I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
,
p
p
.
1
3
5
2
~
1
3
6
3
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v15
i
3
.
pp
1
3
5
2
-
1
3
6
3
1352
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Deep r
einforcem
e
nt
lea
rning
i
nspir
ed optimiza
tion
f
r
a
mewo
rk
using
O
p
tuna for
bra
in t
umo
r
dete
ction
Aa
s
hu
t
o
s
h K
ha
rb,
P
ra
chi C
ha
ud
ha
ry
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
mm
u
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
D
C
R
U
S
T,
M
u
r
t
h
a
l
,
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
Sep
1
0
,
2
0
2
5
R
ev
is
ed
Ma
y
2
4
,
2
0
2
6
Acc
ep
ted
J
u
l 5
,
2
0
2
6
Ac
c
u
ra
te
b
ra
in
tu
m
o
r
d
e
tec
ti
o
n
i
s
e
ss
e
n
ti
a
l
fo
r
e
ffe
c
ti
v
e
c
li
n
ica
l
d
iag
n
o
sis;
h
o
we
v
e
r,
t
h
e
p
e
rfo
rm
a
n
c
e
o
f
d
e
e
p
lea
rn
in
g
m
o
d
e
ls
is
h
ig
h
ly
se
n
siti
v
e
t
o
m
a
n
u
a
ll
y
se
lec
ted
a
rc
h
it
e
c
tu
re
s
a
n
d
h
y
p
e
rp
a
ra
m
e
ters
.
To
a
d
d
re
ss
th
is
c
h
a
ll
e
n
g
e
,
th
is
p
a
p
e
r
p
re
se
n
ts
a
re
in
fo
rc
e
m
e
n
t
lea
rn
in
g
–
i
n
sp
ired
a
u
to
m
a
ted
o
p
ti
m
iza
ti
o
n
fra
m
e
wo
rk
f
o
r
b
ra
in
tu
m
o
r
d
e
tec
ti
o
n
t
h
a
t
e
li
m
in
a
te
s
m
a
n
u
a
l
tri
a
l
-
a
n
d
-
e
rro
r
tu
n
in
g
o
f
h
y
p
e
rp
a
r
a
m
e
ters
.
Th
e
p
ro
p
o
se
d
a
p
p
ro
a
c
h
in
teg
ra
tes
Eff
icie
n
tNe
tB0
a
s
a
fix
e
d
fe
a
tu
re
e
x
trac
to
r
(b
a
se
m
o
d
e
l)
with
a
n
Op
tu
n
a
-
b
a
se
d
re
in
fo
rc
e
m
e
n
t
lea
rn
in
g
st
ra
teg
y
to
jo
in
tl
y
o
p
ti
m
ize
th
e
c
las
sifier
a
rc
h
it
e
c
tu
re
a
n
d
k
e
y
train
i
n
g
h
y
p
e
rp
a
ra
m
e
ters
,
in
c
lu
d
i
n
g
lea
rn
i
n
g
ra
te,
b
a
tc
h
siz
e
,
d
ro
p
o
u
t
ra
te,
a
n
d
n
e
two
r
k
d
e
p
th
.
U
n
li
k
e
e
x
isti
n
g
st
u
d
ies
th
a
t
re
ly
o
n
sta
ti
c
o
r
h
e
u
risti
c
a
ll
y
tu
n
e
d
m
o
d
e
ls,
th
e
p
ro
p
o
se
d
f
ra
m
e
wo
rk
d
y
n
a
m
ica
ll
y
a
d
a
p
ts
m
o
d
e
l
c
o
n
fi
g
u
ra
ti
o
n
s
b
a
se
d
o
n
v
a
li
d
a
ti
o
n
fe
e
d
b
a
c
k
.
Ex
p
e
rime
n
ts
c
o
n
d
u
c
ted
o
n
t
h
e
Bra
TS
2
0
2
0
M
RI
d
a
tas
e
t
d
e
m
o
n
stra
te
t
h
a
t
t
h
e
o
p
t
imiz
e
d
m
o
d
e
l
a
c
h
iev
e
s
a
n
a
c
c
u
ra
c
y
o
f
9
2
%
,
a
n
F
1
-
sc
o
re
o
f
9
2
%
,
a
n
d
a
R
OC
–
AU
C
o
f
0
.
9
6
.
Ad
d
it
i
o
n
a
l
e
v
a
l
u
a
ti
o
n
s
o
n
imb
a
lan
c
e
d
a
n
d
c
r
o
ss
-
d
a
tas
e
t
se
tt
in
g
s
sh
o
w
sta
b
le
m
in
o
rit
y
-
c
las
s
p
e
rfo
r
m
a
n
c
e
a
n
d
g
o
o
d
g
e
n
e
ra
li
z
a
ti
o
n
.
T
h
e
re
su
lt
s
c
o
n
firm
t
h
a
t
th
e
p
r
o
p
o
se
d
a
u
t
o
m
a
ted
o
p
ti
m
iza
ti
o
n
fra
m
e
wo
rk
o
ffe
rs
a
ro
b
u
st,
sc
a
lab
le,
a
n
d
c
li
n
ica
ll
y
r
e
lev
a
n
t
so
lu
ti
o
n
f
o
r
b
ra
in
tu
m
o
r
d
e
tec
ti
o
n
,
re
p
re
se
n
ti
n
g
a
sig
n
ifi
c
a
n
t
a
d
v
a
n
c
e
m
e
n
t
o
v
e
r
m
a
n
u
a
ll
y
t
u
n
e
d
d
e
e
p
lea
rn
in
g
a
p
p
ro
a
c
h
e
s.
K
ey
w
o
r
d
s
:
B
r
ain
tu
m
o
r
d
etec
tio
n
Dee
p
r
ein
f
o
r
ce
m
en
t le
ar
n
in
g
E
f
f
icien
tNet
Hy
p
er
p
ar
a
m
eter
tu
n
in
g
R
ein
f
o
r
ce
m
en
t le
ar
n
i
n
g
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
:
Aash
u
to
s
h
Kh
ar
b
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics an
d
C
o
m
m
u
n
icatio
n
E
n
g
in
ee
r
i
n
g
,
DC
R
U
ST
Mu
r
th
al,
Har
y
a
n
a,
I
n
d
ia
E
m
ail:
aa
s
h
u
to
s
h
k
h
ar
b
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
B
r
ain
tu
m
o
r
s
r
em
ain
am
o
n
g
t
h
e
m
o
s
t
d
ev
astatin
g
n
eu
r
o
lo
g
ical
d
is
o
r
d
er
s
,
with
tim
ely
a
n
d
ac
cu
r
ate
d
etec
tio
n
b
ein
g
cr
itical
to
p
ati
en
t su
r
v
iv
al.
Ho
wev
e
r
,
s
tan
d
ar
d
m
ag
n
etic
r
eso
n
an
ce
im
ag
i
n
g
(
MRI
)
d
iag
n
o
s
tics
d
ep
en
d
h
ea
v
ily
o
n
r
ad
io
lo
g
is
t
ex
p
er
tis
e
an
d
m
ay
b
e
d
ela
y
ed
d
u
e
to
in
c
r
ea
s
in
g
ca
s
elo
ad
s
an
d
s
u
b
jectiv
e
in
ter
p
r
etatio
n
.
Au
to
m
ated
,
r
el
iab
le
s
y
s
tem
s
ar
e
th
er
ef
o
r
e
u
r
g
en
tly
n
ee
d
e
d
to
s
u
p
p
o
r
t
clin
ician
s
b
y
r
ed
u
cin
g
d
iag
n
o
s
tic
v
ar
iab
ilit
y
a
n
d
en
h
an
cin
g
ea
r
ly
in
ter
v
e
n
tio
n
.
I
n
r
ec
en
t
y
ea
r
s
,
d
ee
p
lear
n
i
n
g
h
as
r
ev
o
l
u
tio
n
ized
th
e
f
iel
d
o
f
m
ed
ical
im
ag
e
a
n
aly
s
is
,
o
f
f
er
in
g
r
em
ar
k
ab
le
im
p
r
o
v
em
en
ts
in
a
cc
u
r
ac
y
,
s
p
ee
d
,
an
d
c
o
n
s
is
ten
cy
o
v
er
t
r
ad
itio
n
al
m
eth
o
d
s
.
