
ca
p
illar
i
es,
an
d
b
if
u
r
ca
tio
n
s
tr
u
ctu
r
es
.
T
h
is
clo
s
ed
-
lo
o
p
o
p
tim
izatio
n
-
lear
n
i
n
g
m
ec
h
a
n
is
m
en
h
an
ce
s
v
ess
el
co
n
tin
u
ity
,
r
o
b
u
s
tn
ess
to
illu
m
in
atio
n
v
ar
iatio
n
s
,
an
d
g
en
er
aliza
tio
n
ac
r
o
s
s
h
eter
o
g
en
eo
u
s
r
etin
al
d
atasets
,
d
is
tin
g
u
is
h
in
g
Haw
k
Net
f
r
o
m
ex
is
tin
g
o
p
tim
izatio
n
-
en
h
an
ce
d
U
-
Net
f
r
am
ewo
r
k
s
.
Mo
tiv
ated
b
y
th
ese
lim
itatio
n
s
,
th
is
s
tu
d
y
in
tr
o
d
u
ce
s
Haw
k
Net.
T
h
e
co
r
e
id
ea
b
e
h
in
d
Ha
wk
Net
lies
in
co
m
b
in
in
g
a
lo
ca
lized
p
atch
-
wis
e
lear
n
in
g
s
tr
ateg
y
with
a
p
o
wer
f
u
l
b
io
-
in
s
p
ir
e
d
o
p
ti
m
izatio
n
alg
o
r
ith
m
k
n
o
wn
as
HHO
.
I
n
s
tead
o
f
p
r
o
ce
s
s
in
g
th
e
en
tire
f
u
n
d
u
s
im
a
g
e
g
lo
b
ally
,
th
e
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
d
iv
id
es e
ac
h
im
ag
e
in
to
o
v
er
lap
p
i
n
g
p
atch
e
s
,
en
ab
lin
g
th
e
m
o
d
el
to
f
o
cu
s
o
n
lo
ca
l
v
ess
el
f
ea
tu
r
es
s
u
ch
as
th
in
ca
p
illar
ies,
b
if
u
r
ca
tio
n
s
,
an
d
cu
r
v
e
d
s
tr
u
ctu
r
es
th
at
ar
e
o
f
ten
m
is
s
ed
in
g
lo
b
al
tr
ain
in
g
.
T
h
is
p
atch
-
wis
e
U
-
Net
ap
p
r
o
ac
h
n
o
t
o
n
ly
e
n
h
an
ce
s
f
in
e
v
ess
el
d
etec
tio
n
b
u
t
also
en
s
u
r
es
b
etter
h
an
d
lin
g
o
f
lo
ca
l
c
o
n
t
r
ast
v
ar
iatio
n
s
an
d
illu
m
in
atio
n
in
co
n
s
is
ten
cy
.
T
h
e
HHO
alg
o
r
ith
m
au
to
m
atica
lly
ad
ju
s
ts
k
ey
p
atch
wis
e
U
-
Ne
t
h
y
p
er
p
a
r
am
eter
s
in
clu
d
in
g
lear
n
in
g
r
ate,
f
ilter
s
ize,
an
d
weig
h
t
in
itializatio
n
ac
h
iev
i
n
g
f
aster
c
o
n
v
er
g
e
n
ce
an
d
im
p
r
o
v
ed
ac
cu
r
ac
y
with
o
u
t
m
an
u
al
s
u
p
er
v
is
io
n
.
T
h
e
h
y
b
r
id
izatio
n
o
f
U
-
Net’
s
h
ier
ar
ch
ical
f
ea
tu
r
e
ex
tr
ac
tio
n
with
HHO’
s
d
y
n
am
ic
o
p
tim
izatio
n
en
ab
les
th
e
n
etwo
r
k
to
lea
r
n
b
o
th
g
l
o
b
al
c
o
n
tex
t
an
d
lo
ca
lized
m
icr
o
-
v
ess
el
r
ep
r
esen
tatio
n
s
ef
f
icien
tly
,
r
ed
u
cin
g
s
eg
m
en
tatio
n
d
is
co
n
tin
u
ities
an
d
f
alse d
etec
tio
n
s
.
−
T
o
d
ev
el
o
p
an
ad
ap
tiv
e
Haw
k
Net
ar
ch
itectu
r
e
u
s
in
g
p
atch
-
wis
e
tr
ain
in
g
an
d
b
alan
ce
d
e
n
co
d
er
–
d
ec
o
d
er
s
k
ip
co
n
n
ec
tio
n
s
f
o
r
ac
c
u
r
ate
s
eg
m
e
n
tatio
n
o
f
f
in
e
a
n
d
th
i
ck
r
etin
al
v
ess
els
u
n
d
er
c
h
allen
g
in
g
im
a
g
in
g
co
n
d
itio
n
s
.
−
T
o
in
teg
r
ate
HHO
f
o
r
au
to
m
ated
h
y
p
er
p
ar
am
eter
an
d
weig
h
t
tu
n
in
g
,
im
p
r
o
v
in
g
le
ar
n
in
g
s
tab
ilit
y
,
s
eg
m
en
tatio
n
p
r
ec
is
io
n
,
an
d
r
e
d
u
cin
g
r
elian
ce
o
n
lar
g
e
a
n
n
o
t
ated
r
etin
al
d
atasets
.
−
T
o
ev
alu
ate
Haw
k
Net’
s
r
o
b
u
s
tn
ess
an
d
g
en
er
aliza
tio
n
ab
ilit
y
ac
r
o
s
s
DR
I
VE
,
STA
R
E
,
an
d
C
HASE
_
DB
1
d
atasets
,
d
em
o
n
s
tr
atin
g
s
u
p
er
io
r
q
u
an
titativ
e
p
e
r
f
o
r
m
an
ce
an
d
clin
ically
co
n
s
is
ten
t
v
ess
el
s
eg
m
en
tatio
n
co
m
p
ar
ed
to
ex
is
tin
g
m
et
h
o
d
s
.
T
h
e
r
em
ain
d
er
o
f
th
is
p
ap
e
r
i
s
s
tr
u
ctu
r
ed
as
f
o
llo
ws.
Sectio
n
2
r
e
v
iews
ex
is
tin
g
r
esear
ch
o
n
r
etin
al
v
ess
el
s
eg
m
en
tatio
n
.
Sectio
n
3
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
Haw
k
Net
ar
ch
itectu
r
e,
d
etailin
g
its
p
atch
-
wis
e
lear
n
in
g
s
tr
ateg
y
an
d
HHO
m
ec
h
an
is
m
.
Sectio
n
4
d
escr
ib
e
s
th
e
ex
p
er
im
en
tal
s
etu
p
with
q
u
an
tit
ativ
e
an
d
q
u
alitativ
e
r
esu
lts
,
in
clu
d
in
g
co
m
p
ar
ativ
e
an
aly
s
es
with
s
tate
-
of
-
th
e
-
ar
t
tech
n
iq
u
es.
Fin
ally
,
Secti
o
n
5
co
n
clu
d
es
th
e
p
ap
er
o
n
o
p
tim
izatio
n
-
en
h
an
c
ed
r
etin
al
im
ag
e
s
eg
m
e
n
tatio
n
.
Am
o
n
g
th
e
av
ailab
le
b
io
-
in
s
p
ir
e
d
o
p
tim
izatio
n
alg
o
r
ith
m
s
,
HHO
was
s
elec
ted
b
ec
au
s
e
o
f
its
d
y
n
am
ic
b
alan
ce
b
etwe
en
ex
p
lo
r
atio
n
an
d
ex
p
lo
itatio
n
,
w
h
ich
en
a
b
les
ef
f
ec
tiv
e
a
v
o
id
a
n
ce
o
f
lo
ca
l
m
i
n
im
a
an
d
p
r
o
m
o
tes
f
aster
co
n
v
er
g
e
n
ce
.
C
o
m
p
a
r
ed
with
GW
O
an
d
wh
ale
o
p
tim
izatio
n
alg
o
r
ith
m
(
W
OA)
,
HH
O
ex
h
ib
its
s
u
p
er
io
r
ad
ap
tab
ilit
y
an
d
s
ea
r
ch
d
iv
er
s
ity
,
m
ak
i
n
g
it
p
ar
ticu
lar
ly
s
u
itab
le
f
o
r
o
p
tim
izin
g
d
ee
p
lear
n
in
g
h
y
p
er
p
ar
a
m
eter
s
in
m
ed
ical
im
ag
e
s
eg
m
en
tatio
n
task
s
.
T
h
ese
c
h
ar
ac
ter
is
tics
f
ac
ilit
ate
s
tab
le
lear
n
in
g
an
d
im
p
r
o
v
ed
p
r
eser
v
atio
n
o
f
f
in
e
v
asc
u
lar
s
tr
u
ctu
r
es u
n
d
e
r
v
ar
y
in
g
illu
m
in
atio
n
a
n
d
c
o
n
tr
ast co
n
d
itio
n
s
.
Un
lik
e
ex
is
tin
g
o
p
tim
izatio
n
-
ass
is
ted
U
-
Net
m
o
d
els,
th
e
p
r
o
p
o
s
ed
Haw
k
Net
f
r
am
ewo
r
k
co
m
b
in
es
p
atch
-
wis
e
ad
ap
tiv
e
lear
n
in
g
with
HHO
in
a
u
n
if
ied
s
e
g
m
en
tatio
n
f
r
a
m
ewo
r
k
.
T
h
e
p
atch
-
wis
e
s
tr
ateg
y
en
ab
les
ef
f
ec
tiv
e
p
r
eser
v
atio
n
o
f
th
in
v
ess
els,
ca
p
illar
ies,
an
d
b
if
u
r
ca
tio
n
s
tr
u
ctu
r
es
b
y
f
o
c
u
s
in
g
o
n
lo
ca
lized
r
etin
al
r
eg
io
n
s
,
wh
ile
HHO
au
to
m
atica
lly
o
p
tim
izes
k
ey
n
et
wo
r
k
h
y
p
e
r
p
ar
am
eter
s
an
d
lea
r
n
in
g
weig
h
ts
.
T
h
is
in
teg
r
atio
n
estab
lis
h
es
a
s
y
n
e
r
g
is
tic
r
elatio
n
s
h
ip
b
etwe
en
l
o
ca
l
f
ea
tu
r
e
lear
n
in
g
an
d
ad
a
p
tiv
e
o
p
tim
izatio
n
,
r
esu
ltin
g
in
im
p
r
o
v
e
d
v
ess
el
co
n
tin
u
ity
,
r
o
b
u
s
tn
ess
to
illu
m
in
atio
n
v
a
r
iatio
n
s
,
an
d
en
h
an
ce
d
s
eg
m
en
tatio
n
p
er
f
o
r
m
an
ce
ac
r
o
s
s
h
eter
o
g
en
eo
u
s
r
etin
al
d
atasets
.
2.
