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
.
192
~
20
6
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
192
-
20
6
192
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
A hybrid a
pp
ro
a
ch f
o
r mul
ti
-
v
iew
M
RI Alzehim
er
’
s
detec
tion usin
g
co
nv
o
lutiona
l neural networks
and
bio
-
inspi
red alg
o
r
ithms
I
heb Chem
s
s
E
l D
ine H
a
g
a
ni
1
,
Na
cé
ra
B
ena
m
ra
ne
1
,
L
a
kh
da
r
Sa
is
2
1
S
I
M
P
A
La
b
o
r
a
t
o
r
y
,
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
ma
t
i
c
s,
U
n
i
v
e
r
si
t
y
o
f
S
c
i
e
n
c
e
s
a
n
d
T
e
c
h
n
o
l
o
g
y
o
f
O
r
a
n
-
M
o
h
a
m
e
d
B
o
u
d
i
a
f
,
O
r
a
n
,
A
l
g
e
r
i
a
2
C
e
n
t
r
e
d
e
R
e
c
h
e
r
c
h
e
e
n
I
n
f
o
r
mat
i
q
u
e
d
e
Le
n
s (
C
R
I
L)
,
C
N
R
S
,
U
n
i
v
e
r
s
i
t
é
d
’
A
r
t
o
i
s,
Le
n
s
,
F
r
a
n
c
e
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
r
2
3
,
2
0
2
6
R
ev
is
ed
J
u
n
2
2
,
2
0
2
6
Acc
ep
ted
J
u
n
2
9
,
2
0
2
6
Alz
h
e
ime
r
’
s
d
ise
a
se
(AD
)
is
a
n
e
u
ro
d
e
g
e
n
e
ra
ti
v
e
d
is
o
rd
e
r
th
a
t
re
m
a
in
s
in
c
u
ra
b
le
to
d
a
te.
Th
e
re
fo
re
,
t
h
e
m
o
st
imp
o
rtan
t
ste
p
i
n
trea
tme
n
t
re
m
a
in
s
th
e
e
a
rly
d
e
tec
ti
o
n
o
f
t
h
e
sig
n
s
in
d
ica
ti
n
g
it
s
p
re
se
n
c
e
.
T
h
e
so
o
n
e
r
t
h
e
se
sig
n
s
a
re
d
isc
o
v
e
re
d
,
t
h
e
so
o
n
e
r
p
re
v
e
n
tati
v
e
c
a
re
c
a
n
b
e
a
d
m
in
istere
d
.
Co
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
(CNN
s)
h
a
v
e
d
e
m
o
n
stra
ted
i
m
p
re
ss
iv
e
p
e
rfo
rm
a
n
c
e
in
m
e
d
ica
l
ima
g
e
a
n
a
ly
sis;
h
o
we
v
e
r,
th
e
y
o
ften
s
u
ffe
r
fro
m
su
b
o
p
ti
m
a
l
m
a
n
u
a
l
tu
n
in
g
o
f
th
e
i
r
h
y
p
e
rp
a
ra
m
e
ters
.
Th
e
re
fo
re
,
we
o
p
te
d
fo
r
a
h
y
b
ri
d
m
e
th
o
d
c
o
m
b
in
i
n
g
th
e
m
with
g
e
n
e
ti
c
a
lg
o
ri
th
m
s
(G
A)
a
n
d
p
a
rti
c
le
sw
a
rm
o
p
ti
m
iza
ti
o
n
(P
S
O)
to
a
u
to
m
a
ti
c
a
ll
y
o
p
ti
m
ize
a
rc
h
it
e
c
tu
re
s
a
n
d
fu
sio
n
we
ig
h
ts
fo
r
imp
r
o
v
e
d
AD
d
e
tec
ti
o
n
.
Us
in
g
d
a
ta
o
b
tain
e
d
fr
o
m
AD
NI
a
n
d
Ka
g
g
le,
o
u
r
a
p
p
ro
a
c
h
a
c
h
ie
v
e
d
8
7
.
4
%
a
c
c
u
ra
c
y
,
su
rp
a
ss
i
n
g
c
las
sic
a
l
CNN
s
o
f
th
e
sa
m
e
siz
e
a
n
d
d
e
p
th
.
T
h
e
se
re
su
lt
s
h
ig
h
li
g
h
t
t
h
e
p
o
ten
ti
a
l
o
f
e
v
o
lu
ti
o
n
a
ry
o
p
ti
m
iza
ti
o
n
f
o
r
d
e
v
e
lo
p
i
n
g
re
li
a
b
le
d
ia
g
n
o
stic to
o
ls
.
K
ey
w
o
r
d
s
:
Alzh
eim
er
’
s
d
is
ea
s
e
B
io
-
in
s
p
ir
ed
o
p
tim
izatio
n
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
Gen
etic
alg
o
r
ith
m
MRI
Par
ticle
s
war
m
o
p
tim
izatio
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
:
I
h
eb
C
h
em
s
s
E
l D
in
e
Hag
an
i
SIM
PA L
ab
o
r
ato
r
y
,
Dep
ar
tm
e
n
t o
f
I
n
f
o
r
m
atics
Un
iv
er
s
ity
o
f
Scien
ce
s
an
d
T
e
ch
n
o
lo
g
y
o
f
Or
an
-
Mo
h
am
e
d
B
o
u
d
iaf
Or
an
,
Alg
er
ia
E
m
ail: ih
eb
ch
em
s
s
eld
in
e.
h
ag
a
n
i@
u
n
iv
-
u
s
to
.
d
z
1.
I
NT
RO
D
UCT
I
O
N
Alzh
eim
er
’
s
d
is
ea
s
e
(
AD)
p
r
esen
ts
a
s
ig
n
if
ican
t
ch
allen
g
e
f
o
r
m
o
d
er
n
m
ed
icin
e
,
an
d
m
ag
n
etic
r
eso
n
an
ce
im
a
g
in
g
(
MRI)
r
e
m
ain
s
th
e
m
o
d
ality
o
f
c
h
o
ice
f
o
r
clin
ician
s
.
Ho
wev
er
,
m
an
u
al
an
aly
s
is
o
f
MRIs
is
a
tim
e
-
co
n
s
u
m
in
g
task
.
T
h
is
is
wh
y
th
is
d
ec
ad
e
h
as
s
ee
n
th
e
em
er
g
e
n
ce
o
f
a
r
tific
ial
in
tellig
en
ce
(
AI
)
ap
p
licatio
n
s
in
th
e
m
ed
ical
f
ield
in
g
en
er
al,
a
n
d
m
o
r
e
s
p
ec
if
i
ca
lly
in
im
ag
in
g
.
Dee
p
lear
n
in
g
(
DL
)
,
a
s
u
b
s
et
o
f
AI
,
m
o
r
e
s
p
ec
if
ically
c
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
(
C
NN
)
,
h
as
r
ev
o
lu
tio
n
ized
i
m
ag
e
an
aly
s
is
an
d
,
co
n
s
eq
u
en
tly
,
m
e
d
ical
im
ag
e
an
aly
s
is
.
B
y
au
to
m
atin
g
ce
r
tain
s
tep
s
s
u
ch
as
f
ea
tu
r
e
ex
tr
ac
tio
n
,
s
eg
m
en
tatio
n
an
d
class
if
icatio
n
[
1
]
-
[
5
]
,
th
ese
to
o
ls
h
av
e
allo
wed
s
p
ec
ialis
ts
to
f
o
cu
s
th
eir
ex
p
er
tis
e
o
n
in
ter
p
r
etin
g
r
esu
lts
,
ea
r
ly
d
etec
tin
g
s
u
b
tle
ab
n
o
r
m
alities
,
an
d
d
ev
elo
p
i
n
g
a
p
er
s
o
n
alize
d
ca
r
e
p
lan
f
o
r
th
e
p
at
ien
t.
Desp
ite
th
eir
s
u
cc
ess
,
C
N
N
-
b
ased
ap
p
r
o
ac
h
es
f
o
r
AD
class
if
icatio
n
f
ac
e
t
wo
p
er
s
is
ten
t
lim
itatio
n
s
.
First
,
ar
ch
itectu
r
es
an
d
h
y
p
er
p
ar
am
eter
s
ar
e
t
y
p
ically
d
esig
n
ed
th
r
o
u
g
h
m
an
u
al
tr
ial
-
an
d
-
er
r
o
r
,
a
co
s
tly
an
d
n
o
n
-
r
e
p
r
o
d
u
cib
le
p
r
o
ce
s
s
th
at
r
estricts
th
e
ex
p
lo
r
atio
n
o
f
th
e
d
esig
n
s
p
ac
e.
Seco
n
d
,
w
h
en
p
r
ed
ictio
n
s
f
r
o
m
m
u
ltip
le
an
ato
m
ical
v
iews
ar
e
co
m
b
i
n
ed
,
u
n
if
o
r
m
a
v
er
ag
in
g
is
co
m
m
o
n
ly
ap
p
lied
r
ath
er
th
a
n
a
d
ata
-
d
r
iv
e
n
we
ig
h
tin
g
s
ch
em
e
-
an
ap
p
r
o
ac
h
th
at
d
is
r
eg
ar
d
s
th
e
v
ar
y
in
g
d
is
cr
im
in
ativ
e
p
o
wer
o
f
ea
ch
o
r
ien
tatio
n
.
B
ef
o
r
e
th
e
ad
v
e
n
t
o
f
C
NNs,
b
io
-
in
s
p
ir
ed
a
p
p
r
o
ac
h
es
h
ad
alr
ea
d
y
d
em
o
n
s
tr
ated
th
e
p
o
ten
tial
o
f
s
u
ch
o
p
tim
izatio
n
i
n
th
e
m
ed
ical
f
ield
.
Gen
etic
alg
o
r
ith
m
s
(
GA)
[
6
]
,
[
7
]
a
n
d
p
ar
ticle
s
war
m
o
p
tim
izatio
n
(
PS
O)
[
8
]
o
f
f
er
a
m
u
ltit
u
d
e
o
f
p
o
s
s
ib
le
ap
p
licatio
n
s
[
9
]
,
[
1
0
]
.
B
ased
o
n
th
ese
o
b
s
er
v
atio
n
s
,
we
p
r
o
p
o
s
e
a
h
y
b
r
id
izatio
n
o
f
C
NNs
with
GA
an
d
PS
O,
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
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mu
lti
-
view MR
I
A
lz
eh
imer’
s
d
etec
tio
n
u
s
in
g
…
(
I
h
eb
C
h
emss
E
l D
in
e
Ha
g
a
n
i
)
193
ex
ten
d
in
g
a
p
r
elim
in
ar
y
s
tu
d
y
[
1
1
]
t
h
at
co
m
b
in
ed
C
NNs
with
PS
O
f
o
r
Alzh
eim
er
’
s
d
etec
tio
n
.
T
h
e
k
e
y
co
n
tr
ib
u
tio
n
o
f
th
is
wo
r
k
is
t
wo
f
o
ld
:
(
i)
GA
au
to
m
ates
th
e
s
ea
r
ch
f
o
r
C
NN
ar
ch
itectu
r
e
s
ad
ap
ted
to
ea
ch
an
ato
m
ical
o
r
ien
tatio
n
,
r
e
m
o
v
in
g
th
e
n
ee
d
f
o
r
m
an
u
al
d
es
ig
n
;
an
d
(
ii)
PS
O
is
em
p
lo
y
e
d
n
o
t
as
a
n
etwo
r
k
weig
h
t
o
p
tim
izer
b
u
t
to
d
eter
m
in
e
o
p
tim
al
f
u
s
io
n
weig
h
ts
f
o
r
c
o
m
b
in
i
n
g
o
r
ien
tatio
n
-
s
p
e
cif
ic
p
r
ed
ictio
n
s
-
a
n
o
v
el
ap
p
licatio
n
o
f
PS
O
t
o
th
e
m
u
lti
-
v
iew
f
u
s
io
n
p
r
o
b
l
em
.
W
e
ch
o
s
e
to
wo
r
k
with
2
D
s
lices
ex
tr
ac
ted
f
r
o
m
3
D
v
o
lu
m
es.
T
h
is
ch
o
ice
was
f
ac
ilit
ated
b
y
th
e
lo
wer
co
m
p
u
tatio
n
al
c
o
s
t
o
f
2
D
ar
ch
it
ec
tu
r
es,
a
d
ec
is
iv
e
f
ac
to
r
with
in
th
e
GA
o
p
tim
i
za
tio
n
lo
o
p
wh
er
e
n
u
m
e
r
o
u
s
ca
n
d
i
d
ate
m
o
d
els
a
r
e
tr
ain
e
d
iter
ativ
ely
.