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs),
in
p
ar
ticu
lar
,
h
av
e
d
e
m
o
n
s
tr
ated
s
tr
o
n
g
p
er
f
o
r
m
an
ce
in
d
etec
tin
g
an
d
class
if
y
in
g
b
r
ain
tu
m
o
r
s
f
r
o
m
MRI
s
ca
n
s
,
ef
f
ec
tiv
ely
lear
n
in
g
co
m
p
le
x
s
p
atial
an
d
s
tr
u
ctu
r
al
p
atter
n
s
wit
h
o
u
t
th
e
n
ee
d
f
o
r
h
an
d
cr
a
f
ted
f
ea
tu
r
es.
State
-
of
-
th
e
-
a
r
t
ar
ch
itectu
r
es
s
u
ch
as
E
f
f
icien
tNet
[
1
]
h
av
e
f
u
r
th
er
o
p
tim
ized
p
er
f
o
r
m
an
ce
b
y
b
alan
ci
n
g
m
o
d
el
d
ep
th
,
wid
th
,
an
d
r
es
o
lu
tio
n
.
Stu
d
ies
s
u
ch
as
th
e
B
r
aT
S
b
en
c
h
m
ar
k
ch
allen
g
e
[
2
]
h
av
e
h
ig
h
li
g
h
te
d
th
e
ef
f
ec
tiv
en
ess
o
f
C
NN
-
b
ased
m
o
d
els
in
s
eg
m
en
tin
g
an
d
class
if
y
in
g
b
r
ain
tu
m
o
r
r
eg
io
n
s
,
s
h
o
wca
s
in
g
th
e
p
r
ac
tical
v
alu
e
o
f
d
ee
p
lear
n
i
n
g
in
clin
ical
co
n
tex
ts
.
Hy
b
r
i
d
d
e
ep
C
NN
m
o
d
els
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Dee
p
r
ein
fo
r
ce
men
t le
a
r
n
in
g
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
fr
a
mewo
r
k
u
s
in
g
o
p
tu
n
a
fo
r
…
(
A
a
s
h
u
t
o
s
h
K
h
a
r
b
)
1353
h
av
e
d
em
o
n
s
tr
ated
m
u
lti
-
clas
s
class
if
icatio
n
with
r
em
ar
k
ab
le
ac
cu
r
ac
ies
—
u
p
t
o
9
9
.
5
3
%
f
o
r
t
u
m
o
r
d
etec
tio
n
an
d
9
3
.
8
1
%
f
o
r
tu
m
o
r
-
ty
p
e
cl
ass
if
icatio
n
ac
r
o
s
s
g
lio
m
a,
m
e
n
in
g
io
m
a,
p
itu
itar
y
,
an
d
m
etastatic
tu
m
o
r
s
u
s
in
g
g
r
id
-
s
ea
r
ch
h
y
p
er
p
ar
am
eter
o
p
t
im
izatio
n
[
3
]
.
A
d
d
itio
n
ally
,
lig
h
tweig
h
t
U
-
Net
v
ar
ian
ts
lik
e
L
eU
-
Net
an
d
tr
an
s
f
er
-
lear
n
in
g
-
en
h
an
ce
d
ar
c
h
itectu
r
es
s
u
ch
as
Sq
u
ee
ze
Net
h
av
e
ac
h
ie
v
ed
s
eg
m
e
n
tatio
n
a
cc
u
r
ac
ies
b
etwe
en
9
4
% a
n
d
9
8
%,
o
f
f
er
in
g
f
aster
an
d
r
eso
u
r
ce
-
ef
f
icien
t a
lter
n
ativ
es su
itab
le
f
o
r
clin
ical
s
ettin
g
s
[
4
]
–
[
6
]
.
Sh
awo
n
et
a
l.
[
7
]
tack
led
t
h
e
co
m
m
o
n
is
s
u
e
o
f
class
im
b
ala
n
ce
in
MRI
d
atasets
u
s
in
g
c
o
s
t
-
s
en
s
itiv
e
tu
n
in
g
alo
n
g
s
id
e
p
o
p
u
lar
C
N
N
ar
ch
itectu
r
es
lik
e
I
n
ce
p
tio
n
V3
,
R
e
s
Net5
0
,
an
d
E
f
f
icien
tNetB
0
.
T
h
eir
co
s
t
-
s
en
s
itiv
e
I
n
ce
p
tio
n
V3
attain
e
d
an
im
p
r
ess
iv
e
9
9
.
3
3
%
ac
cu
r
ac
y
o
n
b
alan
ce
d
d
ata,
wh
ile
ex
p
lain
ab
ilit
y
t
o
o
ls
r
ev
ea
led
d
ee
p
er
m
o
d
el
in
s
ig
h
ts
.
E
v
en
o
n
im
b
alan
ce
d
d
ata,
th
ey
m
an
a
g
ed
t
o
in
cr
ea
s
e
ac
cu
r
ac
y
b
y
ap
p
r
o
x
im
ately
4
%,
h
i
g
h
lig
h
ti
n
g
th
e
im
p
o
r
tan
ce
o
f
f
air
n
es
s
in
d
etec
tio
n
s
y
s
tem
s
.
Díaz
-
Per
n
as
et
a
l
.
[
8
]
d
esig
n
ed
a
m
u
ltis
ca
le
C
NN
m
ir
r
o
r
in
g
h
u
m
an
v
is
u
al
p
r
o
ce
s
s
in
g
b
y
an
aly
zin
g
MRI
i
n
p
u
ts
at
th
r
ee
r
eso
lu
tio
n
s
.
Op
er
atin
g
o
n
a
d
ataset
o
f
3
,
0
6
4
im
ag
es,
th
eir
en
d
-
to
-
en
d
m
o
d
el
ac
h
iev
ed
9
7
.
3
%
ac
cu
r
ac
y
ac
r
o
s
s
m
en
in
g
io
m
a
,
g
lio
m
a
,
an
d
p
it
u
itar
y
tu
m
o
r
ty
p
es
with
o
u
t
p
r
ep
r
o
ce
s
s
in
g
—
d
em
o
n
s
tr
atin
g
th
e
v
alu
e
o
f
s
ca
le
-
awa
r
e
f
ea
tu
r
e
ex
tr
ac
tio
n
.
Ka
r
ag
o
z
et
a
l.
[
9
]
in
tr
o
d
u
ce
d
R
esViT,
a
h
y
b
r
id
co
m
b
in
in
g
co
n
v
o
lu
tio
n
al
an
d
tr
an
s
f
o
r
m
er
b
lo
ck
s
.
T
h
e
y
p
r
e
tr
ain
ed
th
e
m
o
d
el
th
r
o
u
g
h
s
e
lf
-
s
u
p
er
v
is
ed
lear
n
i
n
g
(
SS
L
)
,
s
y
n
th
esizin
g
MRI
s
ca
n
s
to
lear
n
g
en
e
r
al
r
ep
r
esen
tatio
n
s
b
ef
o
r
e
f
in
e
-
tu
n
i
n
g
o
n
class
if
icatio
n
.
T
h
e
m
o
d
el
ac
h
i
ev
ed
a
p
p
r
o
x
im
ately
9
8
.
5
%
ac
cu
r
ac
y
,
s
h
o
win
g
t
h
a
t
lim
ited
lab
eled
MRI
d
ataset
s
ca
n
s
till
b
e
lev
er
ag
e
d
ef
f
ec
tiv
ely
th
r
o
u
g
h
SS
L
.
Aam
ir
et
a
l.
[
1
0
]
d
e
v
elo
p
ed
a
C
NN
o
p
tim
ized
th
r
o
u
g
h
s
y
s
te
m
atic
h
y
p
e
r
p
ar
a
m
eter
tu
n
in
g
—
co
v
er
in
g
lear
n
in
g
r
ates,
b
atch
s
izes,
f
ilter
d
im
en
s
io
n
s
,
an
d
d
e
p
th
—
to
ac
h
iev
e
ap
p
r
o
x
im
ately
9
7
% o
n
ac
cu
r
a
cy
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
ac
r
o
s
s
th
r
ee
K
ag
g
le
MRI
d
atasets
.
T
h
is
h
ig
h
lig
h
ts
h
o
w
p
r
o
p
er
ca
lib
r
atio
n
ca
n
s
ig
n
if
ican
tly
ele
v
ate
m
o
d
el
r
o
b
u
s
tn
ess
.
Mu
g
h
al
et
a
l.