RE
L
AT
ED
WO
RK
A
g
r
o
win
g
n
u
m
b
er
o
f
ey
e
a
n
d
s
y
s
tem
ic
d
is
o
r
d
er
s
,
in
cl
u
d
in
g
d
ia
b
etic
r
etin
o
p
ath
y
,
g
lau
co
m
a,
a
n
d
h
y
p
er
ten
s
io
n
,
ar
e
b
ein
g
b
etter
d
etec
ted
an
d
tr
ac
k
e
d
th
r
o
u
g
h
th
e
u
s
e
o
f
r
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
,
wh
ich
h
as
b
ec
o
m
e
an
im
p
o
r
t
an
t
f
ield
o
f
s
tu
d
y
i
n
m
ed
ical
im
ag
e
an
aly
s
is
[
1
2
]
.
Fro
m
b
asic
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
s
to
m
o
r
e
co
m
p
lex
m
eth
o
d
s
lik
e
m
atch
ed
f
ilter
in
g
an
d
m
o
r
p
h
o
lo
g
ical
tr
an
s
f
o
r
m
atio
n
s
,
th
er
e
h
av
e
b
ee
n
a
n
ar
r
ay
o
f
a
p
p
r
o
ac
h
e
s
s
u
g
g
ested
f
o
r
im
a
g
e
p
r
o
c
ess
in
g
in
th
e
last
ten
y
ea
r
s
.
W
h
ile
tr
ad
itio
n
al
ap
p
r
o
ac
h
es
o
f
f
e
r
co
m
p
u
tatio
n
al
s
im
p
licity
,
th
eir
p
er
f
o
r
m
a
n
c
e
is
ch
allen
g
in
g
u
n
d
e
r
ce
r
tain
im
ag
in
g
co
n
d
itio
n
s
ch
ar
ac
ter
ized
b
y
u
n
ev
en
illu
m
in
atio
n
,
lo
w
co
n
tr
ast,
an
d
p
ath
o
lo
g
ical
v
ar
iatio
n
s
[
1
3
]
.
T
h
e
task
o
f
a
u
to
m
at
ed
r
etin
al
im
ag
e
s
eg
m
en
tatio
n
ca
n
b
e
d
iv
id
e
d
in
to
two
ca
te
g
o
r
i
es,
n
am
ely
s
u
p
er
v
is
ed
lear
n
in
g
m
eth
o
d
s
an
d
u
n
s
u
p
er
v
is
ed
lear
n
in
g
m
eth
o
d
s
[
1
4
]
.
Un
s
u
p
e
r
v
is
ed
m
eth
o
d
s
[
1
5
]
in
clu
d
e
r
u
le
-
b
ased
ap
p
r
o
ac
h
es
th
at
m
ay
u
tili
ze
a
m
atch
ed
f
ilter
an
d
m
o
r
p
h
o
lo
g
ical
o
p
er
ati
o
n
s
to
d
etec
t
an
d
s
eg
m
en
t
th
e
v
ess
els.
A
s
u
m
m
ar
y
o
f
r
ep
r
esen
tativ
e
r
etin
al
b
l
o
o
d
v
es
s
el
s
eg
m
e
n
tatio
n
m
eth
o
d
s
,
d
atasets
,
p
er
f
o
r
m
a
n
ce
ch
ar
ac
ter
is
tics
,
an
d
k
e
y
lim
i
tatio
n
s
,
is
p
r
esen
ted
i
n
T
ab
l
e
1
.
T
h
e
c
o
m
p
ar
ativ
e
an
aly
s
is
f
o
cu
s
es
o
n
th
e
ch
allen
g
es
th
at
ar
e
alr
ea
d
y
p
r
esen
t
an
d
en
co
u
r
ag
es
th
e
d
ev
elo
p
m
en
t
o
f
th
e
p
r
o
p
o
s
ed
Ha
wk
Net
f
r
am
ewo
r
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
4
3
,
No
.
1
,
Ju
ly
20
2
6
:
157
-
17
8
160
T
ab
le
1
s
u
m
m
ar
izes
th
e
d
atasets
,
m
ajo
r
ad
v
an
tag
es,
a
n
d
in
h
er
en
t
lim
itatio
n
s
o
f
r
e
p
r
esen
tativ
e
r
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
m
eth
o
d
s
.
T
h
e
co
m
p
a
r
is
o
n
s
h
o
ws
th
e
ad
v
an
tag
es
an
d
d
is
ad
v
an
tag
es
o
f
th
e
cu
r
r
en
t
f
r
am
ewo
r
k
s
,
a
n
d
h
ig
h
lig
h
ts
th
e
im
p
o
r
tan
ce
o
f
d
e
v
elo
p
i
n
g
a
m
o
r
e
p
o
wer
f
u
l
o
p
tim
izatio
n
-
b
ased
f
r
am
ewo
r
k
th
at
ca
n
m
ain
tain
th
e
f
id
elity
o
f
f
in
e
v
ess
el
s
tr
u
ctu
r
es
an
d
e
n
h
an
ce
s
eg
m
e
n
tatio
n
p
er
f
o
r
m
an
ce
wh
en
im
ag
i
n
g
u
n
d
er
d
if
f
er
e
n
t c
o
n
d
itio
n
s
.
T
ab
le
1
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
o
f
r
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
m
eth
o
d
s
R
e
f
e
r
e
n
c
e
M
e
t
h
o
d
D
a
t
a
s
e
t
R
e
s
u
l
t
s
Li
mi
t
a
t
i
o
n
s
S
a
u
e
t
a
l
.
[
1
6
]
R
i
d
ge
-
b
a
se
d
seg
m
e
n
t
a
t
i
o
n
D
R
I
V
E
R
e
l
i
a
b
l
e
d
e
t
e
c
t
i
o
n
o
f
ma
j
o
r
b
l
o
o
d
v
e
sse
l
s
P
o
o
r
t
h
i
n
v
e
ss
e
l
d
e
t
e
c
t
i
o
n
Zh
a
n
g
e
t
a
l
.
[
1
7
]
G
a
b
o
r
f
i
l
t
e
r
w
i
t
h
c
l
a
ss
i
f
i
e
r
D
R
I
V
E
I
mp
r
o
v
e
d
v
e
ss
e
l
c
o
n
t
r
a
st
a
n
d
e
d
g
e
c
l
a
r
i
t
y
R
e
q
u
i
r
e
s m
a
n
u
a
l
f
e
a
t
u
r
e
t
u
n
i
n
g
Li
u
e
t
a
l
.
[
1
8
]
Li
n
e
o
p
e
r
a
t
o
r
w
i
t
h
S
V
M
D
R
I
V
E
A
c
c
u
r
a
t
e
v
e
sse
l
b
o
u
n
d
a
r
y
l
o
c
a
l
i
z
a
t
i
o
n
H
i
g
h
c
o
m
p
u
t
a
t
i
o
n
a
l
c
o
m
p
l
e
x
i
t
y
R
a
d
h
a
a
n
d
K
a
r
u
n
a
[
1
9
]
En
se
mb
l
e
-
b
a
se
d
seg
m
e
n
t
a
t
i
o
n
D
R
I
V
E,
S
TA
R
E
R
o
b
u
st
s
e
g
me
n
t
a
t
i
o
n
a
c
r
o
ss
v
a
r
i
e
d
r
e
t
i
n
a
l
i
m
a
g
e
s
Li
mi
t
e
d
c
r
o
ss
-
d
a
t
a
se
t
g
e
n
e
r
a
l
i
z
a
t
i
o
n
W
a
n
g
e
t
a
l
.
[
2
0
]
D
e
e
p
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
D
R
I
V
E,
S
TA
R
E
Ef
f
e
c
t
i
v
e
a
u
t
o
m
a
t
i
c
h
i
e
r
a
r
c
h
i
c
a
l
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
N
e
e
d
s
l
a
r
g
e
l
a
b
e
l
l
e
d
d
a
t
a
s
e
t
s
O
r
t
i
z
-
F
e
r
e
g
r
i
n
o
e
t
a
l
.
[
2
1
]
U
-
N
e
t
a
r
c
h
i
t
e
c
t
u
r
e
D
R
I
V
E
A
c
c
u
r
a
t
e
e
n
d
-
to
-
e
n
d
r
e
t
i
n
a
l
v
e
ss
e
l
s
e
g
me
n
t
a
t
i
o
n
S
t
r
u
g
g
l
e
s w
i
t
h
t
h
i
n
v
e
sse
l
s
D
u
e
t
a
l
.
[
2
2
]
D
e
n
se
U
-
N
e
t
D
R
I
V
E,
C
H
A
S
E_
D
B
1
En
h
a
n
c
e
d
f
e
a
t
u
r
e
p
r
o
p
a
g
a
t
i
o
n
a
n
d
r
e
u
se
I
n
c
r
e
a
se
d
m
o
d
e
l
c
o
m
p
l
e
x
i
t
y
G
a
r
g
a
r
i
e
t
a
l
.
[
2
3
]
A
t
t
e
n
t
i
o
n
U
-
N
e
t
D
R
I
V
E
I
mp
r
o
v
e
d
f
o
c
u
s
o
n
v
e
sse
l
r
e
g
i
o
n
s
H
i
g
h
e
r
m
e
m
o
r
y
a
n
d
c
o
m
p
u
t
a
t
i
o
n
r
e
q
u
i
r
e
m
e
n
t
s
Y
i
e
t
a
l
.
[
2
4
]
M
u
l
t
i
-
sca
l
e
C
N
N
f
u
si
o
n
D
R
I
V
E,
S
TA
R
E
B
e
t
t
e
r
mu
l
t
i
-
sca
l
e
v
e
sse
l
r
e
p
r
e
s
e
n
t
a
t
i
o
n
S
e
n
s
i
t
i
v
e
t
o
i
l
l
u
mi
n
a
t
i
o
n
v
a
r
i
a
t
i
o
n
s
Ad
o
p
ted
m
u
ltis
ca
le
in
p
u
ts
in
U
-
Net
was
p
r
o
p
o
s
ed
b
y
[
2
5
]
with
d
en
s
e
b
lo
ck
s
,
wh
ile
[
2
6
]
p
r
o
p
o
s
ed
r
esid
u
al
co
n
n
ec
tio
n
s
with
s
eq
u
en
ce
s
o
f
3
×
3
co
n
v
o
l
u
tio
n
s
i
n
th
e
s
k
ip
an
d
en
co
d
er
/d
ec
o
d
e
r
p
ath
s
,
d
e
v
elo
p
i
n
g
Mu
ltiR
esU
-
Ne
t.
A
ca
s
ca
d
ed
n
etwo
r
k
o
f
f
o
u
r
u
n
its
u
s
in
g
d
il
ated
co
n
v
o
lu
tio
n
s
at
v
ar
y
i
n
g
r
ates
was
s
u
g
g
ested
by
[
2
7
]
as
R
ef
in
eNe
t
.
T
h
e
f
in
al
s
eg
m
en
tatio
n
m
ap
was
co
m
p
o
s
ed
o
f
th
e
co
m
b
in
ed
o
u
tp
u
ts
o
f
th
ese
u
n
its
.