Mo
r
e
f
u
n
d
am
e
n
tally
,
t
h
e
m
u
lti
-
v
ie
w
ap
p
r
o
ac
h
its
elf
is
b
u
ilt
ar
o
u
n
d
th
e
2
D
r
ep
r
esen
tatio
n
:
d
ec
o
m
p
o
s
in
g
ea
ch
v
o
lu
m
e
in
to
ax
ial,
co
r
o
n
al,
a
n
d
s
ag
ittal
s
lice
s
is
p
r
ec
is
ely
wh
at
y
ield
s
th
e
th
r
ee
d
is
tin
ct
v
iews
th
at
P
SO
s
u
b
s
eq
u
en
tly
weig
h
ts
an
d
f
u
s
es.
A
v
o
lu
m
e
p
r
o
ce
s
s
ed
d
ir
ec
tly
b
y
a
s
in
g
le
3
D
n
etwo
r
k
wo
u
ld
ex
p
o
s
e
n
o
s
ep
ar
ate
o
r
ien
tatio
n
s
to
co
m
b
in
e,
an
d
th
e
f
u
s
io
n
m
ec
h
an
is
m
at
th
e
co
r
e
o
f
o
u
r
co
n
tr
ib
u
tio
n
wo
u
ld
lo
s
e
its
o
b
ject.
As
in
d
icate
d
i
n
th
e
ti
tle,
“
m
u
lti
-
v
iew
”
th
u
s
r
e
f
er
s
to
th
e
th
r
ee
p
r
in
ci
p
al
an
ato
m
ical
p
lan
es
(
ax
ial,
co
r
o
n
al,
an
d
s
ag
ittal).
I
n
s
ec
tio
n
2
th
en
d
etails
th
e
p
r
o
p
o
s
ed
h
y
b
r
i
d
m
eth
o
d
o
lo
g
y
.
Sectio
n
3
.
p
r
esen
ts
an
d
an
aly
ze
s
th
e
e
x
p
er
im
en
tal
r
e
s
u
lts
,
an
d
s
ec
tio
n
4
.
d
r
aws
co
n
clu
s
io
n
s
an
d
o
u
tlin
es
d
ir
ec
tio
n
s
f
o
r
f
u
tu
r
e
r
esear
ch
.
2.
RE
L
AT
E
D
WO
RK
Fo
r
s
ev
er
al
y
ea
r
s
n
o
w,
th
e
a
p
p
licatio
n
o
f
DL
h
as
r
ea
ch
ed
th
e
ex
p
an
s
io
n
s
tag
e
in
th
e
f
ield
o
f
m
e
d
ical
im
ag
e
an
aly
s
is
.
Am
o
n
g
th
e
m
o
s
t
wid
ely
u
s
ed
m
o
d
els
ar
e
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
,
d
em
o
n
s
tr
atin
g
u
n
p
r
ec
e
d
en
ted
p
o
te
n
tial.
I
n
2
0
2
5
,
a
c
o
m
p
r
eh
e
n
s
iv
e
r
ev
i
ew
o
f
DL
tech
n
iq
u
es
f
o
r
th
e
d
etec
tio
n
o
f
AD
h
ig
h
lig
h
ted
th
at
“
C
N
N
s
a
r
e
c
o
n
s
is
ten
tly
fo
u
n
d
to
o
u
tp
erfo
r
m
tr
a
d
itio
n
a
l
ML
a
p
p
r
o
a
ch
es,
s
h
o
w
ca
s
in
g
th
eir
p
o
ten
tia
l
in
imp
r
o
vin
g
d
ia
g
n
o
s
tic
a
cc
u
r
a
cy
”
[
1
2
]
.
Ho
wev
er
,
in
th
e
s
am
e
s
tu
d
y
,
th
e
au
t
h
o
r
s
n
o
ted
s
ig
n
if
ican
t
an
d
p
er
s
is
ten
t
d
if
f
icu
lties
r
elate
d
to
m
o
d
el
s
elec
tio
n
an
d
o
p
tim
izatio
n
,
lead
in
g
u
s
to
ex
p
lo
r
e
au
to
m
ated
s
o
lu
tio
n
s
.
T
h
is
ch
ap
ter
will
s
er
v
e
as
a
s
y
n
th
esis
o
f
cu
r
r
en
t
r
esear
ch
c
o
m
b
in
in
g
DL
an
d
b
io
-
in
s
p
ir
e
d
m
eth
o
d
s
ap
p
lied
to
th
e
d
etec
tio
n
o
f
AD
.
El
-
Ass
y
et
a
l.
[
1
3
]
,
a
d
u
al
-
C
NN
m
o
d
el
with
d
is
tin
ct
f
ilter
s
co
n
ca
ten
ated
in
to
a
f
iv
e
-
o
u
t
p
u
t
lay
er
(
AD,
L
MCI,
MCI,
E
MCI,
a
n
d
NC
)
on
A
DNI
d
ata
with
A
DASYN
-
b
as
ed
class
-
im
b
alan
ce
co
r
r
ec
tio
n
.
I
t
r
ep
o
r
te
d
u
p
to
9
9
.
4
3
%,
9
9
.
5
7
%,
a
n
d
9
9
.
3
0
%
ac
cu
r
ac
y
f
o
r
3
-
,
4
-
,
a
n
d
5
-
way
class
if
icatio
n
,
r
esp
ec
tiv
ely
.
Sh
ar
m
a
et
a
l.
[
1
4
]
,
th
e
au
th
o
r
s
o
b
tain
e
d
ac
cu
r
ac
y
,
r
ec
all,
a
n
d
F1
-
s
co
r
es
o
f
9
0
.
4
%,
0
.
9
0
5
,
0
.
9
0
4
,
an
d
0
.
9
0
4
,
r
esp
ec
tiv
ely
.
T
h
e
d
ataset
was
co
llected
u
s
in
g
Ka
g
g
le.
I
t
co
m
p
r
is
es
6
,
4
0
0
im
ag
es
d
is
tr
ib
u
ted
ac
r
o
s
s
two
f
o
ld
er
s
(
tr
ain
in
g
an
d
test
)
.
E
ac
h
f
o
ld
er
is
f
u
r
th
er
d
i
v
id
ed
in
to
f
o
u
r
s
u
b
f
o
ld
er
s
: M
ild
-
Dem
en
ted
,
Mo
d
er
ate
-
Dem
en
ted
,
No
n
-
Dem
en
ted
,
an
d
Ver
y
Mild
-
Dem
e
n
ted
.
T
h
e
a
u
th
o
r
s
u
s
ed
th
e
VGG
-
1
6
ar
ch
i
tectu
r
e
as
a
f
ea
tu
r
e
ex
tr
ac
to
r
,
a
n
d
an
o
th
er
C
N
N
was th
en
u
s
ed
as a
class
if
ier
(
4
o
u
tp
u
ts
)
.
C
er
tain
lim
itatio
n
s
h
av
e
b
ee
n
r
ea
ch
ed
with
in
th
e
2
D
r
esear
ch
co
m
m
u
n
ity
.
T
h
is
h
as
d
r
iv
en
th
e
ev
o
lu
tio
n
o
f
r
esear
c
h
to
war
d
s
3
D
p
ar
ad
i
g
m
s
.
A
s
tu
d
y
p
u
b
lis
h
ed
in
2
0
2
5
in
tr
o
d
u
ce
d
th
e
3D
-
C
NN
-
VSwin
Fo
r
m
er
m
o
d
el
[
1
5
]
,
w
h
i
ch
co
m
b
in
es
v
o
lu
m
e
m
et
r
ic
co
n
v
o
lu
tio
n
with
tr
an
s
f
o
r
m
er
-
b
ased
atten
tio
n
m
ec
h
an
is
m
s
,
ac
h
iev
in
g
92
.
92%
ac
cu
r
ac
y
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
p
atien
ts
with
AD
an
d
h
ea
lth
y
in
d
iv
id
u
als.
T
h
e
C
NN
u
s
ed
in
th
is
s
tu
d
y
was
in
s
p
ir
ed
b
y
th
e
m
o
d
e
l
p
r
o
p
o
s
ed
in
2
0
1
8
by
[
1
6
]
,
n
am
ed
C
B
AM
f
o
r
co
n
v
o
l
u
tio
n
al
b
lo
c
k
atten
tio
n
m
o
d
u
le
.
Ho
wev
er
,
th
e
in
itial
m
o
d
el
was
cr
ea
te
d
f
o
r
2
D,
an
d
s
in
ce
th
e
au
th
o
r
s
f
o
cu
s
ed
on
3D
v
o
lu
m
es,
s
o
m
e
m
o
d
if
ica
-
tio
n
s
h
ad
to
b
e
m
ad
e.
R
eg
ar
d
in
g
th
e
v
i
d
eo
Swin
Fo
r
m
er
co
m
p
o
n
en
t,
th
is
b
lo
ck
u
s
es a
n
atten
tio
n
m
ec
h
an
is
m
b
ased
o
n
a
3
D
s
lid
in
g
win
d
o
w
m
u
ltih
ea
d
s
elf
-
atten
tio
n
(
MSA)
.
T
o
co
n
d
u
ct
th
eir
e
x
p
er
im
e
n
t
[
1
7
]
,
u
s
ed
th
e
ADNI
d
atab
ase,
s
p
ec
if
ically
th
e
ADNI
3
co
llec
tio
n
.
T
h
is
s
tu
d
y
is
ch
a
r
ac
ter
ized
b
y
its
o
u
tp
u
t
class
es:
ea
r
ly
m
ild
c
o
g
n
itiv
e
im
p
ai
r
m
en
t
(
E
MCI)
a
n
d
n
o
r
m
al
co
n
tr
o
l
(
NC
)
,
with
th
e
o
b
jectiv
e
o
f
ea
r
ly
d
etec
tio
n
o
f
th
e
d
is
ea
s
e.
I
n
itially
,
th
ey
p
er
f
o
r
m
e
d
p
r
e
-
p
r
o
ce
s
s
in
g
co
n
s
is
tin
g
of
r
eg
is
tr
atio
n
u
s
in
g
th
e
s
tan
d
ar
d
MN
I
1
5
2
m
o
d
el
an
d
s
k
u
ll
ex
tr
ac
tio
n
u
s
in
g
th
e
FMB
I
R
s
o
f
twar
e
lib
r
a
r
y
(
FS
L
)
.
T
h
e
d
ata
was
allo
ca
ted
8
0
%
f
o
r
tr
ain
in
g
a
n
d
v
alid
ati
o
n
an
d
2
0
%
f
o
r
test
in
g
.
T
h
is
m
o
d
el
ac
h
iev
ed
an
ac
cu
r
ac
y
of
81
.
8
0
%,
a
r
ec
all
of
82
.
5
0
%,
an
d
a
s
p
ec
if
icity
of
80
.
5
0
%,
t
h
u
s
ef
f
ec
tiv
ely
d
is
ti
n
g
u
is
h
in
g
b
etwe
en
in
d
iv
id
u
als with
m
ild
co
g
n
itiv
e
im
p
air
m
en
t (
MCI)
a
n
d
th
o
s
e
with
n
o
r
m
al
c
o
g
n
itiv
e
f
u
n
ctio
n
(
NC
F).
W
ith
th
e
r
is
e
o
f
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
an
d
th
e
d
if
f
icu
lties
en
co
u
n
ter
e
d
b
y
class
ical
alg
o
r
ith
m
s
[
1
8
]
,
b
io
-
in
s
p
ir
ed
a
lg
o
r
ith
m
s
h
av
e
r
et
u
r
n
e
d
to
t
h
e
f
o
r
ef
r
o
n
t
o
f
th
e
s
cien
tific
s
ce
n
e
as
a
p
o
wer
f
u
l
co
m
p
u
tatio
n
al
m
eth
o
d
f
o
r
ex
p
lo
r
in
g
co
m
p
lex
an
d
m
u
ltid
im
en
s
io
n
al
p
ar
am
eter
s
[
1
9
]
.
T
h
ese
m
etah
eu
r
is
tics
o
f
f
er
r
o
b
u
s
t
an
d
ad
ap
tiv
e
ap
p
r
o
ac
h
es
to
o
p
tim
izatio
n
p
r
o
b
lem
s
th
at
h
av
e
p
r
o
v
en
d
if
f
i
cu
lt
to
s
o
lv
e
u
s
in
g
tr
ad
itio
n
al
m
eth
o
d
s
.