[
1
1
]
ex
p
l
o
r
ed
th
e
ad
d
itio
n
o
f
a
d
e
n
o
is
in
g
s
tep
—
u
s
in
g
an
is
o
tr
o
p
ic
d
if
f
u
s
io
n
—
b
ef
o
r
e
f
ee
d
i
n
g
s
ca
n
s
in
to
C
NNs
lik
e
E
f
f
icien
tNet,
R
esNet1
5
2
V2
,
VGG,
an
d
v
is
io
n
tr
an
s
f
o
r
m
er
(
ViT
)
.
E
f
f
icien
tNet
led
th
e
p
ac
k
with
9
8
%
ac
cu
r
ac
y
,
u
n
d
er
s
co
r
in
g
th
e
im
p
ac
t
o
f
d
en
o
is
in
g
p
ip
elin
es
o
n
f
in
al
m
o
d
el
p
er
f
o
r
m
an
ce
.
Sin
g
h
an
d
J
ain
[
1
2
]
cr
af
ted
a
lig
h
tweig
h
t
b
in
a
r
y
C
NN
f
o
r
t
u
m
o
r
d
e
tectio
n
o
n
a
Kag
g
le
MRI
d
ataset.
Desp
ite
its
s
im
p
licity
,
it
ac
h
iev
ed
9
9
.
3
1
%
ac
cu
r
ac
y
,
p
r
o
v
in
g
th
at
s
m
aller
m
o
d
els
ca
n
s
till
o
f
f
er
u
ltra
-
h
ig
h
p
er
f
o
r
m
an
ce
an
d
m
ay
b
e
id
ea
l f
o
r
lo
w
-
r
eso
u
r
ce
cl
in
ical
en
v
ir
o
n
m
en
ts
.
Patel
an
d
Gan
d
h
i
[
1
3
]
co
m
b
i
n
ed
C
NN
f
ea
tu
r
e
e
x
tr
ac
to
r
s
with
t
r
an
s
f
o
r
m
er
en
co
d
er
s
,
p
r
etr
ain
ed
v
ia
s
elf
-
s
u
p
er
v
is
ed
m
ask
ed
a
u
to
en
co
d
er
s
(
MA
E
s
)
to
tac
k
le
th
e
f
o
u
r
-
way
class
if
icatio
n
task
.
T
h
eir
h
y
b
r
id
ap
p
r
o
ac
h
ac
h
iev
ed
9
4
%
ac
cu
r
ac
y
an
d
F1
,
o
u
tp
er
f
o
r
m
i
n
g
n
o
n
-
SS
L
an
d
s
tan
d
alo
n
e
ar
ch
itectu
r
es
b
y
ap
p
r
o
x
im
ately
6
%
–
7
%
[
1
3
]
.
Me
s
h
r
a
m
et
a
l.
[
1
4
]
a
p
p
lied
t
h
e
s
win
tr
an
s
f
o
r
m
er
to
n
o
t
o
n
ly
class
if
y
b
u
t
also
lo
ca
lize
b
r
ain
tu
m
o
r
s
,
d
eliv
er
in
g
s
eg
m
en
tatio
n
m
ask
s
an
d
t
u
m
o
r
d
i
m
en
s
io
n
esti
m
ates
v
ia
PDF.
T
h
is
alig
n
s
wel
l
with
r
ad
io
lo
g
is
ts
’
wo
r
k
f
lo
ws,
im
p
r
o
v
in
g
clin
ical
u
tili
ty
th
r
o
u
g
h
co
m
b
in
ed
d
e
tec
tio
n
an
d
s
p
atial
awa
r
en
ess
.
Sev
er
al
r
ec
en
t
s
tu
d
ies
[
3
]
,
[
5
]
,
[
7
]
,
[
9
]
,
[
1
2
]
h
a
v
e
s
h
o
wn
th
at
th
e
r
ef
in
em
e
n
t
in
th
e
C
NN
m
o
d
els
m
ay
y
ield
s
to
an
ac
c
u
r
ac
y
r
an
g
in
g
f
r
o
m
9
5
%
to
9
9
%
ap
p
r
o
x
im
ately
.
I
n
2
0
2
5
,
f
u
r
t
h
er
r
e
f
i
n
em
en
ts
co
n
tin
u
ed
with
a
C
NN
m
o
d
el
tai
lo
r
e
d
f
o
r
b
in
a
r
y
tu
m
o
r
d
etec
tio
n
o
n
K
ag
g
le
MRI
d
atasets
,
ac
h
iev
in
g
9
9
.
3
1
%
ac
c
u
r
ac
y
wh
ile
m
ain
tain
in
g
ar
c
h
itectu
r
a
l simp
licity
[
1
2
]
.
Simu
ltan
eo
u
s
ly
,
s
tu
d
ies
ev
al
u
atin
g
m
u
ltip
le
MRI
m
o
d
al
ities
co
m
b
in
ed
with
tr
an
s
f
er
lear
n
in
g
r
ep
o
r
ted
class
if
icatio
n
ac
cu
r
a
cies
ar
o
u
n
d
9
8
.
6
%
ac
r
o
s
s
T
1
,
T
2
,
an
d
FLAI
R
s
eq
u
e
n
ce
s
,
u
n
d
er
lin
i
n
g
t
h
e
clin
ical
r
elev
an
ce
o
f
s
elec
tin
g
th
e
m
o
s
t
in
f
o
r
m
ativ
e
im
a
g
in
g
co
m
b
in
atio
n
s
[
1
5
]
.
Desp
ite
th
ese
ad
v
an
ce
s
,
d
esig
n
in
g
o
p
tim
al
d
ee
p
lea
r
n
in
g
m
o
d
els
r
em
ain
s
a
ch
allen
g
e
d
u
e
to
th
e
v
ast
n
u
m
b
er
o
f
tu
n
ab
le
h
y
p
er
p
ar
am
eter
s
an
d
ar
ch
ite
ctu
r
al
ch
o
ices.
R
ec
en
t
ap
p
r
o
ac
h
es
h
av
e
ex
p
l
o
r
ed
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
to
au
to
m
ate
th
is
p
r
o
ce
s
s
,
allo
wi
n
g
m
o
d
els
to
ad
ap
tiv
ely
lear
n
th
e
b
est
co
n
f
ig
u
r
atio
n
f
o
r
class
if
icatio
n
task
s
[
1
6
]
–
[
1
8
]
.
Mo
tiv
ated
b
y
th
ese
ad
v
an
ce
s
,
th
is
p
ap
er
p
r
o
p
o
s
es
a
n
o
v
el
ap
p
r
o
ac
h
in
teg
r
atin
g
E
f
f
icien
tNet
-
b
ased
C
NNs
with
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
in
s
p
ir
ed
f
in
e
-
tu
n
i
n
g
.
T
h
e
s
tu
d
y
u
tili
ze
s
th
e
B
r
aT
S
2
0
2
0
d
ataset
to
v
alid
ate
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
,
co
n
tr
i
b
u
tin
g
to
t
h
e
g
r
o
win
g
b
o
d
y
o
f
wo
r
k
f
o
cu
s
ed
o
n
d
ev
elo
p
in
g
in
tellig
en
t,
d
ata
-
d
r
iv
en
to
o
ls
f
o
r
im
p
r
o
v
in
g
ea
r
ly
tu
m
o
r
d
etec
tio
n
an
d
s
u
p
p
o
r
t
in
g
clin
ical
d
ec
is
io
n
-
m
a
k
in
g
.