A
m
u
ltis
ca
le
C
NN
with
b
etter
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
o
b
tain
ed
g
o
o
d
s
en
s
itiv
ity
o
n
v
ascu
lar
s
eg
m
en
t
atio
n
;
[
2
8
]
s
u
g
g
ested
th
is
C
NN;
an
d
[
2
9
]
in
tr
o
d
u
ce
d
h
ie
r
ar
c
h
ical
class
if
icatio
n
to
c
o
m
b
in
e
f
in
d
in
g
s
o
f
two
class
if
ier
s
f
o
r
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
.
B
ec
au
s
e
o
f
t
h
eir
s
u
p
e
r
io
r
p
er
f
o
r
m
an
ce
,
U
-
Nets
an
d
f
u
lly
c
o
n
v
o
lu
tio
n
al
n
etwo
r
k
s
(
FC
Ns)
wer
e
h
ea
v
ily
u
s
ed
f
o
r
d
en
s
e
p
r
e
d
ictio
n
s
[
3
0
]
.
U
-
Net
later
h
ad
en
h
an
ce
m
e
n
t
m
o
d
u
les
in
clu
d
in
g
atten
tio
n
m
ec
h
an
is
m
s
,
d
ilated
co
n
v
o
l
u
tio
n
s
,
an
d
r
esid
u
al
b
lo
ck
s
ad
d
ed
to
it
to
m
ak
e
it
ev
e
n
b
etter
[
3
1
]
.
Mo
r
e
s
p
atial
in
f
o
r
m
atio
n
ca
n
b
e
ca
p
tu
r
ed
,
lo
ca
l
i
n
f
o
r
m
atio
n
lo
s
s
c
an
b
e
r
e
d
u
ce
d
,
an
d
lo
w
-
lev
el
f
ea
tu
r
e
m
ap
s
ca
n
b
e
r
eu
s
ed
f
o
r
ac
cu
r
ate
s
eg
m
e
n
tatio
n
in
a
well
-
d
esig
n
ed
m
o
d
el.
Du
e
to
th
eir
d
ee
p
er
co
n
v
o
lu
tio
n
al
lay
er
s
,
U
-
Net
[
3
2
]
an
d
m
u
lti
-
m
o
d
el
n
etwo
r
k
s
[
3
3
]
o
u
tp
er
f
o
r
m
C
NNs
an
d
FC
Ns
in
s
eg
m
en
tatio
n
o
u
tc
o
m
es.
T
h
is
is
b
ec
au
s
e
th
ese
n
etwo
r
k
s
ar
e
ab
le
to
ex
t
r
ac
t
r
ich
er
in
f
o
r
m
atio
n
.
I
t
is
p
o
s
s
ib
le
to
m
o
d
if
y
th
e
s
eg
m
e
n
tatio
n
r
esu
lts
u
s
in
g
m
atch
in
g
f
ilter
s
o
r
m
o
r
p
h
o
lo
g
ical
tr
an
s
f
o
r
m
s
in
p
o
s
tp
r
o
ce
s
s
in
g
[
3
4
]
.
T
h
is
is
b
ec
au
s
e
t
h
e
r
esu
lts
f
r
eq
u
e
n
tly
in
clu
d
e
n
o
is
e
an
d
i
s
o
lated
tin
y
v
ess
els.
T
h
e
ab
ilit
y
o
f
o
u
r
c
o
m
p
o
s
ite
U
-
Net
I
n
ce
p
tio
n
s
tr
u
ctu
r
e
to
g
ath
er
b
o
th
lo
ca
l a
n
d
g
lo
b
al
tr
aits
allo
wed
it to
o
u
tp
er
f
o
r
m
t
h
e
m
ajo
r
ity
o
f
s
tate
-
of
-
t
h
e
-
ar
t
d
esig
n
s
[
3
5
]
,
[
3
6
]
.
C
o
n
tin
u
o
u
s
ad
v
an
ce
m
en
ts
f
r
o
m
tr
ad
itio
n
al
u
n
s
u
p
e
r
v
is
ed
ap
p
r
o
ac
h
es
to
s
u
p
er
v
is
ed
d
ee
p
lear
n
in
g
m
o
d
els
h
av
e
s
ig
n
if
ican
tly
en
h
an
ce
d
s
eg
m
en
tatio
n
ac
cu
r
ac
y
an
d
ef
f
icien
c
y
.
C
o
n
v
en
tio
n
al
m
eth
o
d
s
s
u
ch
as
ad
ap
tiv
e
th
r
esh
o
l
d
in
g
,
r
eg
io
n
g
r
o
win
g
,
an
d
m
o
r
p
h
o
lo
g
i
ca
l
f
ilt
er
in
g
p
r
o
v
id
ed
ess
en
tial
g
r
o
u
n
d
wo
r
k
b
u
t
s
tr
u
g
g
led
with
n
o
is
e
an
d
f
in
e
v
ess
el
co
n
tin
u
ity
.
Dee
p
lear
n
in
g
m
o
d
els,
p
ar
ticu
lar
l
y
U
-
Net
an
d
its
im
p
r
o
v
ed
v
ar
ian
ts
with
s
k
ip
an
d
r
esid
u
a
l
co
n
n
ec
tio
n
s
,
h
a
v
e
en
ab
le
d
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
b
y
lear
n
in
g
m
u
lti
-
s
ca
le
an
d
co
n
tex
tu
al
f
ea
t
u
r
es.
I
n
te
g
r
ati
n
g
o
p
tim
izatio
n
alg
o
r
ith
m
s
h
as
f
u
r
th
er
r
ef
in
ed
th
ese
m
o
d
els
b
y
ac
ce
ler
atin
g
co
n
v
er
g
en
ce
an
d
im
p
r
o
v
in
g
f
e
atu
r
e
ex
tr
ac
tio
n
.
I
n
o
r
d
er
t
o
b
u
ild
s
u
s
tain
ab
le
an
d
e
f
f
icien
t
elec
tr
ical
en
e
r
g
y
n
etwo
r
k
s
,
in
tellig
en
t
f
o
r
e
ca
s
tin
g
o
f
elec
tr
i
city
d
em
an
d
,
en
er
g
y
m
an
ag
em
en
t,
an
d
in
teg
r
atio
n
o
f
r
en
ewa
b
le
en
er
g
y
r
eso
u
r
ce
s
h
av
e
b
ec
o
m
e
ess
en
tial.
Özü
p
ak
an
d
Ma
n
s
u
r
o
v
[
3
7
]
h
av
e
p
r
esen
ted
a
h
y
b
r
id
d
ee
p
lear
n
in
g
s
y
s
tem
f
o
r
th
e
s
h
o
r
t
-
ter
m
p
r
ed
ic
tio
n
o
f
m
u
lti
-
o
u
tp
u
t
elec
tr
ical
en
er
g
y
d
e
m
an
d
,
h
ig
h
lig
h
tin
g
th
e
a
b
ilit
y
o
f
m
u
ltip
le
n
eu
r
al
a
r
ch
itectu
r
es
to
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
o
f
th
e
elec
tr
icity
d
em
an
d
p
r
ed
ictio
n
an
d
ad
d
r
ess
th
e
co
m
p
lex
tem
p
o
r
al
d
ep
e
n
d
en
cies
am
o
n
g
elec
tr
icity
d
em
a
n
d
p
atter
n
s
.
A
d
if
f
e
r
e
n
t
s
tu
d
y
co
n
d
u
cted
b
y
Özü
p
a
k
[
3
8
]
id
en
tif
ied
th
e
f
ac
t
o
r
s
th
at
af
f
e
ct
th
e
ef
f
ec
tiv
e
n
ess
o
f
th
e
wir
eless
p
o
wer
tr
an
s
m
is
s
io
n
s
y
s
tem
s
f
o
r
n
ex
t
g
en
er
atio
n
elec
tr
ic
v
eh
icles,
in
clu
d
in
g
t
h
e
d
is
tan
ce
b
etwe
en
th
e
co
ils
,
th
e
co
il
d
esig
n
an
d
th
e
o
p
e
r
atin
g
f
r
eq
u
e
n
c
y
o
f
t
h
e
wir
eless
tr
an
s
m
is
s
io
n
s
y
s
tem
.
I
n
v
esti
g
ated
s
m
ar
t
en
er
g
y
m
an
a
g
em
en
t
an
d
o
p
tim
izatio
n
tech
n
iq
u
es
f
o
r
s
u
s
tain
ab
le
p
o
wer
s
y
s
tem
s
,
h
ig
h
lig
h
tin
g
t
h
e
im
p
o
r
tan
ce
o
f
in
tellig
en
t
c
o
n
tr
o
l
alg
o
r
ith
m
s
,
d
em
an
d
r
esp
o
n
s
e
m
ec
h
an
is
m
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Ha
w
kNet:
A
n
in
tellig
en
t b
io
-
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
b
a
s
ed
p
a
tch
w
is
e
a
d
a
p
tive
U
-
N
et
…
(
S
a
b
a
S
h
eib
a
)
161
an
d
r
en
ewa
b
le
en
er
g
y
in
teg
r
at
io
n
in
en
h
an
cin
g
s
y
s
tem
ef
f
ici
en
cy
an
d
r
eliab
ilit
y
.
Ad
v
an
ce
d
co
n
tr
o
l
s
tr
ateg
ies
f
o
r
r
en
ewa
b
le
en
er
g
y
in
te
g
r
at
io
n
wer
e
in
w
h
ich
th
e
y
h
ig
h
li
g
h
ted
t
h
e
u
s
e
o
f
ad
ap
tiv
e
an
d
p
r
ed
ictiv
e
co
n
tr
o
l
m
eth
o
d
s
f
o
r
e
n
s
u
r
in
g
g
r
id
s
ta
b
ilit
y
an
d
p
o
wer
q
u
ality
in
m
o
d
er
n
en
er
g
y
s
y
s
tem
s
.
Fu
r
th
e
r
m
o
r
e,
in
tr
o
d
u
ce
d
a
h
y
b
r
id
d
ee
p
lear
n
in
g
s
y
s
tem
f
o
r
s
o
lar
PV
p
o
wer
f
o
r
ec
asti
n
g
an
d
ac
h
iev
e
d
h
ig
h
er
f
o
r
ec
a
s
tin
g
ac
cu
r
ac
y
an
d
lo
wer
p
r
ed
ictio
n
er
r
o
r
s
,
wh
ich
is
cr
u
cial
f
o
r
th
e
en
er
g
y
s
ch
e
d
u
lin
g
an
d
g
r
id
o
p
e
r
atio
n
.
I
n
ad
d
itio
n
,
th
e
r
ec
en
t
ad
v
an
ce
m
e
n
ts
in
elec
tr
ical
en
er
g
y
s
y
s
tem
s
an
d
s
m
ar
t
g
r
i
d
tech
n
o
l
o
g
ies,
h
i
g
h
lig
h
tin
g
th
e
s
ig
n
if
ican
ce
o
f
d
is
tr
ib
u
ted
g
en
er
atio
n
,
en
er
g
y
s
to
r
ag
e
s
y
s
tem
s
,
in
tel
lig
en
t
m
o
n
ito
r
in
g
,
ar
tific
ial
in
tellig
en
ce
,
an
d
I
n
ter
n
et
o
f
T
h
in
g
s
tech
n
o
l
o
g
ies in
th
e
m
o
d
er
n
izatio
n
o
f
p
o
wer
s
y
s
tem
s
.