I
n
th
is
s
tu
d
y
[
2
0
]
,
f
o
r
ex
am
p
le,
th
e
au
th
o
r
s
co
m
b
in
ed
AI
B
L
an
d
A
DNI
d
ata
in
a
f
o
u
r
-
s
tag
e
f
r
am
ewo
r
k
:
f
u
zz
y
-
lo
g
ic
R
OI
clu
s
ter
in
g
,
f
ea
tu
r
e
ex
tr
ac
tio
n
v
ia
PLT
P,
R
e
s
Net
-
5
0
an
d
VGG
-
1
6
,
f
ea
tu
r
e
s
elec
tio
n
with
th
e
m
o
d
if
ied
g
o
r
illa
tr
o
o
p
o
p
tim
izer
(
MG
T
O)
,
an
d
C
ap
s
Net
class
if
icatio
n
.
T
h
e
m
o
d
el
r
ep
o
r
ted
99
.
7
6
%,
9
9
.
8
8
%,
an
d
9
9
.
9
2
%
ac
cu
r
ac
y
f
o
r
AD
v
s
.
NC
,
MCI
v
s
.
AD,
an
d
NC
v
s
.
MCI,
r
es
p
ec
tiv
ely
.
A
h
y
b
r
id
m
o
d
el
c
o
m
b
in
in
g
a
c
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
a
n
d
a
s
p
ik
in
g
n
eu
r
al
n
etwo
r
k
(
C
NN
-
SNN)
was
p
r
o
p
o
s
ed
in
th
is
s
tu
d
y
[
2
1
]
.
T
h
is
h
y
b
r
id
m
o
d
el
lev
er
ag
es
t
h
e
s
p
atial
f
ea
tu
r
e
ex
tr
ac
tio
n
ca
p
ab
ilit
ies
of
C
NNs
an
d
th
e
tem
p
o
r
al
d
y
n
am
ics
of
SNNs
.
I
n
o
th
er
wo
r
d
s
,
th
e
C
N
N
m
o
d
u
le
ex
tr
ac
ts
f
ea
tu
r
es
f
r
o
m
th
e
in
p
u
t
im
ag
e,
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
:
192
-
20
6
194
wh
ile
th
e
SNN
m
o
d
u
le
p
r
o
c
ess
es
th
em
o
v
er
tim
e.
Acc
o
r
d
in
g
to
th
e
au
th
o
r
s
,
th
is
co
m
b
in
atio
n
allo
ws
th
e
m
o
d
el
to
ca
p
tu
r
e
b
o
th
s
p
atial
an
d
tem
p
o
r
al
in
f
o
r
m
atio
n
,
m
ak
i
n
g
it
s
u
itab
le
f
o
r
task
s
r
eq
u
ir
in
g
an
u
n
d
er
s
tan
d
in
g
of
th
e
s
tr
u
ctu
r
e
an
d
d
y
n
am
ics
o
f
th
e
in
p
u
t
d
ata.
To
h
ig
h
lig
h
t
th
e
im
p
o
r
tan
ce
of
th
e
SNN,
th
e
au
th
o
r
s
co
n
d
u
cted
an
ab
latio
n
s
tu
d
y
:
its
r
em
o
v
al
r
e
d
u
ce
d
ac
c
u
r
ac
y
f
r
o
m
99
.
58%
to
75
.
6
7
%.
T
h
e
p
r
ep
r
o
ce
s
s
in
g
s
tep
co
n
s
is
ted
of
im
ag
e
re
-
im
p
o
r
tin
g
,
d
ata
a
u
g
m
en
tatio
n
,
an
d
u
p
s
am
p
lin
g
(
S
MO
T
E
[
2
2
]
)
.
A
2
0
2
5
s
tu
d
y
f
u
r
th
er
co
n
f
ir
m
ed
th
e
d
iag
n
o
s
tic
p
o
ten
tial
o
f
d
ee
p
C
NN
ar
ch
itectu
r
es
f
o
r
ea
r
ly
AD
d
etec
tio
n
f
r
o
m
MRI
d
ata
[
2
3
]
,
r
ein
f
o
r
cin
g
th
e
r
elev
an
ce
o
f
co
n
v
o
lu
ti
o
n
al
ap
p
r
o
ac
h
es a
s
a
s
o
lid
f
o
u
n
d
atio
n
f
o
r
AD
d
iag
n
o
s
tic
s
y
s
tem
s
.
@
ar
ticlep
eter
s
en
2
0
1
0
alz
h
eim
er
,
tit
le=
Alzh
eim
er
’
s
d
is
ea
s
e
Neu
r
o
im
ag
in
g
I
n
itiativ
e
(
ADNI
)
clin
ical
ch
a
r
ac
ter
izatio
n
,
a
u
th
o
r
=Pete
r
s
en
,
R
o
n
ald
C
ar
l
an
d
Ais
en
,
Pau
l
S
an
d
B
ec
k
ett,
L
au
r
el
A
an
d
Do
n
o
h
u
e
,
Mic
h
ae
l
C
an
d
Gam
s
t,
An
th
o
n
y
C
o
llin
s
an
d
Har
v
ey
,
Dan
ielle
J
an
d
J
ac
k
J
r
,
C
liff
o
r
d
R
an
d
J
ag
u
s
t,
W
ill
iam
J
an
d
Sh
aw,
L
e
s
lie
M
an
d
T
o
g
a,
Ar
th
u
r
W
an
d
o
th
er
s
,
jo
u
r
n
al=
Neu
r
o
l
o
g
y
,
v
o
l
u
m
e=
7
4
,
n
u
m
b
er
=
3
,
p
ag
es=2
0
1
–
2
0
9
,
y
ea
r
=
2
0
1
0
,
p
u
b
lis
h
er
=L
ip
p
i
n
co
tt
W
illi
am
s
&
W
ilk
in
s
T
h
e
in
teg
r
atio
n
o
f
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
with
b
i
o
-
in
s
p
ir
ed
o
p
tim
izatio
n
al
g
o
r
ith
m
s
r
ep
r
esen
ts
a
n
em
e
r
g
in
g
f
r
o
n
tier
in
m
ed
ical
im
ag
in
g
r
esear
ch
,
t
h
o
u
g
h
s
p
e
-
cif
ic
ap
p
licatio
n
s
t
o
AD
d
ete
ctio
n
r
em
ain
r
elativ
ely
u
n
d
er
d
ev
elo
p
e
d
.
C
u
r
r
e
n
t
liter
atu
r
e
r
ev
ea
ls
a
g
r
a
d
u
al
b
u
t
co
n
s
is
ten
t
p
r
o
g
r
ess
io
n
t
o
war
d
au
to
m
ated
ar
c
h
itectu
r
e
d
esig
n
an
d
o
p
tim
izatio
n
,
with
m
etah
eu
r
is
tic
ap
p
r
o
ac
h
e
s
o
f
f
er
in
g
p
r
o
m
is
in
g
p
ath
way
s
to
ad
d
r
ess
th
e
well
-
d
o
cu
m
e
n
ted
lim
itatio
n
s
o
f
m
an
u
ally
en
g
in
ee
r
ed
DL
m
o
d
els.
3.
M
E
T
H
O
D
3
.
1
.
O
v
er
v
iew
of
t
he
pro
po
s
ed
f
r
a
me
w
o
r
k
T
h
e
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
in
teg
r
ates
C
NNs
w
ith
two
b
io
-
in
s
p
ir
ed
alg
o
r
ith
m
s
th
e
GA
an
d
PS
O
to
au
to
m
ate
th
e
d
esig
n
an
d
f
u
s
io
n
o
f
DL
m
o
d
els
f
o
r
AD
class
if
icatio
n
.
As
illu
s
tr
ated
in
Fig
u
r
e
1
,
th
e
p
ip
elin
e
s
tar
ts
with
an
ess
en
tial
p
r
e
p
r
o
ce
s
s
in
g
s
tep
,
f
o
llo
wed
b
y
t
h
r
ee
s
eq
u
e
n
tial
s
tag
es:
a
GA
s
ea
r
ch
es
f
o
r
C
NN
ar
ch
itectu
r
es
ad
ap
ted
t
o
ea
ch
an
ato
m
ical
o
r
ien
tatio
n
;
th
e
o
p
tim
ized
m
o
d
els
ar
e
r
etr
ai
n
ed
an
d
v
alid
ated
o
n
a
lar
g
er
,
in
d
ep
en
d
en
t
d
ataset
to
ass
es
s
g
en
er
aliza
b
ilit
y
;
an
d
f
in
ally
,
PS
O
d
eter
m
in
es
th
e
f
u
s
io
n
weig
h
ts
th
at
co
m
b
in
e
th
e
p
r
ed
ictio
n
s
f
r
o
m
th
e
th
r
ee
an
ato
m
ical
v
iews.
Fig
u
r
e
1
.
Ov
e
r
all
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
C
NN
-
GA
-
PS
O
f
r
am
ewo
r
k
3
.
2
.
Da
t
a
a
cquis
it
io
n a
nd
pr
epro
ce
s
s
ing
3
.
2
.
1
.
Da
t
a
s
et
s
T
h
e
Kag
g
le
d
ataset
[
2
4
]
p
r
o
v
id
es
6
,
4
0
0
p
r
e
p
r
o
ce
s
s
ed
2
D
MRI
s
lices
(
1
2
8
×
1
2
8
p
x
)
d
is
tr
ib
u
ted
ac
r
o
s
s
f
o
u
r
d
em
e
n
tia
s
ev
er
ity
class
es,
r
e
-
lab
elled
in
to
a
b
in
a
r
y
s
ch
em
e
(
Dem
en
ted
v
s
.
No
n
-
Dem
en
ted
)
f
o
r
th
e
GA
ar
ch
itectu
r
e
s
ea
r
ch
p
h
as
e.
T
h
e
ADNI
d
ataset
[
2
5
]
p
r
o
v
id
es
1
,
7
1
5
f
u
ll
3
D
s
tr
u
ct
u
r
al
MRI
v
o
lu
m
es
(
8
2
3
AD,
8
9
2
NC
)
ac
q
u
ir
ed
at
1
.
5
T
an
d
3
T
f
ield
s
tr
e
n
g
th
s
,
p
ar
titi
o
n
ed
in
t
o
tr
ain
in
g
(
6
0
%),
v
alid
atio
n
(
1
5
%),
a
n
d
h
eld
-
o
u
t
test
(
2
5
%
)
s
u
b
s
ets;
s
u
b
ject
co
u
n
ts
p
e
r
s
p
lit
ar
e
g
iv
e
n
in
T
ab
le
1
.
T
wo
d
atasets
wer
e
u
s
ed
ac
r
o
s
s
th
e
d
if
f
er
e
n
t stag
es o
f
t
h
is
s
tu
d
y
T
ab
le
2
.
T
ab
le
1
.
ADNI
s
u
b
ject
d
is
tr
ib
u
tio
n
ac
r
o
s
s
d
ataset
s
p
lits
S
p
l
i
t
A
D
su
b
j
e
c
t
s
N
C
su
b
j
e
c
t
s
To
t
a
l
Tr
a
i
n
i
n
g
(
6
0
%)
4
9
4
5
3
5
1
,
0
2
9
V
a
l
i
d
a
t
i
o
n
(
1
5
%)
1
2
3
1
3
4
2
5
7
Te
st
(
2
5
%)
2
0
6
2
2
3
4
2
9
To
t
a
l
8
2
3
8
9
2
1
,
7
1
5
T
ab
le
2
.
Su
m
m
a
r
y
o
f
d
atasets
u
s
ed
in
th
is
s
tu
d
y
C
h
a
r
a
c
t
e
r
i
s
t
i
c
A
D
N
I
K
a
g
g
l
e
To
t
a
l
su
b
j
e
c
t
s
/
i
ma
g
e
s
1
,
7
1
5
s
u
b
j
e
c
t
s
6
,
4
0
0
i
ma
g
e
s
AD
8
2
3
3
,
2
0
0
NC
8
9
2
3
,
2
0
0
D
a
t
a
t
y
p
e
3
D
v
o
l
u
m
e
s
2
D
sl
i
c
e
s
U
sag
e
Tr
a
i
n
i
n
g
,
v
a
l
i
d
a
t
i
o
n
,
t
e
s
t
G
A
a
r
c
h
i
t
e
c
t
u
r
e
se
a
r
c
h
F
i
e
l
d
s
t
r
e
n
g
t
h
1
.
5
T
a
n
d
3
T
U
n
k
n
o
w
n
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
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mu
lti
-
view MR
I
A
lz
eh
imer’
s
d
etec
tio
n
u
s
in
g
…
(
I
h
eb
C
h
emss
E
l D
in
e
Ha
g
a
n
i
)
195
3
.