T
h
e
m
a
j
o
r
c
o
n
t
r
i
b
u
ti
o
n
s
o
f
t
h
is
w
o
r
k
a
r
e
as
f
o
l
l
o
w
s
:
i
)
p
r
o
p
o
s
e
d
a
n
a
d
a
p
t
i
v
e
d
e
e
p
l
e
a
r
n
i
n
g
f
r
a
m
e
w
o
r
k
f
o
r
b
r
a
i
n
t
u
m
o
r
d
e
t
e
c
t
i
o
n
u
s
i
n
g
M
R
I
i
m
a
g
es
,
w
h
i
c
h
c
o
m
b
i
n
e
s
t
h
e
f
e
a
t
u
r
e
e
x
t
r
a
ct
i
o
n
c
a
p
a
b
i
l
i
t
y
o
f
E
f
f
i
c
i
e
n
t
Ne
tB
0
w
it
h
t
h
e
a
d
a
p
ti
v
e
t
u
n
i
n
g
p
o
w
e
r
o
f
r
ei
n
f
o
r
c
em
e
n
t
l
e
a
r
n
i
n
g
;
i
i
)
r
ei
n
f
o
r
c
e
m
en
t
l
e
a
r
n
i
n
g
i
n
s
p
i
r
e
d
O
p
t
u
n
a
f
r
a
m
e
w
o
r
k
i
s
u
s
e
d
t
o
d
e
t
e
r
m
i
n
e
t
h
e
m
o
s
t
e
f
f
e
ct
i
v
e
c
la
s
s
i
f
i
e
r
a
r
c
h
it
e
c
t
u
r
e
a
n
d
t
o
a
u
to
m
a
t
i
c
a
ll
y
f
i
n
e
-
t
u
n
e
c
r
i
t
i
c
al
h
y
p
e
r
p
a
r
a
m
e
t
e
r
s
s
u
c
h
a
s
l
e
a
r
n
i
n
g
r
a
t
e
,
b
a
t
c
h
s
i
z
e
,
n
u
m
b
e
r
o
f
e
p
o
c
h
s
,
a
n
d
d
r
o
p
o
u
t
r
a
t
e
;
i
i
i
)
t
h
is
d
u
a
l
-
l
e
v
e
l
o
p
t
i
m
i
z
a
ti
o
n
e
n
s
u
r
e
s
t
h
a
t
t
h
e
f
in
a
l
m
o
d
e
l
i
s
b
o
t
h
s
t
r
u
c
t
u
r
al
l
y
o
p
t
i
m
a
l
a
n
d
w
el
l
-
c
a
li
b
r
a
t
e
d
f
o
r
tr
a
i
n
i
n
g
d
y
n
a
m
i
c
s
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
p
r
o
p
o
s
ed
r
esear
ch
m
eth
o
d
o
lo
g
y
f
o
r
au
to
m
ated
b
r
ain
tu
m
o
r
d
etec
tio
n
f
r
o
m
MRI
im
ag
es.
I
t
o
u
tlin
es
th
e
o
v
er
all
f
r
a
m
ewo
r
k
,
in
clu
d
i
n
g
d
ata
p
r
e
p
r
o
ce
s
s
in
g
,
E
f
f
icien
tNet
-
b
ased
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
th
e
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
–
i
n
s
p
ir
ed
Op
tu
n
a
o
p
tim
izatio
n
s
tr
ateg
y
u
s
ed
to
j
o
in
tly
o
p
tim
ize
th
e
class
if
ier
ar
ch
itectu
r
e
an
d
tr
ai
n
in
g
h
y
p
er
p
ar
am
eter
s
as sh
o
w
n
in
Fig
u
r
e
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
3
5
2
-
1
3
6
3
1354
Fig
u
r
e
1
.
Ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
E
f
f
icien
tNet
-
b
ased
r
ei
n
f
o
r
ce
m
e
n
t le
ar
n
in
g
-
in
s
p
ir
ed
o
p
tim
izatio
n
f
r
am
ewo
r
k
f
o
r
b
r
ain
t
u
m
o
r
d
et
ec
tio
n
2
.
1
.
Da
t
a
s
et
des
cr
iptio
n
I
n
th
is
wo
r
k
,
B
r
aT
S
2
0
2
0
d
at
aset
h
as
b
ee
n
u
s
ed
.
T
h
e
B
r
aT
S
2
0
2
0
d
ataset
is
a
wid
ely
u
s
ed
m
ed
ical
im
ag
in
g
d
ataset
f
o
r
b
r
ai
n
tu
m
o
r
d
etec
tio
n
a
n
d
s
eg
m
e
n
tatio
n
.
I
t
in
cl
u
d
es
MRI
s
ca
n
s
in
f
o
u
r
m
o
d
alities
—
T
1
,
T
1
Gd
,
T
2
,
an
d
FLAI
R
—
o
f
p
atien
ts
with
h
ig
h
-
an
d
lo
w
-
g
r
ad
e
g
lio
m
as,
alo
n
g
with
ex
p
e
r
t
-
an
n
o
tated
tu
m
o
r
r
eg
io
n
s
.
I
ts
r
ea
lis
tic
clin
ical
v
ar
iatio
n
s
an
d
r
ich
a
n
n
o
tatio
n
s
m
ak
e
it
a
r
eliab
le
b
en
ch
m
ar
k
f
o
r
tr
ain
in
g
a
n
d
ev
alu
atin
g
d
ee
p
lear
n
i
n
g
m
o
d
els in
b
r
ain
tu
m
o
r
an
aly
s
is
.
2
.
2
.
Da
t
a
pre
pro
ce
s
s
ing
T
h
e
d
ata
p
r
ep
r
o
ce
s
s
in
g
h
as
b
ee
n
d
o
n
e
to
s
tan
d
ar
d
ize
th
e
in
p
u
t,
m
ak
e
it
less
v
ar
iab
le,
an
d
im
p
r
o
v
e
th
e
q
u
ality
o
f
th
e
in
p
u
t
d
ata
a
n
d
im
p
r
o
v
e
m
o
d
el
p
er
f
o
r
m
an
ce
b
ef
o
r
e
d
ee
p
lea
r
n
in
g
o
n
th
e
B
r
aT
S
2
0
2
0
MRI
s
ca
n
s
.
T
h
e
f
o
llo
win
g
s
tep
s
ar
e
p
er
f
o
r
m
ed
:
I
m
a
g
e
r
esizin
g
: a
l
l
o
f
th
e
MRI
s
lices a
r
e
r
esized
to
2
2
4
×
2
2
4
p
ix
els,
s
o
as
to
m
ak
e
s
u
r
e
th
at
all
th
e
s
am
p
les
ar
e
t
h
e
s
am
e
s
ize
wh
ile
k
ee
p
in
g
th
e
im
p
o
r
tan
t
s
p
at
ial
f
ea
tu
r
es
th
at
ar
e
n
ee
d
ed
to
id
en
tif
y
tu
m
o
r
s
.
No
r
m
aliza
tio
n
:
h
er
e,
m
in
-
m
a
x
s
ca
lin
g
is
u
s
ed
to
b
r
in
g
p
ix
el
in
ten
s
ities
d
o
wn
to
a
r
an
g
e
o
f
[
0
,
1
]
.
T
h
is
m
ak
es
it
less
lik
ely
th
at
s
ca
n
s
will
h
a
v
e
d
if
f
er
en
t
lev
els
o
f
b
r
ig
h
tn
ess
a
n
d
c
o
n
tr
ast.
I
t
also
m
ak
es
s
u
r
e
th
at
th
e
in
p
u
t
d
is
tr
ib
u
tio
n
is
alwa
y
s
th
e
s
am
e,
wh
ich
s
p
ee
d
s
u
p
tr
ain
in
g
.
Mo
d
ality
s
elec
tio
n
an
d
s
tack
in
g
:
b
ased
o
n
th
e
e
x
p
e
r
im
en
tal
s
etu
p
,
ch
o
s
en
c
o
m
b
in
atio
n
s
o
f
th
e
f
o
u
r
a
v
ailab
le
MRI
m
o
d
alities
(
T
1
,
T
1
Gd
,
T
2
,
an
d
FLAI
R
)
ar
e
s
tack
ed
as
in
p
u
t
ch
a
n
n
els.
T
h
is
h
elp
s
th
e
m
o
d
el
lear
n
f
r
o
m
d
if
f
er
e
n
t
co
n
tr
asts
th
at
s
h
o
w
d
if
f
er
e
n
t th
in
g
s
ab
o
u
t th
e
tu
m
o
r
.
Sli
ce
e
x
tr
ac
tio
n
an
d
f
ilter
in
g
: in
th
is
wo
r
k
2
D
s
lices h
av
e
b
ee
n
ex
tr
ac
ted
f
r
o
m
th
e
3
D
v
o
lu
m
es
th
at
h
av
e
v
is
ib
le
tu
m
o
r
ar
ea
s
.
T
h
is
s
tep
r
em
o
v
es
a
lo
t
o
f
n
o
is
e
in
th
e
d
ataset
an
d
h
elp
s
th
e
m
o
d
el
f
o
cu
s
o
n
im
p
o
r
ta
n
t
lear
n
in
g
g
o
als.