Ov
er
all,
th
ese
s
tu
d
ies s
u
g
g
est
th
at
th
e
s
y
n
er
g
y
o
f
de
ep
lear
n
i
n
g
m
et
h
o
d
s
,
o
p
tim
i
za
tio
n
tech
n
iq
u
es,
an
d
s
o
p
h
is
t
icate
d
co
n
tr
o
l
s
tr
ateg
ies
s
ig
n
if
ican
tly
co
n
tr
ib
u
tes
to
im
p
r
o
v
ed
en
e
r
g
y
e
f
f
icien
cy
,
p
r
e
d
ictiv
e
ac
cu
r
ac
y
,
an
d
g
r
id
s
tab
ilit
y
.
T
h
e
a
d
v
en
t
o
f
s
m
ar
t
g
r
id
an
d
in
tellig
en
t
en
er
g
y
m
an
ag
e
m
e
n
t
s
y
s
tem
o
f
f
er
s
g
r
ea
t
p
r
o
s
p
e
cts
to
ac
h
iev
e
s
u
s
tain
ab
le
p
o
wer
g
en
er
atio
n
an
d
d
is
tr
ib
u
tio
n
an
d
s
u
p
p
o
r
t
th
e
g
r
o
win
g
in
teg
r
atio
n
o
f
r
en
ewa
b
l
e
en
er
g
y
s
o
u
r
ce
s
an
d
o
v
er
c
o
m
e
th
e
ch
allen
g
es
o
f
f
u
tu
r
e
e
n
er
g
y
d
em
an
d
.
3.
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
ed
w
o
r
k
co
m
b
i
n
e
s
d
ee
p
co
n
v
o
lu
tio
n
al
p
atch
wis
e
U
-
Net
lear
n
in
g
an
d
m
e
tah
eu
r
is
tic
o
p
tim
izatio
n
to
ac
h
ie
v
e
p
r
ec
i
s
e
r
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
.
T
h
e
m
et
h
o
d
o
lo
g
y
f
o
llo
ws
a
f
o
u
r
-
s
tag
e
p
ip
elin
e:
i)
d
ataset
ac
q
u
is
itio
n
an
d
p
r
e
p
r
o
ce
s
s
in
g
,
ii)
p
atc
h
-
wis
e
ad
ap
tiv
e
U
-
Net
lear
n
i
n
g
,
iii)
HHO
-
b
ased
h
y
p
er
p
ar
am
eter
tu
n
in
g
,
a
n
d
i
v
)
Pro
p
o
s
ed
h
y
b
r
i
d
ap
p
r
o
ac
h
with
co
m
b
in
atio
n
o
f
p
atch
wis
e
U
-
Net
+
HHO
v
ess
el
p
r
o
b
ab
ilit
y
m
ap
g
en
er
a
tio
n
with
p
o
s
t
-
p
r
o
ce
s
s
in
g
r
ef
i
n
em
en
t.
E
ac
h
s
tag
e
e
n
s
u
r
es
th
at
r
etin
al
s
tr
u
ctu
r
es
ar
e
ac
cu
r
ately
r
e
p
r
esen
ted
,
il
lu
m
in
atio
n
in
co
n
s
is
ten
cies
ar
e
m
in
im
ized
,
an
d
f
in
e
ca
p
illar
y
s
tr
u
ctu
r
es
ar
e
p
r
eser
v
ed
th
r
o
u
g
h
o
u
t
th
e
s
e
g
m
en
tatio
n
p
r
o
ce
s
s
.
T
h
e
f
l
o
w
o
f
o
p
er
atio
n
s
is
d
esig
n
ed
to
m
ain
tai
n
s
tr
o
n
g
in
ter
co
n
n
ec
tiv
ity
b
etwe
en
d
at
a
p
r
e
p
ar
atio
n
an
d
lear
n
i
n
g
p
h
ases
,
en
s
u
r
in
g
t
h
at
ea
ch
s
tep
d
ir
ec
tly
en
h
a
n
ce
s
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
s
u
b
s
eq
u
en
t o
n
e.
3
.
1
.
Da
t
a
s
et
des
cr
iptio
n
T
h
e
p
r
o
p
o
s
ed
Haw
k
Net
f
r
am
ewo
r
k
is
e
v
alu
ated
o
n
th
r
ee
p
u
b
licly
av
ailab
le
an
d
clin
icall
y
v
er
i
f
ied
r
etin
al
v
ess
el
s
eg
m
en
tatio
n
d
atasets
DR
I
VE
(
d
ig
ital
r
eti
n
al
im
ag
es
f
o
r
v
ess
el
ex
tr
ac
tio
n
)
,
an
d
STARE
(
s
tr
u
ctu
r
ed
a
n
aly
s
is
o
f
th
e
r
etin
a
)
[
3
9
]
–
[
4
1
]
.
T
ab
le
2
d
e
s
cr
ib
es
th
e
d
ataset
attr
ib
u
tes.
T
h
ese
b
en
c
h
m
ar
k
r
ep
o
s
ito
r
ies
co
llectiv
ely
p
r
o
v
i
d
e
a
co
m
p
r
eh
e
n
s
iv
e
r
ep
r
esen
tatio
n
o
f
r
etin
al
im
ag
es
ca
p
tu
r
ed
u
n
d
er
v
ar
i
o
u
s
clin
ical
co
n
d
itio
n
s
,
im
ag
i
n
g
d
ev
ice
s
,
r
eso
lu
tio
n
s
,
an
d
s
u
b
jec
t p
o
p
u
latio
n
s
.
T
ab
le
2
.
Su
m
m
a
r
y
o
f
d
ataset
a
ttrib
u
tes
A
t
t
r
i
b
u
t
e
n
a
me
D
e
scri
p
t
i
o
n
Ex
a
m
p
l
e
v
a
l
u
e
s
I
mag
e
I
D
U
n
i
q
u
e
i
ma
g
e
i
d
e
n
t
i
f
i
e
r
D
R
I
V
E_
0
1
.
t
i
f
,
S
TA
R
E
_
i
m
0
1
2
.
p
n
g
D
a
t
a
s
e
t
S
o
u
r
c
e
R
e
p
o
s
i
t
o
r
y
o
r
i
g
i
n
D
R
I
V
E,
a
n
d
S
TA
R
E
I
mag
e
R
e
so
l
u
t
i
o
n
D
i
me
n
si
o
n
s
o
f
f
u
n
d
u
s
i
ma
g
e
5
6
5
×
5
8
4
(
D
R
I
V
E)
,
7
0
0
×
6
0
5
(
S
TA
R
E)
,
9
9
9
×
9
6
0
(
C
H
A
S
E
_
D
B
1
)
I
mag
i
n
g
D
e
v
i
c
e
A
c
q
u
i
s
i
t
i
o
n
e
q
u
i
p
m
e
n
t
C
a
n
o
n
C
R
5
3
C
C
D
,
To
p
C
o
n
TR
V
-
50
F
i
e
l
d
o
f
V
i
e
w
C
i
r
c
u
l
a
r
r
e
t
i
n
a
l
c
a
p
t
u
r
e
3
5
°
–
45°
P
a
t
i
e
n
t
C
a
t
e
g
o
r
y
H
e
a
l
t
h
t
y
p
e
N
o
r
mal
,
D
i
s
e
a
s
e
d
A
n
n
o
t
a
t
i
o
n
Ty
p
e
B
i
n
a
r
y
v
e
sse
l
m
a
s
k
V
e
ssel
=
1
,
B
a
c
k
g
r
o
u
n
d
=
0
To
t
a
l
I
mag
e
s
I
mag
e
c
o
u
n
t
p
e
r
d
a
t
a
s
e
t
D
R
I
V
E:
4
0
,
a
n
d
S
TA
R
E
:
2
0
D
a
t
a
s
e
t
S
p
l
i
t
Tr
a
i
n
–
t
e
s
t
p
r
o
p
o
r
t
i
o
n
5
0
%
–
5
0
%
C
o
l
o
u
r
C
h
a
n
n
e
l
s
Ex
t
r
a
c
t
e
d
f
r
o
m
R
G
B
R
e
d
,
G
r
e
e
n
,
B
l
u
e
F
o
r
mat
I
mag
e
st
o
r
a
g
e
f
o
r
ma
t
.
t
i
f
,
.
p
n
g
T
h
eir
in
clu
s
io
n
en
s
u
r
es
th
at
th
e
p
r
o
p
o
s
ed
s
y
s
tem
ac
h
iev
es
h
ig
h
g
en
e
r
aliza
tio
n
ca
p
a
b
ilit
y
an
d
r
o
b
u
s
tn
ess
ac
r
o
s
s
h
eter
o
g
e
n
e
o
u
s
en
v
ir
o
n
m
en
ts
ty
p
ical
in
r
ea
l
-
wo
r
ld
c
o
lo
r
f
u
n
d
u
s
im
ag
es
an
d
ex
p
er
t
-
an
n
o
tated
b
in
a
r
y
v
ess
el
m
ask
s
th
at
s
er
v
e
as
g
r
o
u
n
d
tr
u
t
h
f
o
r
o
b
je
ctiv
e
ev
al
u
atio
n
.
T
h
e
DR
I
VE
d
atase
t
co
n
tain
s
4
0
co
lo
r
f
u
n
d
u
s
im
a
g
es
(
2
0
f
o
r
tr
ain
in
g
an
d
2
0
f
o
r
test
in
g
)
o
b
tain
ed
at
a
4
5
°
f
ield
o
f
v
iew.
E
ac
h
im
ag
e
h
as
a
r
eso
lu
ti
o
n
o
f
5
6
5
×5
8
4
p
ix
els
an
d
is
a
cc
o
m
p
a
n
i
ed
b
y
p
r
ec
is
ely
an
n
o
tated
v
es
s
el
m
ask
s
p
r
o
d
u
ce
d
b
y
two
i
n
d
ep
en
d
en
t
o
p
h
th
al
m
o
lo
g
is
ts
,
with
o
n
e
u
s
ed
a
s
th
e
r
ef
er
en
ce
s
tan
d
ar
d
.
T
h
e
STARE
d
ataset
co
m
p
r
is
es
2
0
r
etin
al
im
ag
es
(
1
0
h
ea
lth
y
a
n
d
1
0
p
ath
o
lo
g
ical)
u
n
d
er
a
3
5
°
f
ield
o
f
v
iew
an
d
7
0
0
×6
0
5
r
eso
lu
tio
n
.
T
h
ese
im
ag
es
ex
h
i
b
it
a
wid
e
s
p
ec
tr
u
m
o
f
r
etin
a
l
ab
n
o
r
m
ali
ties
,
in
clu
d
in
g
lesi
o
n
s
,
h
em
o
r
r
h
ag
es,
an
d
v
ar
ia
b
le
illu
m
in
atio
n
z
o
n
e
s
,
p
r
o
v
id
in
g
ch
allen
g
in
g
ca
s
es
f
o
r
ev
alu
atin
g
s
eg
m
e
n
tatio
n
c
o
n
s
is
ten
cy
.
3
.
2
.