2
.
2
.
P
re
pro
ce
s
s
ing
p
ipeli
ne
Pre
p
r
o
ce
s
s
in
g
MRI
im
ag
es
is
a
cr
itical
s
tep
in
th
e
m
eth
o
d
o
l
o
g
ical
ch
ain
,
as
th
e
p
er
f
o
r
m
a
n
ce
o
f
DL
m
o
d
els
d
ep
e
n
d
s
h
ea
v
ily
o
n
t
h
e
q
u
ality
,
co
n
s
is
ten
cy
,
an
d
s
tan
d
ar
d
izatio
n
o
f
th
e
in
p
u
t
d
ata.
R
aw
MRI
v
o
lu
m
es
ex
h
ib
it
v
ar
io
u
s
s
o
u
r
ce
s
of
v
ar
iab
ilit
y
,
r
elate
d
to
d
if
f
er
en
ce
s
in
ac
q
u
is
itio
n
,
s
ca
n
n
er
s
,
n
o
is
e,
n
o
n
-
h
o
m
o
g
en
eo
u
s
in
ten
s
ities
,
an
d
th
e
p
r
esen
ce
o
f
n
o
n
-
b
r
ain
s
tr
u
ctu
r
es.
T
o
r
ed
u
ce
th
ese
b
iases
an
d
en
s
u
r
e
r
o
b
u
s
t
lear
n
in
g
,
a
s
tr
u
ctu
r
ed
p
r
ep
r
o
c
ess
in
g
p
ip
elin
e
is
ap
p
lied
to
all
v
o
lu
m
es.
I
t
r
elies
o
n
f
o
u
r
m
ain
s
tep
s
:
s
k
u
ll
s
tr
ip
p
in
g
,
in
ten
s
ity
n
o
r
m
aliza
t
io
n
,
n
o
is
e
f
ilter
in
g
,
an
d
in
ter
-
s
u
b
ject
h
ar
m
o
n
izatio
n
.
Sk
u
ll
s
tr
ip
p
in
g
,
also
k
n
o
wn
a
s
b
r
ain
s
tr
ip
p
i
n
g
,
i
n
v
o
lv
es
r
e
m
o
v
in
g
ex
tr
ac
er
e
b
r
al
s
tr
u
ct
u
r
es
s
u
ch
as
b
o
n
es,
s
k
i
n
,
an
d
n
o
n
-
n
eu
r
o
n
al
tis
s
u
es
to
f
o
c
u
s
th
e
a
n
aly
s
is
ex
clu
s
iv
ely
o
n
in
t
r
ac
r
an
ial
r
e
g
io
n
s
.
A
DL
m
o
d
el
ca
lled
d
ee
p
b
r
ain
[
2
6
]
,
s
p
ec
ializin
g
in
b
r
ain
s
eg
m
en
tatio
n
,
was
ch
o
s
en
d
u
e
to
its
s
u
p
er
io
r
ac
cu
r
ac
y
an
d
r
o
b
u
s
tn
ess
to
in
ter
-
s
u
b
ject
v
ar
iatio
n
.
T
h
is
ap
p
r
o
ac
h
y
i
el
d
s
s
h
ar
p
b
r
ain
m
ask
s
,
p
r
ese
r
v
in
g
co
r
tical
an
d
s
u
b
co
r
tical
s
tr
u
ctu
r
es wh
ile
ef
f
ec
tiv
ely
elim
in
atin
g
n
o
n
-
i
n
f
o
r
m
ativ
e
r
eg
io
n
s
.
I
n
t
e
n
s
i
t
y
N
o
r
m
a
l
i
z
at
i
o
n
:
T
h
is
s
t
e
p
p
r
o
m
o
t
es
t
h
e
s
t
a
b
il
i
t
y
o
f
n
e
u
r
a
l
n
e
t
w
o
r
k
t
r
a
i
n
i
n
g
a
n
d
i
m
p
r
o
v
e
s
t
h
e
m
o
d
e
l
’
s
a
b
il
i
t
y
t
o
g
e
n
e
r
a
l
iz
e
t
o
n
e
w
s
u
b
j
e
c
ts
.
I
t
r
e
li
es
o
n
l
i
n
e
a
r
s
t
r
e
t
c
h
i
n
g
o
f
i
n
te
n
s
it
i
es,
t
r
a
n
s
f
o
r
m
i
n
g
e
a
c
h
v
o
x
e
l
a
c
c
o
r
d
i
n
g
t
o
t
h
e
m
i
n
im
u
m
a
n
d
m
a
x
i
m
u
m
v
a
l
u
e
s
o
f
t
h
e
v
o
l
u
m
e
.
T
h
i
s
p
r
o
c
es
s
u
n
i
f
ie
s
i
n
t
e
n
s
i
t
y
d
i
s
t
r
i
b
u
ti
o
n
s
a
n
d
l
i
m
it
s
t
h
e
im
p
a
c
t
o
f
n
o
n
-
b
i
o
l
o
g
i
c
a
l
v
a
r
i
a
t
i
o
n
s
,
w
h
i
le
p
r
e
s
e
r
v
i
n
g
i
n
t
e
r
n
a
l
c
o
n
t
r
a
s
ts
b
e
t
we
e
n
b
r
a
i
n
t
i
s
s
u
e
s
.
=
−
−
(
1
)
No
is
e
f
ilter
in
g
an
d
in
ter
-
s
u
b
je
ct
h
ar
m
o
n
izatio
n
:
th
is
n
o
is
e
r
ed
u
ctio
n
im
p
r
o
v
es
th
e
v
is
u
al
q
u
ality
o
f
th
e
v
o
lu
m
es
an
d
e
n
h
an
ce
s
th
e
r
eliab
ilit
y
o
f
t
h
e
f
ea
tu
r
es
ex
tr
ac
ted
in
s
u
b
s
eq
u
e
n
t
s
tep
s
.
T
h
e
g
o
al
is
to
r
em
o
v
e
r
an
d
o
m
v
ar
iatio
n
s
with
o
u
t
al
ter
in
g
s
u
b
tle
m
o
r
p
h
o
l
o
g
ical
p
atter
n
s
,
s
u
ch
as
co
r
tical
th
i
n
n
in
g
o
r
v
o
l
u
m
etr
ic
v
ar
iatio
n
o
f
th
e
h
i
p
p
o
ca
m
p
u
s
,
wh
ich
ar
e
k
e
y
m
ar
k
er
s
o
f
n
eu
r
o
d
eg
en
e
r
atio
n
.
E
v
en
a
f
ter
n
o
r
m
aliza
tio
n
an
d
f
ilter
in
g
,
r
esid
u
al
d
if
f
er
e
n
ce
s
m
ay
r
e
m
ain
b
etwe
en
s
u
b
je
cts
d
u
e
to
s
ca
n
n
er
s
,
ac
q
u
is
itio
n
ce
n
ter
s
,
o
r
ex
p
er
im
en
tal
p
r
o
to
co
ls
.
I
n
t
er
-
s
u
b
ject
h
ar
m
o
n
izatio
n
is
th
er
ef
o
r
e
a
p
p
lied
to
m
itig
ate
th
ese
d
is
cr
ep
a
n
cies
a
n
d
s
tan
d
a
r
d
ize
t
h
e
r
e
p
r
esen
tati
o
n
o
f
v
o
lu
m
es
with
in
th
e
d
ata
s
et.
T
ab
le
3
r
e
p
o
r
ts
th
e
av
er
ag
e
wall
-
clo
c
k
tim
e
p
er
p
r
ep
r
o
ce
s
s
in
g
s
tep
m
ea
s
u
r
e
d
o
n
a
r
ep
r
esen
tativ
e
ADNI
v
o
lu
m
e
u
s
in
g
Go
o
g
le
c
o
lab
(
T
esla T
4
GPU)
.
T
ab
le
3
.
Av
e
r
ag
e
wall
-
clo
ck
ti
m
e
p
er
p
r
ep
r
o
ce
s
s
in
g
s
tep
(
p
er
MRI
v
o
lu
m
e,
Go
o
g
le
C
o
lab
T
4
)
S
t
e
p
A
v
e
r
a
g
e
t
i
me
R
e
o
r
i
e
n
t
a
t
i
o
n
a
n
d
i
s
o
t
r
o
p
i
c
r
e
sam
p
l
i
n
g
1
4
s
B
r
a
i
n
e
x
t
r
a
c
t
i
o
n
(
D
e
e
p
B
r
a
i
n
)
3
2
s
I
n
t
e
n
s
i
t
y
n
o
r
m
a
l
i
sa
t
i
o
n
3
s
N
o
i
se
f
i
l
t
e
r
i
n
g
9
s
I
n
t
e
r
-
su
b
j
e
c
t
h
a
r
m
o
n
i
s
a
t
i
o
n
7
s
2
D
sl
i
c
e
e
x
t
r
a
c
t
i
o
n
(
p
e
r
p
l
a
n
e
)
4
s
To
t
a
l
p
e
r
s
u
b
j
e
c
t
∼
1
mi
n
1
0
s
3
.3
.
St
a
g
e
1
:
G
enet
ic
a
lg
o
rit
hm
f
o
r
CNN
a
rc
hite
ct
ure
o
p
t
im
iza
t
io
n
3
.
3
.
1
.
Arc
hite
ct
ure
enco
din
g
a
nd
s
ea
rc
h sp
a
ce
E
ac
h
ca
n
d
id
ate
C
NN
is
r
ep
r
es
en
ted
as a
ch
r
o
m
o
s
o
m
e,
with
g
en
es e
n
co
d
i
n
g
ar
c
h
itectu
r
al
p
ar
am
eter
s
:
−
Nu
m
b
er
of
c
o
n
v
o
lu
tio
n
al
-
p
o
o
l
in
g
b
l
o
ck
s
(1
-
5)
.
−
Ker
n
el
s
izes
(
e.
g
.
,
3
×
3,
5
×
5)
an
d
f
ilter
co
u
n
ts
p
er
lay
er
(
e.
g
.
,
32,
6
4
,
128)
.
−
R
eg
u
lar
izatio
n
an
d
o
p
tim
izatio
n
p
ar
am
ete
r
s
:
d
r
o
p
o
u
t
r
ate
(
0
.
2
-
0
.
5
)
,
o
p
tim
izer
ty
p
e
(
Ad
a
m
an
d
SGD)
,
an
d
ac
tiv
atio
n
f
u
n
cti
o
n
s
(
R
eL
U
an
d
L
ea
k
y
R
eL
U)
.
T
h
is
s
ea
r
ch
s
p
ac
e
co
m
b
in
es
d
is
cr
ete
o
r
d
in
al
v
ar
iab
les
(
lay
e
r
co
u
n
ts
an
d
f
ilter
s
izes)
an
d
ca
teg
o
r
ical
v
ar
iab
les
(
ac
tiv
atio
n
f
u
n
cti
o
n
s
,
o
p
tim
izer
ty
p
e)
,
f
o
r
wh
ich
g
r
ad
ien
t
-
b
ased
o
p
tim
izatio
n
ca
n
n
o
t
be
ap
p
lied
d
ir
ec
tly
.
E
x
h
au
s
tiv
e
e
n
u
m
er
atio
n
is
co
m
p
u
tatio
n
ally
in
tr
ac
tab
le
g
iv
en
th
e
co
m
b
i
n
ato
r
ial
ex
p
lo
s
io
n
o
f
co
n
f
ig
u
r
atio
n
s
,
an
d
r
a
n
d
o
m
s
ea
r
ch
p
r
o
v
id
es
n
o
m
ec
h
a
n
is
m
to
d
ir
ec
t
ex
p
lo
r
atio
n
to
war
d
h
ig
h
-
f
itn
ess
r
eg
io
n
s
.
GA
ad
d
r
ess
es
th
ese
co
n
-
s
tr
ain
ts
n
atu
r
ally
:
it
o
p
er
ates
with
o
u
t
g
r
ad
ien
ts
,
en
co
d
es
h
eter
o
g
en
e
o
u
s
v
ar
iab
le
ty
p
es
th
r
o
u
g
h
c
h
r
o
m
o
s
o
m
e
r
ep
r
esen
tatio
n
,
m
ai
n
tain
s
p
o
p
u
latio
n
d
iv
er
s
ity
v
ia
m
u
tatio
n
to
av
o
i
d
p
r
em
atu
r
e
co
n
v
er
g
en
ce
,
an
d
s
teer
s
th
e
s
ea
r
ch
p
r
o
g
r
ess
iv
ely
th
r
o
u
g
h
s
elec
tio
n
p
r
ess
u
r
e
an
d
cr
o
s
s
o
v
e
r
r
ec
o
m
b
in
atio
n
.