L
ab
el
en
co
d
in
g
:
g
r
o
u
n
d
-
tr
u
th
lab
els
ar
e
u
s
ed
to
ca
teg
o
r
ize
th
e
d
if
f
er
e
n
t ty
p
e
s
o
f
tu
m
o
r
s
in
to
“
tu
m
o
r
”
an
d
“
n
o
tu
m
o
r
”
.
2
.
3
.
F
e
a
t
ure
ex
t
r
a
ct
io
n us
ing
E
f
f
icient
Net
B
0
T
h
e
E
f
f
icien
tNetB
0
u
s
es
co
m
p
o
u
n
d
s
ca
lin
g
to
f
in
d
th
e
r
ig
h
t
b
alan
ce
b
etwe
en
d
e
p
th
,
wid
th
,
an
d
r
eso
lu
tio
n
,
wh
ich
m
a
k
es
it
f
ast
an
d
ef
f
icien
t
th
er
ef
o
r
e
it
h
as
b
ee
n
u
s
ed
in
th
is
wo
r
k
as
th
e
b
ac
k
b
o
n
e
m
o
d
e
l
to
ex
tr
ac
t
f
ea
tu
r
es
f
r
o
m
t
h
e
2
D
s
lices.
T
h
e
f
ir
s
t
f
ew
lay
er
s
o
f
E
f
f
icien
tNetB
0
h
as
b
ee
n
f
r
o
ze
n
an
d
we
tak
e
f
ea
tu
r
e
em
b
e
d
d
in
g
s
f
r
o
m
t
h
e
lay
er
ju
s
t
b
ef
o
r
e
th
e
last
o
n
e.
T
h
e
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
a
g
en
t
th
en
d
y
n
am
ically
o
p
tim
izes th
ese
em
b
ed
d
in
g
s
b
y
s
en
d
i
n
g
th
e
m
to
a
cu
s
to
m
izab
le
class
if
ier
h
ea
d
.
2
.
4
.
Reinf
o
rc
e
m
ent
lea
rning
ins
pire
d dua
l o
ptim
iza
t
io
n us
ing
O
ptuna
R
ein
f
o
r
ce
m
en
t
lea
r
n
i
n
g
p
lay
s
a
ce
n
tr
al
r
o
le
in
o
u
r
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
b
y
au
t
o
m
atin
g
t
wo
m
ajo
r
asp
ec
ts
o
f
m
o
d
el
d
ev
elo
p
m
en
t:
1
.
Desig
n
in
g
th
e
o
p
tim
al
class
if
ier
s
tr
u
ctu
r
e
(
i.e
.
,
th
e
co
n
f
ig
u
r
atio
n
o
f
f
u
lly
co
n
n
ec
ted
la
y
er
s
)
,
a
n
d
2
.
Se
lectin
g
th
e
b
est
tr
ain
i
n
g
h
y
p
e
r
p
ar
am
eter
s
(
e.
g
.
,
lear
n
in
g
r
ate,
b
atch
s
ize,
n
u
m
b
er
o
f
ep
o
c
h
s
,
o
p
tim
izer
ty
p
e
)
.
T
h
is
in
tellig
en
t
au
to
m
atio
n
a
llo
ws
th
e
s
y
s
tem
to
ex
p
lo
r
e
an
d
id
en
tify
h
ig
h
-
p
er
f
o
r
m
in
g
c
o
n
f
ig
u
r
atio
n
s
w
ith
o
u
t
m
an
u
al
in
ter
v
e
n
tio
n
—
esp
ec
ially
b
en
ef
icial
in
d
ee
p
lear
n
in
g
,
wh
er
e
m
an
u
al
tu
n
i
n
g
is
tim
e
-
co
n
s
u
m
in
g
an
d
o
f
ten
s
u
b
o
p
tim
al.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Dee
p
r
ein
fo
r
ce
men
t le
a
r
n
in
g
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
fr
a
mewo
r
k
u
s
in
g
o
p
tu
n
a
fo
r
…
(
A
a
s
h
u
t
o
s
h
K
h
a
r
b
)
1355
2
.
4
.
1
.
R
einf
o
rc
em
ent
lea
rnin
g
s
et
up
o
v
er
v
iew
I
n
th
is
s
tu
d
y
,
we
u
tili
ze
Op
tu
n
a
as
a
r
ein
f
o
r
ce
m
en
t
lea
r
n
in
g
-
in
s
p
ir
ed
f
r
am
ewo
r
k
to
au
to
m
ate
th
e
d
u
al
task
o
f
(
1
)
class
if
ier
s
tr
u
ctu
r
e
d
esig
n
an
d
(
2
)
tr
ain
i
n
g
h
y
p
e
r
p
ar
am
eter
t
u
n
in
g
.
T
h
is
n
o
t
o
n
ly
r
ep
lace
s
m
an
u
al
ex
p
er
im
en
tatio
n
b
u
t
also
s
ig
n
if
ican
tly
im
p
r
o
v
es
m
o
d
el
p
er
f
o
r
m
a
n
ce
an
d
ef
f
icien
cy
t
h
r
o
u
g
h
d
ata
-
d
r
iv
en
s
ea
r
ch
.
Op
tu
n
a
f
o
r
m
u
lates
o
p
tim
izatio
n
as
a
b
lac
k
-
b
o
x
f
u
n
ctio
n
m
in
im
izatio
n
p
r
o
b
lem
.
I
n
th
is
ca
s
e,
th
e
f
u
n
ctio
n
is
to
m
ax
im
ize
th
e
m
o
d
el
v
alid
atio
n
ac
cu
r
ac
y
,
ev
alu
ated
o
v
er
a
s
et
o
f
h
y
p
er
p
ar
am
eter
s
an
d
ar
ch
itectu
r
e
d
ec
is
io
n
s
.
E
ac
h
tr
ial
in
Op
tu
n
a
r
e
p
r
esen
ts
a
f
u
ll
ep
is
o
d
e
o
f
m
o
d
el
b
u
ild
in
g
an
d
ev
alu
atio
n
:
-
T
r
ial
(
ep
is
o
d
e)
:
s
am
p
le
a
s
et
o
f
v
alu
es
th
at
is
n
u
m
b
er
a
n
d
s
ize
o
f
d
en
s
e
lay
er
s
,
ac
tiv
atio
n
f
u
n
ctio
n
,
d
r
o
p
o
u
t r
ate,
o
p
tim
izer
,
lear
n
i
n
g
r
ate,
b
atch
s
ize,
n
u
m
b
er
o
f
tr
ain
in
g
ep
o
ch
s
-
E
n
v
ir
o
n
m
en
t:
a
d
ee
p
lear
n
i
n
g
m
o
d
el
is
co
n
s
tr
u
cte
d
u
s
in
g
t
h
e
s
elec
ted
co
n
f
ig
u
r
atio
n
,
tr
ai
n
ed
o
n
a
s
m
all
s
u
b
s
et
o
f
th
e
d
ataset,
an
d
ev
al
u
ated
o
n
a
v
alid
atio
n
s
et.
-
R
ewa
r
d
: a
s
ca
lar
v
alu
e
is
ca
lcu
lated
b
ased
o
n
m
o
d
el
p
er
f
o
r
m
an
ce
th
at
is
v
alid
atio
n
ac
c
u
r
ac
y
in
th
is
wo
r
k
.
-
Sam
p
ler
u
p
d
ate:
u
s
e
th
at
r
e
war
d
to
g
u
id
e
t
h
e
n
e
x
t
t
r
ial
v
ia
B
ay
esian
o
p
tim
izatio
n
t
h
at
is
tr
ee
-
s
tr
u
ctu
r
ed
p
ar
ze
n
esti
m
ato
r
(
T
PE
)
in
th
is
wo
r
k
,
im
p
r
o
v
in
g
its
ch
an
ce
s
o
f
s
elec
tin
g
b
etter
co
n
f
ig
u
r
atio
n
s
in
f
u
tu
r
e
e
p
is
o
d
es.
Op
tu
n
a
b
u
ild
s
a
p
r
o
b
ab
ilit
y
m
o
d
el
o
f
p
r
o
m
is
in
g
r
eg
io
n
s
o
f
th
e
s
ea
r
ch
s
p
ac
e,
s
im
ilar
t
o
a
p
o
licy
u
p
d
ate
i
n
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
,
an
d
p
r
io
r
itizes
ex
p
lo
r
in
g
th
e
m
wh
ile
s
till
s
am
p
lin
g
f
r
o
m
u
n
ex
p
lo
r
e
d
ar
ea
s
(
ex
p
lo
r
atio
n
v
s
.
ex
p
lo
itatio
n
)
.