P
re
-
pro
ce
s
s
in
g
E
ac
h
h
awk
in
t
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
c
o
r
r
esp
o
n
d
s
to
a
ca
n
d
id
ate
h
y
p
er
p
ar
am
eter
v
ec
to
r
m
ad
e
u
p
o
f
th
e
lear
n
in
g
r
ate
,
th
e
s
ize
o
f
th
e
f
ilter
,
th
e
d
r
o
p
o
u
t p
r
o
b
ab
i
lity
an
d
th
e
lo
s
s
-
weig
h
t c
o
ef
f
i
cien
ts
.
T
h
e
n
u
m
b
er
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
4
3
,
No
.
1
,
Ju
ly
20
2
6
:
157
-
17
8
162
o
f
HHO
was
to
b
e
2
0
a
n
d
th
e
n
u
m
b
e
r
o
f
o
p
tim
izati
o
n
iter
at
io
n
s
5
0
.
Fo
r
ea
ch
iter
atio
n
,
ca
n
d
id
ate
p
a
r
am
eter
s
ets
wer
e
test
ed
with
a
s
m
al
l
n
u
m
b
er
o
f
e
p
o
ch
s
o
f
th
e
tr
ain
in
g
o
f
th
e
U
-
Net,
a
n
d
th
e
b
est
s
o
lu
tio
n
was
s
elec
ted
to
im
p
r
o
v
e
th
e
g
lo
b
a
l
o
p
tim
u
m
.
T
h
e
o
p
tim
ized
p
a
r
am
eter
s
wer
e
u
s
ed
t
o
r
etr
ain
th
e
last
p
atch
-
wis
e
U
-
Net
to
cr
ea
te
t
h
e
s
eg
m
e
n
tatio
n
m
o
d
el
af
te
r
co
n
v
er
g
en
ce
.
T
h
e
o
p
tim
izatio
n
-
lear
n
i
n
g
i
n
ter
ac
tio
n
allo
ws
f
o
r
s
tab
le
co
n
v
er
g
e
n
ce
an
d
m
o
r
e
ac
cu
r
ate
s
eg
m
en
tatio
n
.
I
m
ag
e
n
o
r
m
aliza
tio
n
an
d
r
esiz
in
g
:
Sin
ce
th
e
d
atasets
v
ar
y
in
r
eso
lu
tio
n
,
all
im
ag
es
(
,
)
ar
e
r
esized
to
a
u
n
if
o
r
m
s
ca
le
o
f
5
1
2
×
5
1
2
p
ix
els
u
s
in
g
b
icu
b
ic
in
ter
p
o
latio
n
to
m
ain
tain
g
eo
m
etr
ic
co
n
s
is
ten
cy
ac
r
o
s
s
d
atasets
:
(
′
,
′
)
=
∑
∑
(
,
)
⋅
ℎ
(
′
−
)
⋅
ℎ
(
′
−
)
)
)
−
1
=
0
−
1
=
0
(
1
)
wh
er
e
ℎ
is
th
e
b
icu
b
ic
in
ter
p
o
latio
n
k
er
n
el.
T
h
is
s
tep
en
s
u
r
es
s
p
atial
u
n
if
o
r
m
ity
an
d
p
r
e
p
ar
e
s
th
e
d
ata
f
o
r
th
e
n
ex
t
s
tag
e
co
lo
u
r
ch
a
n
n
el
ex
tr
ac
tio
n
,
wh
er
e
m
ea
n
in
g
f
u
l
v
ess
el
co
n
tr
ast
i
s
i
s
o
lated
f
o
r
s
eg
m
en
tatio
n
en
h
an
ce
m
e
n
t.
Gr
ee
n
ch
a
n
n
el
ex
tr
ac
tio
n
an
d
in
ten
s
ity
n
o
r
m
aliza
tio
n:
R
etin
al
v
ess
el
s
e
x
h
ib
i
t
th
e
h
ig
h
est
co
n
tr
ast
in
th
e
g
r
ee
n
ch
a
n
n
el
d
u
e
to
h
ae
m
o
g
lo
b
i
n
ab
s
o
r
p
tio
n
p
r
o
p
er
ties
.
Hen
ce
,
th
e
R
GB
im
ag
e
is
d
ec
o
m
p
o
s
ed
,
an
d
th
e
g
r
ee
n
c
o
m
p
o
n
e
n
t
(
,
)
is
ex
tr
ac
ted
.
T
h
e
p
ix
el
in
ten
s
ities
ar
e
n
o
r
m
alize
d
in
to
th
e
[
0
,
1
]
r
an
g
e
v
ia
m
in
–
m
a
x
n
o
r
m
aliza
tio
n
:
(
,
)
=
(
,
)
−
−
(
2
)
T
h
is
n
o
r
m
aliza
tio
n
r
em
o
v
es
in
ter
-
im
ag
e
b
r
ig
h
t
n
ess
d
if
f
er
e
n
ce
s
,
s
tab
ilizin
g
tr
ain
in
g
a
n
d
im
p
r
o
v
in
g
th
e
n
ex
t
s
tag
e
c
o
n
tr
ast
en
h
an
c
em
en
t,
wh
ich
r
e
f
in
es
v
ess
el
v
i
s
ib
ilit
y
b
y
em
p
h
asizin
g
s
u
b
tle
g
r
ad
ien
ts
.
C
o
n
tr
ast
en
h
an
ce
m
e
n
t
u
s
in
g
c
o
n
tr
ast
-
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
e
q
u
aliza
tio
n
(
C
L
AHE
)
:
T
o
co
u
n
t
er
lo
w
co
n
t
r
ast
an
d
illu
m
in
atio
n
im
b
alan
ce
,
C
L
AHE
is
ap
p
lied
lo
ca
lly
:
ℎ
(
,
)
=
(
(
,
)
,
)
(
3
)
wh
er
e
r
ep
r
esen
ts
th
e
clip
p
i
n
g
lim
it
th
at
co
n
tr
o
ls
o
v
e
r
-
am
p
l
if
icatio
n
o
f
co
n
t
r
ast.
T
h
e
o
u
tc
o
m
e
en
h
an
ce
s
m
icr
o
-
v
ess
els
in
d
ar
k
r
eg
io
n
s
with
o
u
t
am
p
lify
in
g
n
o
is
e,
cr
ea
tin
g
o
p
tim
al
in
p
u
t
f
o
r
th
e
f
o
llo
win
g
n
o
is
e
r
ed
u
ctio
n
s
tep
th
at
en
s
u
r
es
s
tr
u
ctu
r
al
clar
ity
i
n
f
in
e
v
ess
els.
No
is
e
Su
p
p
r
ess
io
n
v
ia
Gau
s
s
ian
Fil
ter
in
g
:
No
is
e
f
r
o
m
im
a
g
in
g
h
ar
d
war
e
is
s
u
p
p
r
ess
ed
u
s
in
g
a
Gau
s
s
ian
s
m
o
o
th
in
g
f
ilter
:
(
,
)
=
1
2
2
−
2
+
2
2
2
(
4
)
an
d
th
e
s
m
o
o
th
ed
im
a
g
e
is
o
b
t
ain
ed
b
y
:
ℎ
(
,
)
=
ℎ
(
,
)
∗
(
,
)
(
5
)
wh
er
e
co
n
tr
o
ls
th
e
f
ilter
’
s
s
p
r
ea
d
.
T
h
is
s
tep
r
etain
s
es
s
en
ti
al
v
ess
el
ed
g
es
wh
ile
r
ed
u
cin
g
h
ig
h
-
f
r
eq
u
e
n
c
y
d
is
tu
r
b
an
ce
s
,
f
o
r
m
in
g
t
h
e
f
o
u
n
d
atio
n
f
o
r
t
h
e
p
atch
e
x
tr
a
ctio
n
s
tag
e
th
at
tar
g
ets
lo
ca
l
s
tr
u
ctu
r
es.
Patch
ex
tr
ac
tio
n
f
o
r
lo
ca
lized
lear
n
i
n
g
:
Af
ter
n
o
is
e
s
u
p
p
r
ess
io
n
,
e
ac
h
en
h
an
ce
d
im
ag
e
is
d
iv
id
e
d
in
to
o
v
e
r
lap
p
in
g
p
atch
es o
f
64
×
64
p
ix
els to
f
o
cu
s
o
n
l
o
ca
lized
v
ess
el
r
eg
io
n
s
:
=
ℎ
(
:
+
63
,
:
+
63
)
(
6
)
T
h
is
p
r
o
ce
s
s
en
h
an
ce
s
th
e
d
et
ec
tio
n
o
f
t
h
in
ca
p
illar
ie
s
an
d
b
if
u
r
ca
tio
n
s
o
f
ten
lo
s
t
in
g
lo
b
al
tr
ain
in
g
.
T
h
ese
p
atch
es
s
er
v
e
as
tr
ai
n
in
g
u
n
its
f
o
r
th
e
ad
ap
tiv
e
U
-
Ne
t
m
o
d
el.
T
o
f
u
r
th
er
s
tr
en
g
th
e
n
lear
n
i
n
g
,
th
e
n
ex
t
s
tep
ap
p
lies
d
ata
au
g
m
en
tatio
n
to
d
i
v
er
s
if
y
th
e
p
atch
d
ataset
an
d
en
s
u
r
e
r
o
tatio
n
an
d
s
ca
le
in
v
ar
ian
ce
.
Data
au
g
m
en
tatio
n
:
T
o
im
p
r
o
v
e
m
o
d
el
r
o
b
u
s
tn
ess
an
d
p
r
e
v
en
t
o
v
er
f
itti
n
g
,
tr
an
s
f
o
r
m
atio
n
s
s
u
ch
as
r
o
tatio
n
(
)
,
s
ca
lin
g
(
)
,
an
d
h
o
r
izo
n
tal
f
lip
p
in
g
(
ℎ
)
ar
e
ap
p
lied
:
=
ℎ
(
(
(
)
)
)
(
7
)
T
h
ese
au
g
m
en
tatio
n
s
ex
p
an
d
t
h
e
tr
ain
in
g
d
ata,
e
n
s
u
r
in
g
th
at
th
e
m
o
d
el
lear
n
s
v
ess
el
co
n
tin
u
ity
ir
r
esp
ec
tiv
e
o
f
o
r
ien
tatio
n
o
r
p
atien
t
-
s
p
ec
if
ic
v
ar
iatio
n
.
T
h
e
au
g
m
en
ted
p
a
tch
es
ar
e
th
en
s
tan
d
ar
d
ized
t
o
en
s
u
r
e
co
n
s
is
ten
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Ha
w
kNet:
A
n
in
tellig
en
t b
io
-
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
b
a
s
ed
p
a
tch
w
is
e
a
d
a
p
tive
U
-
N
et
…
(
S
a
b
a
S
h
eib
a
)
163
in
ten
s
ity
d
is
tr
ib
u
tio
n
f
o
r
o
p
tim
izati
on.
Data
s
et
Stan
d
a
r
d
izatio
n
:
T
h
e
au
g
m
en
ted
i
m
ag
e
p
atc
h
es
ar
e
s
tan
d
ar
d
ized
u
s
in
g
ze
r
o
-
m
ea
n
an
d
u
n
it
-
v
ar
ian
ce
n
o
r
m
aliza
tio
n
:
(
,
)
=
(
,
)
−
(
8
)
wh
er
e
an
d
r
ep
r
esen
t
th
e
m
ea
n
an
d
s
tan
d
a
r
d
d
e
v
iatio
n
o
f
t
h
e
d
ataset.