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
:
192
-
20
6
196
3
.
3
.
2
.
E
v
o
lutio
na
ry
o
pti
m
iza
t
io
n pro
ce
s
s
T
h
e
GA
b
eg
in
s
by
in
itializin
g
a
p
o
p
u
latio
n
of
r
a
n
d
o
m
ar
c
h
itectu
r
es.
E
ac
h
ca
n
d
id
ate
is
tr
ain
ed
o
n
a
s
u
b
s
et
o
f
th
e
Kag
g
le
d
ataset
,
an
d
its
v
alid
atio
n
ac
c
u
r
ac
y
s
er
v
es
as
th
e
f
itn
ess
m
etr
ic.
S
elec
tio
n
,
cr
o
s
s
o
v
er
,
an
d
m
u
tatio
n
o
p
er
ato
r
s
ar
e
ap
p
lied
iter
ativ
ely
to
ev
o
lv
e
th
e
p
o
p
u
latio
n
o
v
er
m
u
ltip
le
g
en
er
atio
n
s
.
Selectio
n
p
r
io
r
itizes
h
ig
h
-
f
itn
ess
in
d
iv
id
u
als;
cr
o
s
s
o
v
er
r
ec
o
m
b
i
n
es
ar
ch
itectu
r
al
co
m
p
o
n
en
ts
f
r
o
m
p
ar
en
t
m
o
d
els;
an
d
m
u
tatio
n
i
n
tr
o
d
u
ce
s
r
an
d
o
m
v
ar
iatio
n
s
to
m
ain
tain
d
i
v
er
s
ity
.
T
h
e
p
r
o
ce
s
s
ter
m
in
ates
wh
en
f
itn
ess
im
p
r
o
v
em
en
t
s
tag
n
ates
o
r
a
m
ax
im
u
m
g
en
er
atio
n
co
u
n
t
is
r
ea
ch
ed
.
T
h
e
o
p
tim
izatio
n
wo
r
k
f
l
o
w
is
d
ep
icted
in
Fig
u
r
e
2
.
T
ab
le
4
s
u
m
m
ar
is
es
th
e
co
m
p
l
ete
s
et
o
f
h
y
p
e
r
p
ar
am
eter
s
g
o
v
er
n
in
g
th
e
GA
o
p
tim
izatio
n
p
r
o
ce
d
u
r
e.
Fig
u
r
e
2
.
Gen
etic
a
lg
o
r
ith
m
o
p
tim
izatio
n
wo
r
k
f
lo
w
f
o
r
C
NN
ar
ch
itectu
r
e
d
esig
n
T
ab
le
4
.
Hy
p
er
p
ar
a
m
eter
s
o
f
t
h
e
g
en
etic
alg
o
r
ith
m
f
o
r
C
NN
ar
ch
itectu
r
e
s
ea
r
ch
P
a
r
a
me
t
e
r
V
a
l
u
e
P
o
p
u
l
a
t
i
o
n
s
i
z
e
20
M
a
x
i
m
u
m
g
e
n
e
r
a
t
i
o
n
s
50
S
e
l
e
c
t
i
o
n
m
e
c
h
a
n
i
s
m
To
u
r
n
a
me
n
t
(
k
=
3
)
C
r
o
ss
o
v
e
r
o
p
e
r
a
t
o
r
U
n
i
f
o
r
m
c
r
o
sso
v
e
r
C
r
o
ss
o
v
e
r
r
a
t
e
0
.
8
0
M
u
t
a
t
i
o
n
r
a
t
e
0
.
1
0
El
i
t
i
sm
2
i
n
d
i
v
i
d
u
a
l
s
Tr
a
i
n
i
n
g
b
u
d
g
e
t
p
e
r
c
a
n
d
i
d
a
t
e
1
0
e
p
o
c
h
s
Tr
a
i
n
i
n
g
su
b
s
e
t
(
K
a
g
g
l
e
)
6
,
4
0
0
i
ma
g
e
s
(
1
2
8
×
1
2
8
p
x
,
b
i
n
a
r
y
l
a
b
e
l
s)
F
i
t
n
e
ss
met
r
i
c
V
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
S
t
o
p
p
i
n
g
c
r
i
t
e
r
i
o
n
M
a
x
g
e
n
e
r
a
t
i
o
n
s
o
r
1
0
st
a
g
n
a
t
i
n
g
g
e
n
e
r
a
t
i
o
n
s
3
.
3
.
3
.
O
rient
a
t
io
n
-
s
pecif
ic
a
r
chit
ec
t
ure
s
ea
rc
h
A
u
n
iq
u
e
asp
ec
t
o
f
th
is
f
r
am
e
wo
r
k
is
th
e
ex
ec
u
tio
n
o
f
in
d
e
p
en
d
en
t
GA
r
u
n
s
f
o
r
ax
ial,
co
r
o
n
al,
an
d
s
ag
ittal
MRI
p
lan
es.
T
h
is
ac
k
n
o
wled
g
es
th
at
ea
c
h
o
r
ie
n
tatio
n
ca
p
tu
r
es
d
is
tin
ct
n
eu
r
o
an
at
o
m
ical
f
ea
tu
r
es
an
d
m
ay
r
eq
u
ir
e
s
p
ec
ialized
f
e
atu
r
e
ex
tr
ac
tio
n
h
ier
a
r
c
h
ies.
T
h
e
b
est
-
p
e
r
f
o
r
m
in
g
a
r
ch
i
tectu
r
e
f
r
o
m
ea
c
h
o
r
ien
tatio
n
is
r
etain
ed
f
o
r
s
u
b
s
eq
u
en
t stag
es.
3
.
4
.
St
a
g
e
2
:
CNN
t
ra
ini
ng
a
nd
v
a
lid
a
t
io
n
T
h
e
ar
ch
itectu
r
es
o
p
tim
ized
b
y
th
e
GA
wer
e
r
etr
ai
n
ed
u
s
in
g
3
D
MRI
d
ata
f
r
o
m
th
e
ADNI
d
atab
ase
to
ev
alu
ate
th
eir
g
en
e
r
aliza
b
ilit
y
to
r
ea
lis
tic
clin
ical
co
n
d
itio
n
s
.
Fig
u
r
e
3
s
h
o
ws
r
ep
r
esen
tativ
e
co
r
o
n
al,
s
ag
ittal,
an
d
ax
ial
s
lices
ex
tr
ac
ted
f
r
o
m
th
ese
v
o
lu
m
es:
Fig
u
r
e
3
(
a)
d
ep
icts
s
lices
f
r
o
m
a
s
u
b
ject
d
iag
n
o
s
ed
with
AD
,
wh
ile
Fig
u
r
e
3
(
b
)
s
h
o
ws th
e
co
r
r
esp
o
n
d
in
g
s
lices
f
o
r
a
co
g
n
itiv
ely
NC
.
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
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mu
lti
-
view MR
I
A
lz
eh
imer’
s
d
etec
tio
n
u
s
in
g
…
(
I
h
eb
C
h
emss
E
l D
in
e
Ha
g
a
n
i
)
197
E
ac
h
o
r
ien
tatio
n
-
s
p
ec
if
ic
C
NN
was
tr
ain
ed
in
d
ep
e
n
d
en
tly
f
o
r
6
0
ep
o
ch
s
u
s
in
g
t
h
e
Ad
a
m
o
p
tim
izer
with
a
lear
n
in
g
r
ate
o
f
1
×
1
0
−4
,
b
atch
s
ize
o
f
3
2
,
a
n
d
ca
te
g
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
.
T
o
m
itig
ate
o
v
er
f
itti
n
g
,
ea
r
ly
s
to
p
p
i
n
g
(
p
atien
ce
=
1
0
ep
o
ch
s
)
a
n
d
lear
n
in
g
r
ate
r
e
d
u
ctio
n
(
f
ac
t
o
r
=
0
.
5
)
wer
e
a
p
p
li
ed
wh
en
v
alid
atio
n
p
er
f
o
r
m
an
ce
p
latea
u
ed
.
A
s
u
b
ject
-
in
d
ep
e
n
d
en
t
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
s
ch
em
e
en
s
u
r
ed
n
o
d
ata
leak
ag
e
b
etwe
en
tr
ain
in
g
an
d
v
alid
atio
n
s
p
lits
.
T
h
e
f
in
al
m
o
d
el
f
o
r
ea
ch
v
iew
w
as
s
e
lecte
d
b
ased
o
n
th
e
h
ig
h
est
m
ea
n
v
alid
atio
n
ac
cu
r
ac
y
ac
r
o
s
s
f
o
ld
s
.
(
a)
(
b
)
Fig
u
r
e
3
.
R
ep
r
esen
tativ
e
MRI
s
lices f
r
o
m
th
e
ADNI
d
atab
ase,
illu
s
tr
atin
g
co
r
o
n
al,
s
ag
ittal,
an
d
ax
ial
o
r
ien
tatio
n
s
;
(
a
)
f
o
r
b
o
th
AD
p
atien
ts
an
d
(
b
)
h
ea
lth
y
c
o
n
tr
o
l
s
[
2
7
]
3
.
5
.
St
a
g
e
3
m
ulti
-
v
iew
f
us
io
n v
ia
pa
rt
icle
s
wa
rm
o
pti
m
iz
a
t
io
n
3
.
5
.
1
.
F
us
io
n m
o
t
iv
a
t
io
n a
nd
f
o
rm
ula
t
i
o
n
W
h
ile
ea
ch
o
r
ien
tatio
n
-
s
p
ec
if
i
c
C
NN
p
r
o
v
id
es
v
alu
ab
le
d
is
cr
im
in
ato
r
y
in
f
o
r
m
atio
n
,
co
m
b
in
in
g
th
ei
r
p
r
ed
ictio
n
s
ca
n
ca
p
tu
r
e
co
m
p
lem
en
tar
y
asp
ec
ts
o
f
AD
-
r
ela
ted
n
eu
r
o
an
ato
m
ical
ch
an
g
es.
R
ath
er
th
an
u
s
in
g
u
n
if
o
r
m
av
e
r
ag
in
g
,
we
f
o
r
m
u
l
ate
v
iew
f
u
s
io
n
as
an
o
p
tim
iz
atio
n
p
r
o
b
lem
wh
er
e
PS
O
d
et
er
m
in
es
th
e
o
p
tim
al
weig
h
tin
g
s
ch
em
e
t
h
at
m
ax
im
izes
class
if
icat
io
n
ac
cu
r
ac
y
.
PS
O
is
s
elec
ted
f
o
r
th
is
task
f
o
r
two
r
ea
s
o
n
s
.
First,
th
e
f
itn
ess
s
ig
n
al
-
class
if
icati
o
n
ac
cu
r
ac
y
is
a
n
o
n
-
d
if
f
er
en
tiab
le
s
tep
f
u
n
ctio
n
o
f
th
e
f
u
s
io
n
weig
h
ts
at
t
h
e
d
ec
is
io
n
b
o
u
n
d
ar
y
,
wh
ic
h
r
u
le
s
o
u
t
g
r
a
d
ien
t
-
b
ased
o
p
tim
izer
s
s
u
ch
as
Ad
am
o
r
SGD.
Seco
n
d
,
th
e
s
u
m
-
to
-
o
n
e
co
n
s
tr
ain
t
(
∑
=
1
,
≥
0
)
is
h
an
d
led
n
ativ
e
ly
b
y
n
o
r
m
alizin
g
p
ar
ticle
p
o
s
itio
n
s
af
ter
ea
ch
v
elo
city
u
p
d
ate,
with
o
u
t
r
eq
u
ir
in
g
p
en
alty
ter
m
s
o
r
p
r
o
jectio
n
alg
o
r
ith
m
s
.
T
h
ese
p
r
o
p
er
ties
m
ak
e
PS
O
a
p
r
in
cip
led
an
d
lig
h
tweig
h
t c
h
o
ice
f
o
r
th
is
lo
w
-
d
im
en
s
io
n
al,
c
o
n
s
tr
ain
ed
,
n
o
n
-
d
if
f
er
en
tiab
le
o
p
tim
izatio
n
p
r
o
b
lem
.
3
.
5
.
2
.
F
us
io
n m
ec
ha
nis
m
A
s
ep
ar
ate
in
d
e
p
en
d
e
n
t
d
atase
t
o
f
5
0
0
MRI
s
lices
-
d
is
tin
ct
f
r
o
m
all
d
ata
u
s
ed
in
p
r
e
v
io
u
s
s
tag
es
-
was
u
tili
ze
d
to
tr
ai
n
th
e
PS
O
f
u
s
io
n
m
o
d
u
le.