2
.
4
.
2
.
I
nte
g
ra
t
io
n
re
info
rc
e
m
ent
lea
rning
in t
he
pip
eline
a.
Step
1
:
d
esig
n
o
f
o
b
jectiv
e
f
u
n
ctio
n
T
h
e
o
b
jectiv
e
f
u
n
ctio
n
is
k
e
y
co
m
p
o
n
en
t
o
f
t
h
e
r
ein
f
o
r
c
em
en
t
lear
n
in
g
-
i
n
s
p
ir
ed
o
p
tim
izatio
n
f
r
am
ewo
r
k
,
as
it
g
u
id
es
t
h
e
s
e
ar
ch
to
war
d
o
p
tim
al
class
if
ier
ar
ch
itectu
r
es
an
d
h
y
p
er
p
ar
am
eter
co
n
f
ig
u
r
atio
n
s
.
I
n
th
is
wo
r
k
,
th
e
o
b
jectiv
e
f
u
n
ctio
n
p
er
f
o
r
m
s
th
e
f
o
llo
win
g
t
ask
s
:
-
B
u
ild
s
th
e
m
o
d
el
u
s
in
g
s
am
p
l
ed
co
n
f
i
g
u
r
atio
n
.
-
T
r
ain
s
o
n
a
s
u
b
s
et
o
f
d
ata.
-
E
v
alu
ates v
alid
atio
n
ac
cu
r
ac
y
,
ar
ea
u
n
d
er
th
e
c
u
r
v
e
(
AUC
)
,
an
d
lo
s
s
.
-
R
etu
r
n
s
a
co
m
p
o
s
ite
r
ewa
r
d
s
co
r
e.
I
n
th
is
s
tu
d
y
,
v
alid
atio
n
ac
cu
r
ac
y
is
s
elec
ted
as
th
e
p
r
im
ar
y
o
p
tim
izatio
n
cr
iter
io
n
d
u
e
t
o
its
ef
f
ec
tiv
en
ess
in
m
ea
s
u
r
in
g
o
v
er
all
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
=
{
η
,
,
,
}
(
1
)
W
h
er
e
θ
d
en
o
te
s
a
ca
n
d
id
ate
co
n
f
ig
u
r
atio
n
s
am
p
led
b
y
th
e
Op
tu
n
a
o
p
tim
izer
,
η
r
ep
r
esen
t
s
th
e
lear
n
in
g
r
ate
,
b
atch
s
ize
(
b
)
,
d
r
o
p
o
u
t
r
ate
(
d
)
,
an
d
n
u
m
b
er
o
f
class
if
ier
lay
er
s
(
n
)
.
Fo
r
ea
ch
tr
ial,
th
e
m
o
d
el
is
tr
ain
ed
u
s
in
g
co
n
f
ig
u
r
atio
n
θ
a
n
d
e
v
alu
ated
o
n
a
v
alid
atio
n
d
ataset.
T
h
e
o
b
jectiv
e
fu
n
ctio
n
is
th
u
s
d
ef
in
ed
as:
ma
x
(
)
=
(
)
(
2
)
T
h
e
v
alid
atio
n
ac
cu
r
ac
y
(
)
is
tr
ea
ted
as
th
e
r
ewa
r
d
s
ig
n
al
an
d
r
etu
r
n
ed
to
th
e
o
p
tim
izer
a
f
ter
ea
ch
tr
ial.
E
ar
ly
s
to
p
p
in
g
is
em
p
lo
y
ed
d
u
r
in
g
tr
ain
i
n
g
to
p
r
e
v
en
t
o
v
er
f
it
tin
g
,
en
s
u
r
in
g
th
at
t
h
e
o
b
jectiv
e
f
u
n
ctio
n
r
ef
lects
tr
u
e
g
en
er
aliza
tio
n
ca
p
ab
ilit
y
r
ath
er
th
an
m
em
o
r
izatio
n
o
f
tr
ain
in
g
d
ata.
T
h
is
f
o
r
m
u
latio
n
en
ab
les
ef
f
icien
t
ex
p
lo
r
atio
n
o
f
th
e
s
ea
r
ch
s
p
ac
e
an
d
s
er
v
es
as
th
e
f
o
u
n
d
atio
n
f
o
r
s
u
b
s
eq
u
en
t
o
p
tim
iz
atio
n
s
tep
s
in
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
.
b.
Step
2
: sear
ch
s
p
ac
e
d
esig
n
T
h
e
s
ea
r
ch
s
p
ac
e
is
f
o
r
m
u
late
d
as
a
m
u
lti
-
d
i
m
en
s
io
n
al
v
ec
t
o
r
wh
e
r
e
ea
c
h
d
im
en
s
io
n
co
r
r
esp
o
n
d
s
to
a
m
o
d
el
d
ec
is
io
n
.
T
h
e
d
ec
i
s
io
n
o
p
tio
n
s
ar
e
s
u
m
m
ar
is
ed
in
T
ab
le
1
.
E
ac
h
c
o
n
f
ig
u
r
atio
n
is
o
n
e
tr
ial
.
Af
ter
s
am
p
lin
g
s
ev
er
al
co
n
f
ig
u
r
atio
n
s
,
th
e
s
am
p
ler
lear
n
s
a
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
th
at
f
av
o
r
s
h
ig
h
-
r
ewa
r
d
(
i.e
.
,
h
ig
h
-
p
er
f
o
r
m
a
n
ce
)
c
h
o
ic
es.
T
ab
le
1
.
Sear
ch
s
p
ac
e
u
s
ed
b
y
th
e
Op
tu
n
a
f
r
am
ewo
r
k
f
o
r
ar
c
h
itectu
r
e
an
d
h
y
p
e
r
p
ar
am
eter
o
p
tim
izatio
n
D
e
n
se
l
a
y
e
r
s
{1
-
5}
Le
a
r
n
i
n
g
r
a
t
e
{1
e
-
6
-
1e
-
2}
N
e
u
r
o
n
s
p
e
r
l
a
y
e
r
{6
4
-
2
5
6
}
B
a
t
c
h
si
z
e
{1
6
,
3
2
,
6
4
}
A
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
{Re
LU
,
E
LU
,
S
e
l
u
}
O
p
t
i
mi
z
e
r
{A
d
a
m,
S
G
D
,
R
M
S
p
r
o
p
}
D
r
o
p
o
u
t
r
a
t
e
{0
.
2
-
0
.
5
}
Ep
o
c
h
s
{1
0
-
1
0
0
}
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
3
5
2
-
1
3
6
3
1356
c.
Step
3
:
o
p
tim
izatio
n
p
r
o
ce
s
s
Op
tu
n
a
r
u
n
s
a
n
u
m
b
er
o
f
tr
ial
s
,
an
d
ea
ch
o
n
e
test
s
a
d
if
f
e
r
e
n
t
g
r
o
u
p
o
f
h
y
p
er
p
ar
am
eter
s
.
I
t
d
o
esn
’
t
u
s
e
b
r
u
te
f
o
r
ce
m
eth
o
d
s
lik
e
Gr
id
Sear
ch
o
r
R
an
d
o
m
Sear
ch
in
s
tead
,
it
u
s
es
s
m
ar
t
alg
o
r
ith
m
s
,
m
o
s
tly
th
e
T
PE,
to
h
el
p
with
th
e
s
ea
r
ch
.
T
PE
lo
o
k
s
at
th
e
r
esu
lts
o
f
p
r
ev
i
o
u
s
tr
ials
to
f
in
d
p
r
o
m
is
i
n
g
ar
ea
s
in
th
e
h
y
p
er
p
ar
am
eter
s
p
ac
e
an
d
s
u
g
g
ests
n
ew
co
m
b
in
atio
n
s
o
f
h
y
p
er
p
ar
am
eter
s
th
at
ar
e
m
o
r
e
li
k
ely
to
wo
r
k
b
etter
.
T
h
is
lets
Op
tu
n
a
q
u
ick
ly
m
o
v
e
th
r
o
u
g
h
th
e
s
ea
r
c
h
s
p
ac
e
an
d
f
in
d
th
e
b
est h
y
p
e
r
p
ar
am
eter
s
.