T
h
is
tr
an
s
f
o
r
m
atio
n
en
s
u
r
es
s
tab
le
co
n
v
er
g
en
ce
d
u
r
in
g
n
etwo
r
k
tr
ain
in
g
a
n
d
p
r
ep
a
r
es
th
e
d
ata
f
o
r
f
ea
tu
r
e
e
x
tr
ac
tio
n
b
y
th
e
p
atch
-
wis
e
ad
ap
tiv
e
U
-
Net
in
th
e
n
ex
t
p
h
ase.
3
.
3
.
P
r
o
po
s
ed
m
o
del
T
h
e
p
r
o
p
o
s
ed
Haw
k
Net
f
r
a
m
ewo
r
k
in
teg
r
ates
s
y
n
e
r
g
y
b
et
wee
n
U
-
Net
an
d
HHO
f
o
r
m
s
a
h
y
b
r
id
o
p
tim
izatio
n
-
d
r
iv
en
lea
r
n
in
g
p
ar
ad
ig
m
,
d
esig
n
e
d
to
ac
h
iev
e
h
ig
h
p
r
ec
is
io
n
,
r
o
b
u
s
tn
ess
,
an
d
g
e
n
er
aliza
tio
n
ac
r
o
s
s
d
iv
er
s
e
r
etin
al
im
ag
e
d
atasets
wh
ile
m
ain
tain
in
g
co
m
p
u
tatio
n
al
ef
f
icien
c
y
an
d
s
t
ab
ilit
y
d
u
r
in
g
m
o
d
el
co
n
v
er
g
en
ce
.
3
.
3
.
1
.
P
a
t
ch
-
wis
e
U
-
Net
a
rc
h
it
ec
t
ure
T
h
e
p
atch
wis
e
U
-
Net
ar
ch
it
ec
tu
r
e
(
s
ee
Fig
u
r
e
2
)
,
o
r
ig
i
n
ally
d
esig
n
ed
f
o
r
th
e
s
eg
m
e
n
tatio
n
o
f
n
eu
r
o
n
al
s
tr
u
ctu
r
es
in
elec
tr
o
n
m
icr
o
s
co
p
ic
s
tack
s
,
h
as
b
ec
o
m
e
o
n
e
o
f
th
e
m
o
s
t
wid
ely
u
s
ed
co
n
v
o
lu
tio
n
al
ar
ch
itectu
r
es
in
b
io
m
ed
ical
im
ag
e
an
aly
s
is
.
I
t
e
x
ten
d
s
th
e
co
n
v
en
tio
n
al
f
u
lly
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
FC
N)
b
y
in
tr
o
d
u
cin
g
a
s
y
m
m
etr
ic
en
co
d
er
-
d
ec
o
d
er
s
tr
u
ctu
r
e
th
at
ca
p
t
u
r
es
b
o
th
co
n
tex
t
u
al
an
d
s
p
atial
in
f
o
r
m
atio
n
(
s
ee
Alg
o
r
ith
m
1
)
.
T
h
e
f
u
n
d
am
en
ta
l c
o
n
v
o
l
u
tio
n
o
p
er
atio
n
i
n
U
-
Net
is
m
ath
em
atica
lly
r
ep
r
esen
ted
as:
F
(
x
,
y
)
=
(
I
*
K
)(
x
,
y
)
=
∑
m
∑
n
I
(
x
-
m
,
y
-
n
)
K
(
m
,
n
)
(
9
)
In
(
9
)
g
iv
en
as
th
e
s
u
m
m
atio
n
o
f
elem
en
t
-
wis
e
m
u
ltip
licatio
n
s
b
etwe
en
th
e
in
p
u
t
im
ag
e
(
,
)
an
d
th
e
co
n
v
o
l
u
tio
n
al
k
er
n
el
(
,
)
,
wh
er
e
,
r
ep
r
esen
t
k
er
n
el
in
d
ice
s
an
d
(
,
)
d
en
o
tes
th
e
r
esu
ltin
g
f
ea
tu
r
e
m
a
p
o
b
tain
ed
b
y
s
lid
i
n
g
th
e
k
e
r
n
el
ac
r
o
s
s
th
e
im
a
g
e
to
ex
tr
ac
t
l
o
ca
l
s
p
atial
f
ea
tu
r
es.
T
o
in
tr
o
d
u
ce
non
-
lin
ea
r
ity
an
d
en
a
b
le
th
e
n
etwo
r
k
to
lear
n
co
m
p
lex
h
ie
r
ar
ch
ical
r
ep
r
esen
tatio
n
s
,
th
e
ac
tiv
atio
n
f
u
n
ctio
n
ap
p
lied
is
th
e
R
e
ctif
ied
L
in
ea
r
Un
it (
R
eL
U)
,
ex
p
r
ess
ed
as:
f
(
x
)
=
m
ax
(
0
,
x
),
(
1
0
)
In
(
1
0
)
g
iv
e
n
as
a
r
ec
tifie
r
f
u
n
ctio
n
wh
er
e
x
r
ep
r
esen
ts
th
e
weig
h
ted
s
u
m
o
f
i
n
p
u
ts
to
a
n
eu
r
o
n
,
an
d
th
e
f
u
n
ctio
n
o
u
tp
u
ts
x
wh
en
p
o
s
itiv
e
an
d
ze
r
o
o
th
er
wis
e,
th
u
s
en
s
u
r
in
g
n
o
n
-
s
atu
r
atin
g
g
r
a
d
ien
t
f
lo
w
d
u
r
in
g
b
ac
k
p
r
o
p
ag
atio
n
.
I
n
th
e
e
n
co
d
er
p
ath
,
th
e
n
etwo
r
k
p
r
o
g
r
ess
iv
ely
r
e
d
u
ce
s
th
e
s
p
atial
d
im
e
n
s
io
n
s
th
r
o
u
g
h
m
ax
-
p
o
o
lin
g
,
d
ef
in
e
d
m
ath
em
atica
l
ly
as:
(
,
)
=
m
ax
(
m
,
n
)
∈
R
(
i
,
j
)
F
(
m
,
n
)
(
1
1
)
In
(
1
1
)
g
iv
en
as
th
e
s
elec
tio
n
o
f
th
e
m
ax
im
u
m
ac
tiv
atio
n
v
alu
e
with
in
a
r
ec
ep
tiv
e
f
ield
(
,
)
ce
n
ter
ed
ar
o
u
n
d
p
o
s
itio
n
(
,
)
,
wh
er
e
(
,
)
ar
e
p
ix
el
in
ten
s
ities
o
r
f
e
atu
r
e
ac
tiv
atio
n
s
,
th
er
eb
y
en
h
an
cin
g
tr
an
s
latio
n
al
in
v
ar
ian
ce
an
d
r
ed
u
cin
g
co
m
p
u
tat
io
n
al
c
o
m
p
l
ex
ity
wh
ile
r
etain
in
g
k
e
y
c
o
n
t
ex
tu
al
in
f
o
r
m
atio
n
.
T
h
e
d
ec
o
d
e
r
p
ath
r
ec
o
n
s
tr
u
c
ts
th
e
o
r
ig
in
al
s
p
atial
r
eso
lu
tio
n
u
s
in
g
u
p
-
s
am
p
lin
g
o
p
er
atio
n
s
,
wh
ich
ar
e
r
ep
r
esen
ted
as:
(
,
)
=
F
(
x
s
,
y
s
)
(
1
2
)
In
(
1
2
)
g
iv
e
n
as
th
e
en
lar
g
em
e
n
t
o
f
f
ea
tu
r
e
m
ap
s
b
y
a
s
ca
le
f
ac
to
r
s
,
wh
er
e
(
,
)
d
en
o
tes
th
e
in
p
u
t
f
ea
tu
r
e
m
ap
an
d
(
,
)
r
ep
r
esen
ts
t
h
e
u
p
-
s
am
p
led
o
u
t
p
u
t,
r
esto
r
i
n
g
f
in
e
d
etails
lo
s
t
d
u
r
in
g
d
o
wn
-
s
am
p
lin
g
.
T
o
m
ain
tain
f
ea
tu
r
e
co
n
ti
n
u
it
y
an
d
p
r
ev
e
n
t
lo
s
s
o
f
lo
ca
lizatio
n
in
f
o
r
m
atio
n
,
s
k
ip
co
n
n
ec
t
io
n
s
ar
e
in
tr
o
d
u
ce
d
b
etwe
en
co
r
r
esp
o
n
d
in
g
lay
er
s
o
f
th
e
en
c
o
d
er
an
d
d
ec
o
d
e
r
.
T
h
is
m
ec
h
an
is
m
is
f
o
r
m
u
lated
a
s
:
(
,
)
=
(
,
)
⊕
(
,
)
(
1
3
)
In
(
1
3
)
g
iv
en
as
an
elem
en
t
-
wis
e
co
n
ca
ten
atio
n
(
⊕
)
o
f
th
e
d
ec
o
d
er
'
s
u
p
-
s
am
p
led
f
ea
tu
r
e
m
ap
(
,
)
an
d
th
e
en
co
d
e
r
'
s
co
r
r
esp
o
n
d
in
g
f
ea
tu
r
e
m
a
p
(
,
)
,
allo
win
g
th
e
m
o
d
el
to
co
m
b
i
n
e
lo
w
-
lev
el
s
p
atial
d
etails
with
h
ig
h
-
le
v
el
s
em
an
tic
co
n
te
x
t,
th
e
r
eb
y
im
p
r
o
v
in
g
s
eg
m
e
n
tatio
n
ac
cu
r
ac
y
at
o
b
ject
b
o
u
n
d
ar
ies
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
4
3
,
No
.
1
,
Ju
ly
20
2
6
:
157
-
17
8
164
Fig
u
r
e
2
.
Pro
p
o
s
ed
p
atch
wis
e
U
-
n
et
ar
ch
itectu
r
e
f
o
r
r
etin
al
b
lo
o
d
v
ess
el
s
eg
m
en
tatio
n
Alg
o
r
ith
m
1
.