E
ac
h
o
p
tim
ized
C
NN
g
en
er
ated
p
r
o
b
ab
ilit
y
s
co
r
es
f
o
r
th
ese
s
lices,
f
o
r
m
in
g
a
m
atr
ix
∑
500
3
wh
er
e
c
o
lu
m
n
s
co
r
r
esp
o
n
d
to
a
x
ial,
co
r
o
n
al,
an
d
s
ag
ittal
p
r
ed
ictio
n
s
,
a
n
d
r
o
ws
r
ep
r
esen
t sli
ce
s
.
=
[
1
,
1
,
1
,
2
,
2
,
2
,
⋮
500
,
⋮
500
,
⋮
500
,
]
T
h
e
PS
O
alg
o
r
ith
m
s
ee
k
s
an
o
p
tim
al
weig
h
t v
ec
to
r
,
=
[
,
,
]
th
at
s
atis
f
ies:
+
+
=
1
,
≥
0
f
o
r
ea
ch
s
lice
,
th
e
f
in
al
f
u
s
ed
p
r
o
b
a
b
ilit
y
is
co
m
p
u
ted
as:
̂
=
∑
,
3
=
1
with
th
e
f
in
al
p
r
e
d
ictio
n
d
ete
r
m
in
ed
b
y
:
̂
=
{
1
,
̂
≥
0
.
5
0
,
ℎ
.
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
:
192
-
20
6
198
3
.
5
.
3
.
P
SO
o
ptim
iz
a
t
io
n pro
c
edure
E
ac
h
p
ar
ticle
in
th
e
s
war
m
r
e
p
r
esen
ts
a
ca
n
d
id
ate
weig
h
t
v
ec
to
r
.
Par
ticle
f
itn
ess
i
s
d
ef
in
ed
as
th
e
class
if
icatio
n
ac
cu
r
ac
y
o
n
t
h
e
in
d
ep
en
d
en
t f
u
s
io
n
d
ataset:
(
)
=
1
/
1
∑
∦
=
1
(
̂
=
)
a
t e
ac
h
iter
atio
n
,
p
ar
ticles u
p
d
ate
th
eir
v
elo
city
a
n
d
p
o
s
itio
n
ac
co
r
d
in
g
to
:
,
(
+
1
)
=
,
(
)
+
1
1
(
,
−
,
(
)
)
+
2
2
(
−
,
(
)
)
,
(
+
1
)
=
,
(
)
+
,
(
+
1
)
wh
er
e
is
th
e
in
er
tia
weig
h
t,
1
an
d
2
ar
e
ac
ce
ler
atio
n
co
e
f
f
icien
ts
,
1
,
2
~
(
0
,
1
)
an
d
r
ep
r
esen
t
p
er
s
o
n
al
an
d
g
l
o
b
al
b
est
p
o
s
itio
n
s
,
r
esp
ec
tiv
ely
.
W
eig
h
ts
ar
e
n
o
r
m
alize
d
af
te
r
ea
ch
u
p
d
ate
to
s
atis
f
y
th
e
co
n
s
tr
ain
t.
3
.
6
.
I
m
ple
m
ent
a
t
io
n det
a
ils
a
nd
re
pro
du
cib
ility
T
ab
le
5
co
n
s
o
lid
ates
all
h
y
p
e
r
p
ar
am
eter
s
an
d
d
ataset
s
p
lits
r
eq
u
ir
ed
to
r
ep
r
o
d
u
ce
th
e
ex
p
er
im
en
ts
.
GA
-
s
p
ec
if
ic
p
ar
am
eter
s
ar
e
f
u
r
th
er
d
etailed
in
T
ab
le
4
.
T
ab
le
5
.
C
o
m
p
lete
h
y
p
er
p
ar
a
m
eter
an
d
d
ataset
co
n
f
ig
u
r
atio
n
f
o
r
r
ep
r
o
d
u
cib
ilit
y
C
o
m
p
o
n
e
n
t
P
a
r
a
me
t
e
r
V
a
l
u
e
K
a
g
g
l
e
d
a
t
a
se
t
To
t
a
l
i
m
a
g
e
s
6
,
4
0
0
(
1
2
8
×
1
2
8
p
x
)
B
i
n
a
r
y
s
p
l
i
t
D
e
men
t
e
d
:
3
,
2
0
0
/
n
o
n
-
d
e
m
e
n
t
e
d
:
3
,
200
A
D
N
I
d
a
t
a
se
t
A
D
/
N
C
s
u
b
j
e
c
t
s
8
2
3
A
D
/
8
9
2
N
C
(
1
,
7
1
5
t
o
t
a
l
)
Tr
a
i
n
/
V
a
l
/
Te
s
t
6
0
%/
1
5
%
/
2
5
%
(
1
,
0
2
9
/
2
5
7
/
4
2
9
su
b
j
e
c
t
s)
2
D
sl
i
c
e
s
p
e
r
p
l
a
n
e
∼
5
3
4
0
0
(
t
r
a
i
n
i
n
g
)
,
∼
2
2
3
0
0
(
t
e
st
)
C
N
N
t
r
a
i
n
i
n
g
(
S
t
a
g
e
2
)
Ep
o
c
h
s
60
O
p
t
i
mi
z
e
r
A
d
a
m
Le
a
r
n
i
n
g
r
a
t
e
1
×
1
0
−4
B
a
t
c
h
si
z
e
32
Lo
ss f
u
n
c
t
i
o
n
C
a
t
e
g
o
r
i
c
a
l
c
r
o
ss
-
e
n
t
r
o
p
y
Ea
r
l
y
st
o
p
p
i
n
g
P
a
t
i
e
n
c
e
=
1
0
e
p
o
c
h
s
LR
r
e
d
u
c
t
i
o
n
F
a
c
t
o
r
0
.
5
o
n
p
l
a
t
e
a
u
C
r
o
ss
-
v
a
l
i
d
a
t
i
o
n
5
-
f
o
l
d
(
s
u
b
j
e
c
t
-
i
n
d
e
p
e
n
d
e
n
t
)
P
S
O
(
S
t
a
g
e
3
)
S
w
a
r
m si
z
e
30
M
a
x
i
t
e
r
a
t
i
o
n
s
40
I
n
e
r
t
i
a
w
e
i
g
h
t
0
.
7
C
o
g
n
i
t
i
v
e
c
o
e
f
f
.
1
1
.
5
S
o
c
i
a
l
c
o
e
f
f
.
2
1
.
5
F
u
si
o
n
d
a
t
a
se
t
s
i
z
e
5
0
0
M
R
I
sl
i
c
e
s
R
a
n
d
o
m s
e
e
d
s
G
A
/
C
N
N
/
P
S
O
4
2
(
f
i
x
e
d
)
3
.
7
.
Co
m
pu
t
a
t
io
na
l r
eso
urc
es
T
ab
le
6
r
ep
o
r
ts
th
e
c
o
m
p
u
tatio
n
al
r
eso
u
r
ce
s
co
n
s
u
m
ed
at
ea
c
h
s
tag
e
o
f
th
e
f
r
am
ewo
r
k
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
ct
ed
o
n
Go
o
g
le
C
o
lab
o
r
ato
r
y
,
u
s
in
g
an
NVI
DI
A
T
esla
T
4
GP
U
(
1
6
GB
VR
AM
,
7
0
W
T
DP)
with
an
I
n
tel
Xeo
n
C
PU
an
d
1
2
GB
R
AM
all
o
ca
ted
b
y
th
e
p
latf
o
r
m
.
T
h
e
t
o
tal
wall
-
clo
ck
tim
e
was
ap
p
r
o
x
im
ately
5
0
h
o
u
r
s
,
d
is
tr
ib
u
ted
ac
r
o
s
s
m
u
ltip
le
C
o
lab
s
ess
io
n
s
.
E
s
tim
ated
en
er
g
y
c
o
n
s
u
m
p
tio
n
,
d
er
iv
ed
f
r
o
m
th
e
T
4
’
s
r
ated
T
DP,
is
ap
p
r
o
x
im
ately
5
0
h
×
7
0
W
≈
3
.
5
k
W
h
.
T
h
e
co
m
p
lete
en
d
-
to
-
e
n
d
wo
r
k
f
lo
w,
f
r
o
m
MRI
in
p
u
t to
f
in
al
d
iag
n
o
s
is
an
d
in
ter
p
r
eta
b
ilit
y
v
is
u
aliza
tio
n
,
is
s
u
m
m
ar
ized
in
Fig
u
r
e
4.
T
ab
le
6
.
C
o
m
p
u
tatio
n
al
co
s
t o
f
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
S
t
a
g
e
C
a
n
d
i
d
a
t
e
s/
i
t
e
r
a
t
i
o
n
s
G
P
U
-
h
o
u
r
s
GA
—
A
x
i
a
l
2
0
×
3
5
=
7
0
0
c
a
n
d
i
d
a
t
e
s
~
9
GA
—
C
o
r
o
n
a
l
2
0
×
3
8
=
7
6
0
c
a
n
d
i
d
a
t
e
s
~
1
0
GA
—
S
a
g
i
t
t
a
l
2
0
×
4
1
=
8
2
0
c
a
n
d
i
d
a
t
e
s
~
1
0
S
t
a
g
e
2
(
A
D
N
I
,
5
-
f
o
l
d
)
5
×
3
=
1
5
t
r
a
i
n
i
n
g
r
u
n
s
(
∼
5
3
0
0
0
c
o
u
p
e
s
/
f
o
l
d
)
~
2
1
P
S
O
f
u
si
o
n
3
0
p
a
r
t
i
c
l
e
s
×
2
5
i
t
e
r
<
0
.
1
To
t
a
l
~
5
0
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
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mu
lti
-
view MR
I
A
lz
eh
imer’
s
d
etec
tio
n
u
s
in
g
…
(
I
h
eb
C
h
emss
E
l D
in
e
Ha
g
a
n
i
)
199
Fig
u
r
e
4
.
E
n
d
-
to
-
en
d
w
o
r
k
f
l
o
w
o
f
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
f
r
am
ewo
r
k
f
o
r
AD
class
if
icatio
n
4.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
a
co
m
p
r
eh
en
s
iv
e
ev
alu
atio
n
o
f
o
u
r
C
NN
-
GA
-
PS
O
h
y
b
r
id
f
r
am
ewo
r
k
f
o
r
AD
class
if
icatio
n
.
W
e
s
y
s
tem
atic
ally
ev
alu
ate
ea
ch
m
eth
o
d
o
l
o
g
ical
s
tep
:
C
NN
ar
ch
itectu
r
e
o
p
tim
izatio
n
u
s
in
g
GA,
m
o
d
el
tr
ain
in
g
o
n
th
e
e
n
tire
ADNI
d
ataset,
a
n
d
m
u
lti
-
v
iew
f
u
s
io
n
u
s
in
g
PS
O.
T
h
is
ev
alu
atio
n
r
elieso
n
q
u
an
titativ
e
m
etr
ics
an
d
v
is
u
al
an
aly
s
is
.
All
ex
p
er
im
en
ts
ad
d
r
ess
a
b
in
ar
y
class
if
icatio
n
ta
s
k
.
Du
r
in
g
th
e
GA
ar
ch
itectu
r
e
s
ea
r
ch
p
h
ase,
th
e
f
o
u
r
-
class
Kag
g
le
d
ataset
(
m
ild
d
em
en
ted
:
8
9
6
,
m
o
d
er
ate
d
em
en
ted
:
6
4
,
v
er
y
m
ild
d
em
e
n
ted
:
2
,
2
4
0
,
n
o
n
d
e
m
en
ted
:
3
,
2
0
0
;
to
tal
6
,
4
0
0
im
ag
es
r
esized
to
1
2
8
×
1
2
8
p
x
)
is
r
e
-
lab
elled
in
to
two
b
alan
ce
d
ca
teg
o
r
ies:
d
e
m
en
ted
(
m
ild
+
m
o
d
er
ate
+
v
er
y
m
ild
;
n
=
3
,
2
0
0
)
v
er
s
u
s
n
o
n
d
em
en
te
d
(
n
=
3
,
2
0
0
)
.
Fin
al
r
etr
ain
in
g
a
n
d
ev
alu
atio
n
u
s
e
th
e
ADNI
d
ataset
u
n
d
er
th
e
s
am
e
b
in
ar
y
s
ch
em
e:
AD
v
er
s
u
s
co
g
n
itiv
ely
NC
.