Op
tu
n
a
also
u
s
es
a
p
r
u
n
in
g
m
ec
h
an
is
m
to
f
in
d
a
n
d
s
to
p
u
n
p
r
o
m
is
in
g
tr
ials
ea
r
ly
in
o
r
d
er
to
s
av
e
tim
e
an
d
co
m
p
u
tin
g
p
o
wer
.
T
h
is
ea
r
ly
s
to
p
p
in
g
is
b
ased
o
n
k
ee
p
i
n
g
an
ey
e
o
n
th
e
in
t
er
m
ed
iate
o
b
jectiv
e
v
alu
es
d
u
r
in
g
a
tr
ial
an
d
en
d
in
g
th
o
s
e
th
at
d
o
n
’
t
m
ee
t
a
ce
r
tai
n
co
n
d
itio
n
.
T
h
is
h
elp
s
k
e
ep
r
eso
u
r
ce
s
f
r
o
m
b
ein
g
wasted
o
n
test
s
th
at
ar
e
u
n
lik
ely
to
f
in
d
th
e
b
est s
o
lu
tio
n
s
.
T
h
e
alg
o
r
ith
m
f
o
r
o
p
tim
izatio
n
u
s
in
g
Op
tu
n
a
is
s
u
m
m
ar
ized
in
Alg
o
r
ith
m
1
.
Alg
o
r
ith
m
1
.
Op
tim
izatio
n
u
s
i
n
g
Op
tu
n
a
Step1: Initialization
a.
Specify
the search space (Table1)
b.
Define objective function: maximize validation accuracy
c.
Identify
starting
point(X
):
Execute
a
small
numbe
r
of
initial
trials
with
randomly
chosen
s
s.
Step2: Sampling the search space (TPE Sampler and Pruning)
Repeat for each
trial
{
a. For each x є X
{
a.
Compute
1
=
(
|
<
∗
)
b.
Compute
2
=
(
|
≥
∗
)
c.
Compute
(
)
=
(
1
2
)
d.
If score(x) > Threshold then add x to l(x), list of parameters yielding high reward.
Else add x to g(x), list of parameters yielding low reward.
}
b. New parameters will be selected a
t the point where the ratio of l(x) and g(x) is
maximized.
c. Evaluate the
trial
by executing the objective function
}
2
.
5
.
F
ina
l
m
o
del selec
t
io
n
Af
ter
th
e
s
ea
r
ch
is
d
o
n
e,
th
e
t
o
p
th
r
ee
co
n
f
ig
u
r
atio
n
s
ar
e
te
s
ted
with
f
u
ll
tr
ain
i
n
g
c
y
cles
to
s
ee
h
o
w
well
th
ey
wo
r
k
.
T
h
e
co
n
f
i
g
u
r
a
tio
n
th
at
p
er
f
o
r
m
s
b
est
o
n
th
e
v
alid
atio
n
an
d
test
s
ets
in
ter
m
s
o
f
g
en
er
aliza
tio
n
is
ch
o
s
en
as th
e
f
in
al
m
o
d
el
to
b
e
u
s
ed
.
2
.
6
.
E
v
a
lua
t
i
o
n pa
ra
m
e
t
er
s
T
h
e
f
in
al
m
o
d
el
h
as
b
ee
n
ev
a
lu
ated
u
s
in
g
m
u
ltip
le
m
etr
ics.
Acc
u
r
ac
y
,
Pre
cisi
o
n
,
R
ec
all,
F1
-
s
co
r
e,
an
d
r
ec
eiv
e
r
o
p
er
atin
g
ch
ar
ac
t
er
is
tic
-
ar
ea
u
n
d
er
th
e
c
u
r
v
e
(
R
OC
-
AUC
)
ar
e
th
e
q
u
an
titativ
e
m
etr
ics
u
s
ed
.
T
h
e
C
o
n
f
u
s
io
n
Ma
tr
ix
s
h
o
ws
h
o
w
m
an
y
tr
u
e
an
d
f
alse
p
o
s
itiv
es
an
d
n
e
g
ativ
es
th
er
e
a
r
e
f
o
r
ea
ch
class
is
u
s
ed
f
o
r
v
is
u
aliza
tio
n
an
d
th
e
R
OC
cu
r
v
e
is
u
s
ed
to
r
ep
r
esen
t th
e
m
o
d
els ca
p
ac
ity
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
tu
d
y
in
v
esti
g
ates
th
e
ef
f
ec
tiv
en
ess
o
f
an
a
u
to
m
ated
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
–
in
s
p
ir
ed
o
p
tim
izat
io
n
f
r
a
m
ewo
r
k
f
o
r
b
r
ain
tu
m
o
r
d
etec
tio
n
f
r
o
m
MRI
im
ag
es.
W
h
ile
p
r
ev
io
u
s
s
tu
d
ies
h
av
e
ex
ten
s
iv
ely
ex
p
l
o
r
ed
d
ee
p
le
ar
n
in
g
–
b
ased
ap
p
r
o
ac
h
es
u
s
in
g
p
r
etr
ain
ed
C
NN
s
s
u
ch
as
VGG1
6
,
R
esNet5
0
,
Den
s
eNe
t1
2
1
,
an
d
E
f
f
icie
n
tNet,
m
o
s
t
ex
is
tin
g
m
eth
o
d
s
r
ely
o
n
m
an
u
ally
d
e
s
ig
n
ed
class
if
ier
ar
ch
itectu
r
es
a
n
d
h
eu
r
is
tically
tu
n
ed
h
y
p
er
p
ar
a
m
eter
s
.
I
n
co
n
tr
ast,
th
is
wo
r
k
d
ir
ec
tly
ad
d
r
ess
es
th
is
g
ap
b
y
o
p
tim
izin
g
b
o
th
th
e
ar
ch
itectu
r
e
o
f
th
e
class
if
ier
a
n
d
th
e
tr
ain
i
n
g
h
y
p
er
p
ar
a
m
eter
s
in
a
d
ata
-
d
r
iv
en
way
.
T
h
is
m
ak
es
it
p
o
s
s
ib
le
to
s
y
s
tem
atica
lly
test
p
er
f
o
r
m
an
c
e,
r
o
b
u
s
tn
ess
,
an
d
g
en
e
r
aliza
tio
n
in
s
itu
atio
n
s
th
at
ar
e
r
ea
lis
tic.
3
.
1
.
Reinf
o
rc
e
m
ent
lea
rning
-
ins
pire
d c
la
s
s
if
ier
o
ptim
iza
t
io
n
W
e
u
s
ed
Op
tu
n
a,
a
cu
ttin
g
-
e
d
g
e
h
y
p
er
p
a
r
am
eter
o
p
tim
iza
tio
n
f
r
am
ewo
r
k
b
ased
o
n
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
,
to
m
ak
e
th
e
ar
c
h
itectu
r
al
d
esig
n
an
d
h
y
p
e
r
p
ar
a
m
eter
tu
n
in
g
o
f
th
e
b
r
ain
tu
m
o
r
d
etec
tio
n
m
o
d
e
l
ea
s
ier
.
T
h
e
s
ea
r
ch
s
p
ac
e
in
clu
d
ed
im
p
o
r
tan
t
d
esig
n
p
ar
a
m
eter
s
lik
e
th
e
n
u
m
b
er
o
f
d
e
n
s
e
lay
er
s
,
ac
tiv
atio
n
f
u
n
ctio
n
s
,
an
d
d
r
o
p
o
u
t
r
ates.
T
h
e
o
p
tim
izatio
n
p
r
o
ce
s
s
u
s
e
d
v
alid
atio
n
ac
cu
r
ac
y
as
th
e
o
b
jectiv
e
f
u
n
ctio
n
t
o
ju
d
g
e
ea
c
h
t
r
ial
co
n
f
ig
u
r
ati
o
n
.
As
s
h
o
wn
in
T
ab
le
2
,
O
p
tu
n
a
was
a
b
le
to
f
in
d
h
ig
h
-
p
er
f
o
r
m
in
g
m
o
d
el
co
n
f
ig
u
r
atio
n
s
b
y
ex
p
lo
r
in
g
a
n
d
cu
ttin
g
o
u
t
tr
ials
th
at
wer
e
n
’
t
wo
r
k
in
g
well.