Patch
-
wis
e
U
-
Net
In
pu
t:
Re
ti
na
l
im
ag
e
da
ta
se
t
=
{
1
,
2
,
…
,
}
,
pa
tc
h
si
ze
×
,
le
ar
ni
ng
ra
te
,
nu
mb
er
of
epochs
Output: Trained U
-
Net model with optimized weights
Begin
Step 1: Preprocessing Phase
for each image
in dataset
do
Normalize pixel intensity values between [0, 1]
Apply contrast enhancement and noise reduction
Divide
into non
-
overlapping patche
s of size
×
Store all patches as
=
{
1
,
2
,
…
,
}
end for
Step 2: Network Initialization
Initialize U
-
Net encoder
–
decoder architecture with weights
Define loss function
=
1
×
CrossEntropyLoss
+
2
×
DiceLoss
Initialize optimizer (e.g., Adam) with learning rate
Step 3: Training Phase
for epoch = 1 to
do
for each batch of patches
in
do
Perform
forward
propagation
through
encoder:
Appl
y
convolution,
ReLU
acti
vation,
and
max
-
pooling
Perform decoding: Apply up
-
sampling, concatenation via skip connections, and convolution
Compute prediction
̂
for
Calculate total loss
using ground tr
uth labels
Perform backpropagation and update weights
−
∂
∂
end for
end for
Step 4: Reconstruction Phase
for each image
in dataset
do
Predict segmentation for all patches using trained U
-
Net
Reconstruct full segmented image by merging p
atch outputs
end for
Step 5: Output final segmented images and trained model parameters
end
Du
r
in
g
m
o
d
el
t
r
ain
in
g
,
th
e
p
ix
el
-
wis
e
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
is
em
p
l
o
y
ed
to
m
ea
s
u
r
e
th
e
d
is
cr
ep
an
cy
b
etwe
en
p
r
e
d
icted
an
d
g
r
o
u
n
d
tr
u
th
lab
els,
d
e
f
i
n
ed
as:
L
CE
=
-
∑
N
i
=1
∑
C
c
=1
,
lo
g
(
,
)
(
1
4
)
In
(
1
4
)
g
iv
en
as
th
e
n
eg
ativ
e
lo
g
ar
ith
m
ic
lo
s
s
co
m
p
u
ted
o
v
er
N
p
ix
els
an
d
C
class
es,
wh
er
e
,
i
s
th
e
g
r
o
u
n
d
tr
u
th
lab
el
(
1
if
p
i
x
el
i
b
elo
n
g
s
to
class
c
,
o
th
er
wis
e
0
)
a
n
d
,
is
th
e
p
r
e
d
icted
p
r
o
b
a
b
ilit
y
o
f
p
ix
el
i
b
elo
n
g
in
g
to
class
c
;
th
is
p
en
alize
s
in
co
r
r
ec
t
class
if
icatio
n
s
an
d
d
r
i
v
es
th
e
m
o
d
el
to
war
d
ac
cu
r
ate
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Ha
w
kNet:
A
n
in
tellig
en
t b
io
-
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
b
a
s
ed
p
a
tch
w
is
e
a
d
a
p
tive
U
-
N
et
…
(
S
a
b
a
S
h
eib
a
)
165
p
r
ed
ictio
n
s
.
T
o
h
a
n
d
le
class
im
b
alan
ce
an
d
im
p
r
o
v
e
r
eg
io
n
o
v
er
lap
d
u
r
in
g
s
eg
m
en
tatio
n
,
th
e
Dice
lo
s
s
i
s
u
tili
ze
d
,
r
ep
r
esen
ted
as:
L
Dic
e
= 1
-
2
∑
i
p
i
y
i
+
ϵ
∑
i
p
i
2
+
∑
i
y
i
2
+
ϵ
(
1
5
)
T
h
e
g
iv
en
in
(
1
5
)
is
a
d
if
f
er
e
n
tiab
le
lo
s
s
m
etr
ic
in
wh
ich
t
h
e
p
r
ed
icted
an
d
tr
u
e
b
in
a
r
y
m
ask
v
alu
es
ar
e
d
en
o
ted
b
y
"p
"
_
"i"
a
n
d
"y
"
_
"i"
r
esp
ec
tiv
ely
,
an
d
a
s
m
all
s
m
o
o
th
in
g
co
n
s
tan
t
"
ϵ
"
is
u
s
ed
t
o
a
v
o
id
d
iv
is
io
n
b
y
ze
r
o
.
T
h
is
f
u
n
ctio
n
m
ak
es
th
e
m
o
s
t
o
f
th
e
s
p
atial
o
v
er
lap
b
etwe
en
th
e
p
r
ed
ict
ed
an
d
g
r
o
u
n
d
tr
u
th
r
eg
io
n
s
,
w
h
ich
is
esp
ec
ially
u
s
ef
u
l
f
o
r
m
ed
ical
d
atasets
th
at
ar
e
im
b
alan
ce
d
.
I
n
o
r
d
er
t
o
ac
h
ie
v
e
a
co
m
p
r
o
m
is
e
b
etwe
en
class
if
ic
atio
n
ac
cu
r
ac
y
a
n
d
r
e
g
io
n
co
n
s
is
ten
cy
,
th
e
to
tal
lo
s
s
f
u
n
ctio
n
in
co
r
p
o
r
ates
b
o
th
cr
o
s
s
-
en
tr
o
p
y
an
d
d
ice
co
m
p
o
n
en
ts
.
I
t is r
ep
r
esen
ted
as
:
L
tot
a
l
=
α
L
CE
+
β
L
Dic
e
(
16
)
In
(
1
6
)
g
i
v
en
as
a
weig
h
ted
s
u
m
m
atio
n
wh
er
e
α
an
d
β
ar
e
s
ca
lar
co
e
f
f
icien
ts
th
at
co
n
tr
o
l
th
e
co
n
tr
ib
u
tio
n
o
f
ea
c
h
co
m
p
o
n
en
t,
e
n
ab
lin
g
f
lex
ib
le
o
p
ti
m
izatio
n
b
ased
o
n
d
atas
et
ch
ar
ac
ter
is
tics
an
d
s
eg
m
en
tatio
n
r
eq
u
ir
e
m
en
ts
.
I
n
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
th
e
in
p
u
t
im
ag
e
is
d
iv
id
ed
in
to
n
o
n
-
o
v
er
lap
p
in
g
p
atch
es
b
ef
o
r
e
b
ein
g
p
r
o
ce
s
s
ed
b
y
th
e
U
-
Net.
T
h
is
p
atc
h
-
wis
e
ap
p
r
o
ac
h
en
h
a
n
ce
s
lo
ca
l
f
ea
t
u
r
e
r
eten
tio
n
,
r
ed
u
ce
s
GPU
m
em
o
r
y
c
o
n
s
u
m
p
tio
n
,
an
d
allo
ws
th
e
m
o
d
el
to
f
o
cu
s
o
n
f
i
n
er
d
etails
o
f
ten
m
is
s
ed
wh
en
tr
ain
in
g
o
n
en
tire
im
ag
es.
B
y
m
o
d
if
y
in
g
th
e
co
n
v
en
tio
n
al
b
in
ar
y
U
-
Net
in
to
a
m
u
lti
-
class
f
r
am
ewo
r
k
,
th
e
m
o
d
el
ef
f
icien
tly
s
eg
m
en
ts
co
m
p
lex
b
io
m
ed
ical
s
tr
u
ctu
r
es su
ch
as
r
etin
al
o
r
b
r
ain
r
e
g
io
n
s
.
3
.
3
.
2
.
H
H
O
f
o
r
weig
ht
up
da
t
e
a
nd
hy
perpa
ra
m
et
er
t
un
ing
T
h
e
HHO
alg
o
r
ith
m
is
a
well
-
k
n
o
wn
ex
a
m
p
le
o
f
a
p
o
p
u
latio
n
-
b
ased
m
etah
e
u
r
is
tic.
Ad
d
it
io
n
ally
,
it
h
as
a
s
ea
r
ch
m
ec
h
an
is
m
d
er
iv
ed
f
r
o
m
n
atu
r
e
th
at
m
im
ics
th
e
b
eh
av
io
u
r
o
f
Har
r
is
H
awk
s
in
h
u
n
tin
g
p
r
ey
,
wh
ich
ca
n
b
e
r
ep
r
esen
ted
b
y
t
h
e
f
o
llo
win
g
:
=
2
0
(
1
−
)
(
1
7
)
0
=
2
1
−
1
(
1
8
)
W
h
er
e
E
is
th
e
p
r
ey
'
s
f
lig
h
t
p
o
wer
,
t
is
th
e
in
itial
r
ep
etitio
n
n
u
m
b
er
,
T
is
th
e
m
a
x
im
u
m
r
ep
etitio
n
n
u
m
b
er
,
0
is
th
e
in
itial
p
r
ey
p
o
wer
with
a
r
an
d
o
m
in
teg
er
b
e
twee
n
-
1
an
d
1
,
an
d
1
is
a
r
an
d
o
m
p
ar
am
ete
r
b
etwe
en
0
an
d
1
.
Du
e
to
its
r
elatio
n
s
h
ip
to
th
e
p
r
e
y
'
s
p
o
wer
d
u
r
in
g
f
lig
h
t,
th
e
v
alu
e
o
f
E
d
ec
r
ea
s
es
s
tead
il
y
with
ea
ch
r
ep
etitio
n
.
T
o
r
ep
h
r
ase,
an
ex
p
l
o
r
atio
n
s
ea
r
ch
is
c
ar
r
ied
o
u
t
wh
en
th
e
h
awk
s
s
ea
r
ch
f
o
r
th
eir
p
r
e
y
in
d
if
f
er
en
t
p
lace
s
an
d
th
e
a
b
s
o
lu
te
v
alu
e
o
f
E
is
g
r
ea
ter
th
an
o
r
eq
u
al
to
1
.
T
h
e
h
awk
s
a
r
e
p
r
ep
ar
e
d
to
ca
p
tu
r
e
p
r
ey
if
th
e
ab
s
o
lu
te
v
alu
e
o
f
E
is
le
s
s
th
an
1
,
wh
ich
is
k
n
o
wn
as
an
ex
p
lo
itatio
n
s
ea
r
ch
.
B
o
th
th
e
ex
p
lo
r
atio
n
an
d
e
x
p
lo
itatio
n
s
ea
r
ch
s
tr
ateg
ies
ar
e
d
escr
ib
ed
i
n
f
u
r
th
e
r
d
ep
th
i
n
th
e
s
ec
tio
n
s
th
at
f
o
l
lo
w.
Fig
u
r
e
3
is
a
co
n
ce
p
tu
al
r
e
p
r
esen
tatio
n
o
f
th
e
h
awk
'
s
s
tr
ateg
y
ad
ap
tatio
n
in
r
esp
o
n
s
e
to
t
h
e
p
r
e
y
'
s
escap
e
en
er
g
y
a
n
d
lo
ca
tio
n
u
n
ce
r
t
ain
t
y
;
it
clea
r
ly
illu
s
tr
ates
th
e
tr
an
s
itio
n
d
y
n
am
ics
b
etwe
en
th
e
ex
p
lo
r
ati
o
n
an
d
e
x
p
lo
itatio
n
s
tag
es o
f
th
e
HHO
p
r
o
ce
s
s
.
a.
E
x
p
lo
r
atio
n
p
h
ase
T
h
e
f
ir
s
t
s
tag
e
o
f
h
u
n
tin
g
is
o
u
tlin
ed
h
er
e
,
an
d
it
en
tails
k
ee
p
in
g
a
n
ey
e
o
u
t
f
o
r
p
o
te
n
tial
p
r
ey
,
f
o
llo
win
g
its
p
ath
,
an
d
e
v
en
tu
ally
s
p
o
ttin
g
it.
T
h
is
s
tag
e
in
HHO
r
ep
r
esen
ts
th
e
m
ec
h
a
n
is
m
f
o
r
e
x
p
lo
r
atio
n
.