Per
f
o
r
m
a
n
ce
is
r
ep
o
r
ted
as
b
o
th
ag
g
r
eg
a
te
an
d
p
er
-
class
m
etr
ics
to
c
h
ar
ac
ter
ize
m
o
d
el
b
eh
av
io
r
o
n
ea
c
h
d
iag
n
o
s
tic
c
ateg
o
r
y
.
4
.
1
.
G
enet
ic
a
lg
o
rit
h
m
f
o
r
CNN
a
rc
hite
ct
ure
o
ptim
iza
t
io
n
4
.
1
.
1
.
Co
nv
er
g
ence
b
eha
v
io
r
T
h
e
GA
d
em
o
n
s
tr
ated
co
n
s
is
ten
t
co
n
v
er
g
e
n
ce
f
o
r
all
th
r
e
e
an
ato
m
ical
o
r
ien
tatio
n
s
,
al
b
eit
alo
n
g
d
if
f
er
en
t
tr
ajec
to
r
ies.
Star
tin
g
with
an
ac
cu
r
ac
y
o
f
ap
p
r
o
x
im
ately
4
9
%,
all
o
r
ien
tati
o
n
s
s
h
o
wed
s
tead
y
im
p
r
o
v
em
e
n
t o
v
e
r
g
e
n
er
atio
n
s
.
T
h
e
ax
ial
b
r
an
ch
r
ea
ch
e
d
c
o
n
v
er
g
e
n
ce
m
o
s
t
r
ap
id
ly
,
s
tab
ilizin
g
b
y
g
en
er
atio
n
3
5
wit
h
a
f
i
n
al
v
alid
atio
n
ac
c
u
r
ac
y
o
f
7
5
.
3
%
.
T
h
e
co
r
o
n
al
o
r
ien
tatio
n
r
e
q
u
ir
ed
s
lig
h
tly
m
o
r
e
g
en
er
atio
n
s
,
co
n
v
er
g
in
g
b
y
g
en
er
atio
n
3
8
with
an
ac
cu
r
a
cy
o
f
7
3
.
6
%.
T
h
e
s
ag
ittal
b
r
a
n
ch
e
x
h
ib
ited
th
e
s
lo
west
co
n
v
er
g
en
c
e,
r
ea
c
h
in
g
s
tab
ilit
y
b
y
g
en
er
atio
n
4
1
with
an
ac
cu
r
ac
y
o
f
7
2
.
4
%.
T
h
is
p
r
o
g
r
ess
iv
e
r
ef
in
em
e
n
t
co
n
f
ir
m
s
th
e
GA
ab
il
ity
to
s
y
s
tem
atica
lly
ev
o
lv
e
th
e
ar
ch
itectu
r
e.
4
.
1
.
2
.
O
ptim
ized
a
rc
hite
ct
ur
es
T
h
e
GA
-
o
p
tim
ized
ar
ch
itect
u
r
es
ex
h
ib
ited
d
is
tin
ct
ch
ar
ac
ter
is
ti
cs
tailo
r
ed
to
th
eir
r
esp
ec
tiv
e
an
ato
m
ical
p
lan
es,
as
s
u
m
m
a
r
ized
in
T
a
b
le
7
.
T
h
e
ax
ial
ar
ch
itectu
r
e,
with
f
o
u
r
co
n
v
o
lu
tio
n
al
lay
er
s
a
n
d
m
o
d
er
ate
r
eg
u
la
r
izatio
n
,
a
ch
i
ev
ed
th
e
h
ig
h
est
p
er
f
o
r
m
a
n
ce
.
T
h
e
co
r
o
n
al
m
o
d
el
ad
o
p
ted
a
d
ee
p
e
r
f
i
v
e
-
lay
er
s
tr
u
ctu
r
e
with
s
tr
o
n
g
er
r
e
g
u
l
ar
izatio
n
,
wh
ile
th
e
s
ag
ittal
n
etwo
r
k
em
p
lo
y
ed
a
m
o
r
e
c
o
m
p
ac
t
th
r
ee
-
lay
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
:
192
-
20
6
200
d
esig
n
with
m
in
i
-
m
al
d
r
o
p
o
u
t.
T
h
ese
ar
ch
itectu
r
al
d
if
f
er
en
ce
s
r
ef
lect
th
e
v
ar
y
in
g
co
m
p
lex
ity
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
r
e
q
u
ir
ed
f
o
r
d
if
f
er
e
n
t a
n
ato
m
ical
v
iews.
T
ab
le
7
.
S
tr
u
ctu
r
al
p
ar
am
ete
r
s
an
d
p
e
r
f
o
r
m
an
ce
o
f
GA
-
o
p
ti
m
ized
C
NN
ar
ch
itectu
r
es
O
r
i
e
n
t
a
t
i
o
n
F
i
l
t
e
r
p
r
o
g
r
e
ssi
o
n
D
r
o
p
o
u
t
r
a
t
e
A
c
c
u
r
a
c
y
(
%)
A
x
i
a
l
[
3
2
,
6
4
,
1
2
8
,
2
5
6
]
0
.
2
8
7
5
.
3
C
o
r
o
n
a
l
[
3
2
,
6
4
,
1
2
8
,
2
5
6
,
5
1
2
]
0
.
4
0
7
3
.
6
S
a
g
i
t
t
a
l
[
1
6
,
3
2
,
6
4
]
0
.
2
0
7
2
.
4
4
.
1
.
3
.
P
er
f
o
r
m
a
nce
lim
it
a
t
io
ns
in o
ptim
iza
t
io
n pha
s
e
T
h
e
m
o
d
e
r
ate
ac
cu
r
ac
y
lev
els
ac
h
iev
ed
d
u
r
in
g
th
e
GA
o
p
ti
m
izatio
n
p
h
ase
(
7
2
.
4
–
7
5
.
3
%)
p
r
im
ar
ily
r
ef
lect
th
e
co
m
p
u
tatio
n
al
co
n
s
tr
ain
ts
o
f
ev
o
lu
tio
n
ar
y
s
ea
r
c
h
.
T
h
e
u
s
e
o
f
a
r
ed
u
ce
d
d
atas
et
was
n
ec
es
s
ar
y
to
m
ain
tain
f
ea
s
ib
le
o
p
tim
izatio
n
tim
es,
as
ea
ch
g
en
er
atio
n
r
e
q
u
ir
ed
tr
ain
in
g
an
d
e
v
alu
atin
g
m
u
ltip
le
ca
n
d
id
ate
ar
ch
itectu
r
es.
C
o
n
s
eq
u
en
tly
,
t
h
ese
ac
cu
r
ac
ies
r
ep
r
esen
t
ar
ch
itectu
r
al
p
o
ten
tial
r
ath
er
th
a
n
f
in
al
p
er
f
o
r
m
an
ce
,
with
th
e
tr
u
e
ca
p
a
b
ilit
y
o
f
th
e
o
p
tim
ized
n
etwo
r
k
s
b
ec
o
m
i
n
g
ap
p
a
r
en
t
o
n
ly
af
ter
r
etr
ain
in
g
o
n
th
e
co
m
p
lete
ADNI
d
ataset.
4
.
2
.
P
er
f
o
r
m
a
nce
ev
a
lua
t
io
n o
n AD
NI da
t
a
s
et
R
etr
ain
in
g
th
e
g
en
etica
lly
o
p
tim
ized
ar
ch
itectu
r
es
o
n
th
e
ADNI
d
ataset
r
esu
lted
in
s
u
b
s
tan
tial
p
er
f
o
r
-
m
an
ce
im
p
r
o
v
em
e
n
ts
a
cr
o
s
s
all
o
r
ien
tatio
n
s
.
T
h
e
ax
i
al
n
etwo
r
k
m
ain
tain
e
d
its
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
,
ac
h
iev
in
g
8
5
.
0
%
v
alid
atio
n
a
cc
u
r
ac
y
a
n
d
8
4
.
6
%
test
ac
cu
r
ac
y
.
T
h
e
co
r
o
n
al
a
n
d
s
ag
itt
al
m
o
d
els
s
h
o
wed
co
m
p
ar
ab
le
im
p
r
o
v
em
e
n
t
tr
e
n
d
s
,
with
test
ac
cu
r
ac
ies
o
f
8
3
.
1
%
an
d
8
2
.
0
%
r
esp
ec
tiv
e
ly
.
T
h
e
co
n
s
is
ten
t
p
er
f
o
r
m
an
ce
g
a
p
b
etwe
en
v
a
lid
atio
n
an
d
test
m
etr
ics
(
ap
p
r
o
x
im
ately
0
.
4
-
0
.
6
p
er
ce
n
ta
g
e
p
o
in
ts
)
in
d
icate
s
ef
f
ec
tiv
e
g
e
n
er
aliza
tio
n
with
m
in
im
al
o
v
e
r
f
itti
n
g
.
T
ab
le
8
r
ep
o
r
ts
th
e
v
alid
atio
n
ac
cu
r
ac
y
,
test
ac
cu
r
ac
y
,
an
d
v
alid
atio
n
lo
s
s
o
b
tain
ed
f
o
r
ea
ch
o
r
ien
tatio
n
-
s
p
ec
if
ic
m
o
d
el
af
ter
f
u
ll r
etr
ain
i
n
g
o
n
th
e
AD
NI
d
ataset.
T
ab
le
8
.
Per
f
o
r
m
an
ce
m
etr
ics af
ter
r
etr
ain
in
g
o
n
ADNI
d
ataset
O
r
i
e
n
t
a
t
i
o
n
V
a
l
.
A
c
c
.
(
%)
Te
st
A
c
c
.
(
%)
V
a
l
.
Lo
ss
A
x
i
a
l
8
5
.
0
8
4
.
6
0
.
3
6
C
o
r
o
n
a
l
8
3
.
8
8
3
.
1
0
.
4
1
S
a
g
i
t
t
a
l
8
2
.
6
8
2
.
0
0
.
4
4
4
.
3
.
M
ulti
-
v
iew
f
us
io
n v
ia
pa
rt
icle
s
wa
rm
o
pti
m
iza
t
io
n
4
.
3
.
1
.
Co
nv
er
g
ence
a
nd
weig
ht
dis
t
ributio
n
PS
O
d
em
o
n
s
tr
ated
ef
f
icien
t
co
n
v
er
g
en
ce
,
g
en
er
ally
r
e
ac
h
in
g
s
tab
ilit
y
in
2
5
iter
a
tio
n
s
with
3
0
p
ar
ticles
an
d
a
m
a
x
im
u
m
o
f
4
0
iter
atio
n
s
.
T
h
e
co
n
v
er
g
en
ce
p
r
o
f
ile
in
Fig
u
r
e
5
e
x
h
i
b
its
two
s
u
cc
ess
iv
e
p
h
ases
.
Du
r
in
g
th
e
in
itial
iter
atio
n
s
,
p
ar
ticles
s
wee
p
a
b
r
o
a
d
r
eg
io
n
o
f
th
e
f
ea
s
ib
le
weig
h
t
s
p
ac
e
an
d
f
itn
ess
im
p
r
o
v
es st
ee
p
ly
as p
r
o
g
r
ess
iv
ely
b
etter
f
u
s
io
n
co
n
f
ig
u
r
atio
n
s
ar
e
id
en
tifie
d
.
Fig
u
r
e
5
.
C
o
n
v
er
g
e
n
ce
b
eh
a
v
i
o
r
o
f
PS
O
o
p
tim
izatio
n
f
o
r
m
u
lti
-
v
iew
f
u
s
io
n
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
A
h
yb
r
id
a
p
p
r
o
a
ch
fo
r
mu
lti
-
view MR
I
A
lz
eh
imer’
s
d
etec
tio
n
u
s
in
g
…
(
I
h
eb
C
h
emss
E
l D
in
e
Ha
g
a
n
i
)
201
T
h
is
ex
p
lo
r
ato
r
y
p
h
ase
tr
a
n
s
itio
n
s
in
to
an
ex
p
lo
itatio
n
p
h
ase
ar
o
u
n
d
iter
atio
n
1
0
,
wh
er
e
im
p
r
o
v
em
e
n
ts
b
ec
o
m
e
in
c
r
ea
s
in
g
ly
in
cr
em
en
tal
as
th
e
s
war
m
co
n
tr
ac
ts
to
war
d
t
h
e
g
lo
b
al
b
est.
A
s
tab
le
p
latea
u
is
r
ea
ch
e
d
at
a
p
p
r
o
x
im
ately
iter
atio
n
2
5
-
6
2
.