Har
r
is
h
awk
s
m
ay
s
ea
r
ch
f
o
r
p
r
ey
f
o
r
h
o
u
r
s
.
T
h
er
e
f
o
r
e,
t
h
e
Har
r
is
h
awk
s
(
i.e
.
,
p
o
ten
tial
s
o
lu
tio
n
s
)
d
eter
m
in
e
th
e
lik
elih
o
o
d
o
f
l
o
ca
tin
g
th
e
p
r
ey
(
i.e
.
,
tar
g
et
)
.
T
h
e
r
ef
o
r
e
,
i
t
s
tan
d
s
to
r
ea
s
o
n
th
at
t
h
e
o
p
ti
m
al
s
o
lu
tio
n
s
h
o
u
l
d
tar
g
et
th
e
o
n
e
clo
s
est
to
it.
Ha
r
r
is
h
awk
s
will
eith
er
wait
in
clo
s
e
p
r
o
x
im
ity
to
o
th
er
m
em
b
er
s
o
f
th
eir
f
am
ily
in
o
r
d
er
to
attac
k
at
th
e
s
am
e
m
o
m
en
t,
o
r
th
e
y
will
wait
in
a
v
ar
iety
o
f
s
itu
atio
n
s
,
s
u
ch
as
h
ig
h
tr
ee
s
,
to
ca
tch
p
r
ey
.
B
o
th
ca
s
es a
r
e
m
o
d
eled
i
n
th
e
f
o
llo
win
g
:
(
+
1
)
=
{
r
a
nd
(
)
−
2
|
r
a
nd
(
)
−
2
3
(
)
|
≥
0
.
5
(
pr
e
y
(
)
−
(
)
)
−
4
(
+
5
(
−
)
)
<
0
.
5
(
1
9
)
wh
er
e
X(
t+1
)
d
en
o
tes th
e
p
o
s
itio
n
s
o
f
th
e
n
ewly
iter
ated
h
a
wk
s
,
pr
e
y
r
ef
er
s
to
th
e
p
r
ey
'
s
lo
ca
ti
o
n
,
d
en
o
tes
th
e
av
er
ag
e
p
o
s
itio
n
o
f
th
e
in
itial
p
o
p
u
latio
n
,
r
a
nd
d
en
o
tes
a
h
awk
ch
o
s
en
at
r
an
d
o
m
f
r
o
m
th
e
s
ea
r
ch
s
p
ac
e,
an
d
X(
t)
d
e
n
o
tes
th
e
p
o
s
itio
n
s
o
f
th
e
in
itial
h
awk
s
,
wh
ich
ar
e
d
eter
m
in
ed
ac
co
r
d
in
g
to
E
q
.
1
9
.
Fo
r
ea
ch
iter
atio
n
t,
th
e
v
alu
es
o
f
q
,
2
,
3
,
4
,
an
d
5
ar
e
im
p
r
o
v
e
d
,
an
d
th
e
s
e
v
alu
es
ar
e
r
an
d
o
m
ly
c
h
o
s
e
n
f
r
o
m
th
e
in
ter
v
al
(
0
,
1
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
4
3
,
No
.
1
,
Ju
ly
20
2
6
:
157
-
17
8
166
(
)
=
1
∑
=
1
(
)
(
2
0
)
wh
er
e
r
ef
er
s
to
th
e
p
o
s
itio
n
o
f
h
awk
in
r
e
p
etitio
n
an
d
is
th
e
to
tal
am
o
u
n
t
o
f
h
aw
k
s
.
Fig
u
r
e
3
.
Pro
p
o
s
ed
HHO
wo
r
k
f
lo
w
b
.
E
x
p
l
o
itatio
n
p
h
ase
T
h
e
d
is
co
v
er
y
o
f
th
e
p
r
ey
d
u
r
in
g
th
e
ex
p
lo
r
atio
n
s
tag
e
(
also
k
n
o
wn
as
th
e
wid
e
s
ea
r
ch
)
m
ar
k
s
th
e
b
eg
in
n
in
g
o
f
th
is
s
tag
e.
A
Har
r
is
h
awk
will
attem
p
t
a
s
u
d
d
en
p
o
u
n
ce
o
n
its
p
r
ey
.
T
h
e
o
p
p
o
s
ite
is
tr
u
e
wh
en
th
e
v
ictim
m
ak
es
a
n
ef
f
o
r
t
t
o
f
le
e,
a
s
itu
atio
n
k
n
o
wn
as
s
ev
en
k
ills
[
3
0
]
.
HHO
cr
ea
te
d
a
m
o
d
el
o
f
f
o
u
r
p
o
s
s
ib
le
way
s
o
f
h
u
n
tin
g
an
d
ev
ad
in
g
ca
p
tu
r
e.
W
h
er
e
th
at
was
p
r
o
p
o
s
ed
as
a
r
an
d
o
m
n
u
m
b
er
t
o
r
ef
e
r
th
e
o
p
p
o
r
tu
n
ity
o
f
p
r
e
y
i
n
s
u
cc
ess
f
u
lly
f
lig
h
t
(
<
0
.
5
)
,
in
c
o
n
tr
ast
(
≥
0
.
5
)
in
t
h
e
ev
e
n
t
o
f
f
ailu
r
e.
Als
o
,
th
e
h
awk
s
will b
e
b
ased
o
n
a
s
o
f
t o
r
h
ar
d
b
lo
ck
a
d
e
to
ca
p
tu
r
e
th
e
p
r
ey
b
ased
o
n
th
e
p
r
e
y
p
o
we
r
E
.
Fo
r
in
s
tan
ce
,
if
th
e
b
lo
ck
a
d
e
is
s
o
f
t th
e
r
ep
r
esen
tatio
n
will b
e
|
|
≥
0
.
5
,
o
th
er
wis
e
|
|
<
0
.
5
.
So
f
t
b
esieg
e
:
I
n
th
e
ca
s
e
o
f
|
E
|
≥
0
.
5
an
d
r
≥
0
.
5
.
As
a
r
esu
lt,
th
e
p
r
ey
h
as
ad
eq
u
ate
e
n
er
g
y
to
escap
e
f
r
o
m
th
e
h
awk
s
b
y
u
tili
zin
g
a
v
ar
iety
o
f
d
ec
eiv
in
g
ju
m
p
s
an
d
u
n
p
r
ed
ictab
le
p
atter
n
s
.
T
h
e
f
ac
t
th
at
th
e
h
ar
r
is
h
awk
s
wo
u
ld
d
r
ain
its
v
itality
b
y
en
cir
clin
g
it
an
d
th
en
s
u
r
p
r
is
e
attac
k
in
g
en
s
u
r
es
t
h
at
it
will
co
llap
s
e.
T
h
is
b
eh
av
io
r
ca
n
b
e
s
ee
n
in
th
e
m
o
d
el
in
(
2
1
)
.
(
+
1
)
=
Δ
−
|
pr
e
y
(
)
−
(
)
|
(
2
1
)
Δ
(
)
=
pr
e
y
(
)
−
(
)
(
2
2
)
=
2
×
(
1
−
6
)
(
2
3
)
wh
er
e
Δ
X
is
th
e
p
o
s
itio
n
o
f
th
e
d
is
tin
ctio
n
r
elativ
e
to
th
e
p
r
ey
s
an
d
th
e
s
tar
tin
g
p
o
in
t
in
r
ep
etitio
n
t,
6
is
an
ar
b
itra
r
y
in
teg
e
r
b
etwe
en
0
a
n
d
1
,
an
d
J
is
th
e
p
r
ey
'
s
r
an
d
o
m
ju
m
p
,
wh
e
r
e
it
(
i.e
.
,
J
)
ch
an
g
es
at
r
an
d
o
m
t
o
im
itate
th
e
p
r
ey
'
s
m
o
v
em
en
ts
.
Ha
r
d
b
esieg
e
:
T
o
b
e
s
p
ec
if
ic
,
r
≥
0
.
5
an
d
E
<
0
.
5
.
So
,
th
e
p
r
ey
ca
n
'
t
g
et
awa
y
s
in
ce
it
d
o
esn
'
t
h
av
e
en
o
u
g
h
e
n
er
g
y
.
Alo
n
g
with
t
h
at,
h
awk
s
ar
e
p
r
ac
tically
p
r
ep
ar
e
d
to
s
u
r
p
r
is
e
th
eir
p
r
e
y
b
y
cir
clin
g
it.
In
(
2
4
)
s
h
o
ws h
o
w
th
e
p
r
esen
t lo
ca
tio
n
s
o
f
th
is
cir
cu
m
s
tan
ce
h
av
e
b
ee
n
u
p
d
ate
d
.
(
+
1
)
=
pr
e
y
(
)
−
|
Δ
(
)
|
(
2
4
)
So
f
t
b
esieg
e
with
p
r
o
g
r
ess
iv
e
r
ap
id
d
iv
es
:
T
h
is
s
ce
n
ar
io
is
m
o
r
e
co
m
p
lex
th
an
th
e
ea
r
lier
ca
s
es
an
d
o
cc
u
r
s
wh
en
∣
∣
≥
0
.
5
an
d
<
0
.
5
.
Un
d
er
th
e
s
e
co
n
d
itio
n
s
,
th
e
p
r
ey
p
o
s
s
ess
es
en
o
u
g
h
s
tr
en
g
th
to
e
v
ad
e
ca
p
tu
r
e
s
u
cc
ess
f
u
lly
.
Nev
e
r
th
eless
,
th
e
h
awk
s
p
er
s
is
t
b
y
p
e
r
f
o
r
m
in
g
m
u
ltip
le
r
ap
id
d
iv
es
to
f
o
r
ce
th
e
p
r
ey
to
alter
its
tr
ajec
to
r
y
an
d
b
ec
o
m
e
d
is
o
r
ien
ted
.
T
h
is
s
tr
ateg
y
is
r
ep
ea
ted
u
n
til
an
o
p
tim
al
m
o
m
en
t
ar
is
es
f
o
r
ca
p
tu
r
in
g
th
e
p
r
ey
.
T
h
e
m
ath
e
m
atica
l
ex
p
r
ess
io
n
b
elo
w
r
ep
r
esen
ts
th
e
h
awk
s
’
m
o
v
e
m
en
t
d
ec
is
io
n
d
u
r
i
n
g
th
e
s
o
f
t e
n
cir
clem
en
t p
h
ase.
=
pr
e
y
(
)
−
|
pr
e
y
(
)
−
(
)
|
(
2
5
)
I
f
th
e
h
awk
s
f
ee
l
th
at
th
e
p
r
e
y
p
r
ef
o
r
m
s
m
is
lead
in
g
m
o
v
e
m
en
ts
an
d
it
alm
o
s
t
escap
e,
t
h
ey
will
in
cr
ea
s
e
th
e
ab
r
u
p
t,
ir
r
eg
u
lar
,
an
d
r
ap
id
d
i
v
es.
T
h
e
n
ew
h
awk
s
'
tech
n
iq
u
e
d
ep
en
d
s
o
n
lev
y
f
lig
h
ts
(
L
F)
as
s
h
o
wn
in
th
e
f
o
llo
win
g
eq
u
atio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.