5
% o
f
th
e
m
ax
im
u
m
b
u
d
g
et
-
in
d
icatin
g
th
at
th
e
s
war
m
o
f
3
0
p
a
r
ticles
lo
ca
tes
a
h
ig
h
-
q
u
ality
s
o
lu
tio
n
well
b
ef
o
r
e
th
e
iter
atio
n
lim
it
is
r
ea
ch
ed
.
T
h
e
s
m
o
o
th
,
m
o
n
o
to
n
ic
p
r
o
g
r
ess
io
n
an
d
th
e
ab
s
en
ce
o
f
o
s
cillatio
n
s
in
th
e
later
ite
r
atio
n
s
co
n
f
ir
m
th
at
th
e
PS
O
co
n
tr
o
l
p
ar
am
eter
s
m
ain
tain
a
p
r
o
d
u
ctiv
e
b
ala
n
ce
b
etwe
en
ex
p
lo
r
atio
n
an
d
ex
p
lo
itatio
n
th
r
o
u
g
h
o
u
t
th
e
r
u
n
.
T
h
is
r
ap
id
co
n
v
er
g
en
ce
is
co
n
s
is
ten
t
wit
h
th
e
lo
w
ef
f
ec
tiv
e
d
im
en
s
io
n
ality
o
f
th
e
p
r
o
b
lem
:
th
e
s
u
m
-
to
-
o
n
e
co
n
s
tr
ain
t
o
n
th
e
th
r
ee
f
u
s
io
n
weig
h
ts
r
ed
u
ce
s
th
e
s
ea
r
ch
to
a
two
-
d
im
en
s
io
n
al
s
im
p
lex
,
wh
ich
is
well
with
in
th
e
ca
p
ac
ity
o
f
a
m
o
d
er
ate
-
s
ized
s
war
m
.
T
h
e
weig
h
t
d
is
tr
ib
u
tio
n
in
th
e
last
iter
atio
n
ass
ig
n
s
th
e
g
r
ea
test
weig
h
t
to
th
e
ax
ial
v
iew
(
0
.
4
6
)
,
f
o
llo
wed
b
y
th
e
co
r
o
n
al
(
0
.
3
3
)
an
d
s
ag
ittal
(
0
.
2
1
)
v
iews,
wh
ich
co
r
r
esp
o
n
d
s
to
th
e
in
d
iv
id
u
al
p
er
f
o
r
m
an
ce
h
i
er
ar
ch
y
o
f
C
NNs
s
p
ec
if
ic
to
ea
ch
o
r
ien
tatio
n
.
T
h
is
asy
m
m
etr
ic
weig
h
tin
g
s
ch
em
e
d
em
o
n
s
tr
ates
th
e
alg
o
r
ith
m
’
s
ab
ilit
y
to
q
u
a
n
titativ
ely
ass
ess
th
e
r
elativ
e
d
is
cr
im
in
ativ
e
p
o
wer
o
f
ea
c
h
an
ato
m
ical
v
iew,
r
ath
e
r
th
an
r
ely
in
g
o
n
u
n
if
o
r
m
a
v
er
ag
in
g
.
4
.
3
.
2
.
F
us
io
n
p
er
f
o
rm
a
nce
T
h
e
PS
O
-
b
ased
f
u
s
io
n
m
ec
h
an
is
m
y
ield
ed
s
ig
n
if
ican
t
p
er
f
o
r
m
an
ce
g
ain
s
.
T
h
e
f
u
s
ed
m
o
d
el
ac
h
iev
ed
8
7
.
4
%
ac
cu
r
ac
y
,
r
e
p
r
esen
tin
g
a
2
.
8
p
er
ce
n
tag
e
p
o
in
t
i
m
p
r
o
v
e
m
en
t
o
v
er
th
e
b
est
in
d
iv
id
u
al
m
o
d
el
(
ax
ial
C
NN)
.
Mo
r
e
im
p
o
r
tan
tly
,
th
e
b
alan
ce
d
p
er
f
o
r
m
a
n
c
e
ac
r
o
s
s
p
r
ec
is
io
n
(
8
6
.
8
%),
r
ec
all
(
8
7
.
0
%),
a
n
d
F1
-
s
co
r
e
(
8
6
.
9
%)
i
n
d
icate
s
th
at
th
e
f
u
s
io
n
m
ec
h
an
is
m
ef
f
ec
tiv
ely
lev
er
a
g
es
co
m
p
le
m
en
tar
y
in
f
o
r
m
atio
n
ac
r
o
s
s
v
iews with
o
u
t in
tr
o
d
u
ci
n
g
s
ig
n
if
ican
t
b
ias.
T
ab
le
9
p
r
o
v
id
es
a
d
etailed
co
m
p
a
r
is
o
n
o
f
class
if
icat
io
n
m
etr
ics
b
etwe
en
ea
ch
in
d
iv
id
u
al
o
r
ien
tatio
n
-
s
p
ec
if
ic
C
NN
an
d
th
e
PS
O
-
f
u
s
ed
m
o
d
el.
T
o
f
u
r
t
h
er
a
n
aly
ze
t
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
f
u
s
ed
m
o
d
el
ac
r
o
s
s
in
d
iv
id
u
al
d
iag
n
o
s
tic
ca
teg
o
r
ies,
T
ab
le
1
0
r
ep
o
r
ts
th
e
p
er
-
class
p
r
ec
is
io
n
,
r
ec
all,
a
n
d
F1
-
s
co
r
e
o
b
tain
e
d
on
th
e
ADNI
test
s
et.
T
h
e
r
es
u
lts
d
em
o
n
s
tr
ate
b
alan
ce
d
p
er
f
o
r
m
an
ce
f
o
r
b
o
th
NC
an
d
A
D
class
es,
in
d
icatin
g
th
at
th
e
p
r
o
p
o
s
ed
f
u
s
io
n
s
tr
ateg
y
im
p
r
o
v
es
o
v
er
all
cla
s
s
if
icatio
n
p
er
f
o
r
m
a
n
ce
with
o
u
t
in
tr
o
d
u
cin
g
a
n
o
ticea
b
le
b
ias to
war
d
eith
e
r
c
lass
.
T
ab
le
9
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
b
etwe
en
in
d
i
v
id
u
al
C
NNs a
n
d
PS
O
-
f
u
s
ed
m
o
d
el
M
o
d
e
l
A
c
c
.
(
%)
P
r
e
c
.
(
%)
R
e
c
.
(
%)
F
1
(
%)
A
U
C
(
%)
A
x
i
a
l
C
N
N
8
4
.
6
8
4
.
0
8
4
.
2
8
4
.
1
8
7
.
3
C
o
r
o
n
a
l
C
N
N
8
3
.
1
8
2
.
3
8
1
.
9
8
2
.
1
8
6
.
2
S
a
g
i
t
t
a
l
C
N
N
8
2
.
0
8
1
.
3
8
0
.
9
8
1
.
1
8
4
.
8
P
S
O
f
u
si
o
n
8
7
.
4
8
6
.
8
8
7
.
0
8
6
.
9
9
0
.
1
T
ab
le
1
0
.
Per
-
class
class
if
icat
i
o
n
m
etr
ics o
f
t
h
e
PS
O
-
f
u
s
ed
m
o
d
el
o
n
th
e
ADNI
test
s
et
C
l
a
s
s
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
sc
o
r
e
(
%)
N
C
(
N
o
r
ma
l
c
o
n
t
r
o
l
)
8
8
.
4
8
5
.
2
8
6
.
8
A
D
(
A
l
z
h
e
i
m
e
r
’
s)
8
5
.
2
8
8
.
8
8
7
.
0
M
a
c
r
o
a
v
g
.
8
6
.
8
8
7
.
0
8
6
.
9
4
.
4
.
Vis
ua
l In
t
er
pret
a
t
io
n v
i
a
G
ra
dient
-
weig
hte
d CA
M
Ap
p
ly
in
g
g
r
ad
ien
t
-
weig
h
ted
c
lass
ac
tiv
atio
n
m
ap
p
in
g
(
Gr
a
d
-
C
AM
)
p
r
o
v
id
e
d
a
b
etter
u
n
d
er
s
tan
d
in
g
o
f
th
e
m
o
d
el
’
s
d
ec
is
io
n
-
m
a
k
in
g
p
r
o
ce
s
s
.
Gr
ad
-
C
AM
was
ap
p
lied
to
th
e
last
co
n
v
o
lu
ti
o
n
al
b
lo
ck
o
f
ea
c
h
GA
-
o
p
tim
ize
d
C
NN
-
th
e
f
o
u
r
th
,
f
if
t
h
,
an
d
th
ir
d
b
lo
ck
f
o
r
th
e
ax
ial,
co
r
o
n
al,
a
n
d
s
ag
ittal
n
etwo
r
k
s
,
r
esp
ec
tiv
ely
T
ab
le
7
im
m
ed
i
ately
p
r
io
r
to
th
e
g
lo
b
al
av
e
r
ag
e
p
o
o
lin
g
o
p
er
atio
n
.
T
h
is
ch
o
ice
is
s
tan
d
ar
d
p
r
ac
tice:
th
e
last
co
n
v
o
lu
tio
n
al
lay
er
r
etain
s
th
e
r
ich
est
s
em
an
tic
co
n
ten
t
wh
ile
p
r
eser
v
i
n
g
s
u
f
f
icien
t
s
p
atial
r
eso
lu
tio
n
f
o
r
m
ea
n
i
n
g
f
u
l
lo
ca
lizatio
n
-
ea
r
lier
lay
e
r
s
ca
p
tu
r
e
lo
w
-
lev
el
te
x
tu
r
es
b
u
t
lack
d
is
cr
im
in
ativ
e
p
o
wer
,
wh
ile
p
o
s
t
-
p
o
o
lin
g
la
y
er
s
d
is
ca
r
d
s
p
atial
in
f
o
r
m
atio
n
en
tire
ly
.
Fr
o
m
th
e
1
2
8
×
1
2
8
in
p
u
t,
th
es
e
b
lo
ck
s
p
r
o
d
u
ce
ac
tiv
atio
n
m
ap
s
o
f
8
×
8
,
4
×
4
,
an
d
1
6
×
1
6
f
o
r
th
e
ax
ial,
co
r
o
n
al,
an
d
s
ag
ittal
m
o
d
els,
wh
ich
we
r
e
b
ilin
ea
r
ly
u
p
s
am
p
le
d
to
th
e
i
n
p
u
t
r
eso
lu
tio
n
an
d
o
v
er
laid
o
n
th
e
s
lice
f
o
r
v
is
u
al
in
ter
p
r
etatio
n
.
Fig
u
r
e
6
illu
s
tr
ates
r
ep
r
esen
tativ
e
v
is
u
aliza
tio
n
s
f
o
r
b
o
th
co
r
r
ec
t
a
n
d
in
co
r
r
ec
t
class
if
icatio
n
s
.
I
n
f
alse
n
eg
ativ
e
ca
s
es
(
Fig
u
r
e
6
(
a)
)
,
th
e
m
o
d
el
ex
h
i
b
ited
d
if
f
u
s
e
an
d
p
o
o
r
ly
lo
ca
l
ized
ac
tiv
atio
n
p
atter
n
s
,
s
u
g
g
esti
n
g
th
e
m
o
d
el
’
s
u
n
ce
r
tain
ty
wh
e
n
d
is
cr
im
in
a
tin
g
f
ea
tu
r
es
wer
e
less
p
r
o
n
o
u
n
ce
d
.
I
n
c
o
n
tr
ast,
in
tr
u
e
p
o
s
itiv
e
ca
s
es
(
Fig
u
r
e
6
(
b
)
)
,
th
e
m
o
d
el
co
n
s
is
ten
tly
f
o
cu
s
ed
o
n
r
eg
io
n
s
o
f
m
ed
ical
in
ter
est,
in
clu
d
in
g
th
e
m
ed
ial
tem
p
o
r
al
lo
b
e
an
d
h
ip
p
o
ca
m
p
u
s
,
r
eg
i
o
n
s
clin
ically
ass
o
ciate
d
with
AD
p
ath
o
l
o
g
y
.
T
h
ese
v
is
u
aliza
tio
n
s
n
o
t
o
n
l
y
en
h
an
ce
in
te
r
p
r
etab
ilit
y
b
u
t a
l
s
o
p
r
o
v
id
e
clin
ical
v
alid
atio
n
b
y
d
em
o
n
s
tr
atin
g
th
at
th
e
m
o
d
el
’
s
atten
tio
n
alig
n
s
with
k
n
o
wn
n
eu
r
o
p
ath
o
l
o
g
ical
m
ar
k
er
s
o
f
AD.
Evaluation Warning : The document was created with Spire.PDF for Python.