T
h
e
b
est
class
if
ier
d
esig
n
h
a
d
th
r
ee
f
u
lly
c
o
n
n
ec
te
d
lay
er
s
w
ith
R
eL
U
ac
tiv
atio
n
,
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
3
,
a
n
d
a
b
atch
s
ize
o
f
3
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Dee
p
r
ein
fo
r
ce
men
t le
a
r
n
in
g
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
fr
a
mewo
r
k
u
s
in
g
o
p
tu
n
a
fo
r
…
(
A
a
s
h
u
t
o
s
h
K
h
a
r
b
)
1357
T
ab
le
2
.
Sam
p
le
Op
t
u
n
a
tr
ials
an
d
co
r
r
esp
o
n
d
in
g
v
alid
atio
n
p
er
f
o
r
m
an
ce
Tr
i
a
l
0
1
2
3
5
6
11
12
15
16
29
S
t
a
t
u
s
o
f
t
r
i
a
l
F
i
n
i
sh
e
d
F
i
n
i
sh
e
d
F
i
n
i
sh
e
d
F
i
n
i
sh
e
d
P
r
u
n
e
d
P
r
u
n
e
d
F
i
n
i
sh
e
d
F
i
n
i
sh
e
d
P
r
u
n
e
d
P
r
u
n
e
d
F
i
n
i
sh
e
d
O
b
j
e
c
t
i
v
e
f
u
n
c
t
i
o
n
0
.
6
9
0
.
6
8
0
.
7
0
0
.
6
4
P
r
u
n
e
d
P
r
u
n
e
d
0
.
6
9
0
.
7
0
P
r
u
n
e
d
P
r
u
n
e
d
0
.
9
5
N
u
mb
e
r
o
f
l
a
y
e
r
s
1
1
3
3
e
p
o
c
h
0
e
p
o
c
h
0
2
2
e
p
o
c
h
0
e
p
o
c
h
0
3
U
n
i
t
s
1
0
5
1
6
6
1
9
1
2
2
9
70
66
1
2
8
A
c
t
i
v
a
t
i
o
n
F
u
n
c
t
i
o
n
_
h
i
d
d
e
n
l
a
y
e
r
S
e
l
u
ELU
S
e
l
u
R
e
LU
S
e
l
u
S
e
l
u
R
e
LU
D
r
o
p
o
u
t
0
.
4
0
.
4
0
.
5
0
.
4
0
.
3
0
.
3
0
.
3
Le
a
r
n
i
n
g
r
a
t
e
1
.
0
9
E
-
05
8
.
5
9
E
-
04
3
.
1
1
E
-
04
8
.
5
9
E
-
04
1
.
3
9
E
-
04
2
.
2
6
E
-
04
5
.
0
0
E
-
04
B
a
t
c
h
si
z
e
64
64
32
32
32
32
32
O
p
t
i
mi
z
e
r
R
M
S
p
r
o
p
S
G
D
A
d
a
m
S
G
D
R
M
S
p
r
o
p
R
M
S
p
r
o
p
A
d
a
m
Ep
o
c
h
s
10
12
12
12
9
8
10
B
e
st
t
r
i
a
l
0
0
2
2
10
10
29
O
b
j
e
c
t
i
v
e
F
u
n
c
t
i
o
n
_
b
e
st
0
.
6
9
0
.
6
9
0
.
7
0
0
.
7
0
0
.
7
0
0
.
7
0
0
.
9
5
3
.
2
.
H
y
perpa
ra
m
e
t
er
t
un
ing
t
hro
ug
h
re
info
rc
em
ent
lea
r
nin
g
I
n
ad
d
itio
n
to
ar
ch
itectu
r
al
t
u
n
in
g
,
th
e
r
ei
n
f
o
r
ce
m
en
t
lea
r
n
in
g
f
r
am
ewo
r
k
was
also
t
ask
ed
wi
th
ad
ju
s
tin
g
k
e
y
tr
ain
i
n
g
h
y
p
e
r
p
ar
am
eter
s
s
u
ch
as
lear
n
in
g
r
a
te,
n
u
m
b
er
o
f
ep
o
c
h
s
,
b
atch
s
ize
an
d
o
p
tim
izer
ch
o
ice.
B
y
ass
ig
n
in
g
h
i
g
h
er
r
ewa
r
d
s
to
p
ar
am
eter
s
ets
th
at
led
to
f
aster
co
n
v
er
g
en
ce
an
d
h
ig
h
e
r
v
alid
atio
n
ac
cu
r
ac
y
,
th
e
ag
e
n
t
lear
n
e
d
t
o
f
av
o
r
a
lear
n
in
g
r
ate
o
f
5
e
-
4
,
1
0
tr
ain
in
g
ep
o
ch
s
,
a
n
d
th
e
u
s
e
o
f
th
e
Ad
a
m
o
p
tim
izer
.
T
h
is
d
y
n
am
ic
s
ea
r
ch
allo
wed
th
e
m
o
d
el
to
a
v
o
id
s
u
b
o
p
tim
al
co
n
f
ig
u
r
atio
n
s
th
at
ar
e
o
f
ten
en
co
u
n
ter
e
d
in
g
r
id
o
r
r
a
n
d
o
m
s
ea
r
ch
ap
p
r
o
ac
h
es.
Ov
er
t
h
e
co
u
r
s
e
o
f
3
0
ep
is
o
d
es,
th
e
ag
en
t
im
p
r
o
v
ed
v
alid
atio
n
ac
cu
r
ac
y
b
y
n
ea
r
l
y
2
5
%,
as f
ewe
r
ep
o
c
h
s
wer
e
r
e
q
u
ir
ed
to
r
ea
ch
p
ea
k
p
er
f
o
r
m
a
n
ce
.
3
.
3
.
E
v
a
lua
t
i
o
n
a
nd
perf
o
r
m
a
nce
T
h
e
tr
a
in
in
g
an
d
v
alid
atio
n
ac
cu
r
ac
y
an
d
l
o
s
s
p
lo
ts
in
Fig
u
r
e
2
an
d
Fig
u
r
e
3
s
h
o
w
s
m
o
o
t
h
an
d
s
tab
le
co
n
v
er
g
en
ce
,
with
n
o
s
ig
n
s
o
f
o
v
er
f
itti
n
g
,
r
ef
lectin
g
th
e
s
tab
ilit
y
o
f
th
e
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
-
tu
n
e
d
co
n
f
ig
u
r
atio
n
.
T
h
e
m
o
d
el
r
e
ac
h
ed
h
ig
h
ac
c
u
r
ac
y
with
in
th
e
f
ir
s
t
f
ew
ep
o
ch
s
,
an
d
co
n
tin
u
ed
to
im
p
r
o
v
e
g
r
ad
u
ally
.
Fig
u
r
e
4
s
h
o
ws
th
at
th
e
f
in
al
m
o
d
el,
tr
ain
e
d
with
t
h
e
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
-
o
p
tim
ized
p
ar
a
m
eter
s
,
ac
h
iev
es
an
ac
cu
r
ac
y
o
f
9
2
%,
an
AUC
s
co
r
e
o
f
0
.
9
3
,
an
d
an
F1
-
s
co
r
e
o
f
9
2
%
as
s
h
o
wn
in
T
a
b
le
3
.
T
h
e
d
r
o
p
o
u
t
lay
er
co
n
tr
i
b
u
ted
to
r
ed
u
cin
g
o
v
e
r
f
itti
n
g
,
esp
e
cially
wh
en
th
e
d
ataset
was
m
o
d
er
ately
s
m
all
in
s
ize.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
as
s
h
o
wn
in
Fig
u
r
e
5
in
d
icate
d
in
s
tr
o
n
g
p
er
f
o
r
m
an
ce
ac
r
o
s
s
b
o
th
tu
m
o
r
an
d
n
o
n
-
tu
m
o
r
class
es,
with
m
in
im
al
f
alse
p
o
s
itiv
es
an
d
n
eg
ativ
e
s
.
T
h
ese
f
in
d
in
g
s
v
alid
ate
th
e
ef
f
ec
tiv
en
ess
o
f
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
as
a
m
eta
-
co
n
tr
o
ller
f
o
r
o
p
tim
izi
n
g
d
ee
p
lear
n
i
n
g
p
ip
elin
es
in
m
ed
ical
im
a
g
in
g
ap
p
licatio
n
s
.
Fig
u
r
e
2
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
ac
cu
r
ac
y
p
lo
t
f
o
r
p
r
o
p
o
s
ed
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
3
5
2
-
1
3
6
3
1358
Fig
u
r
e
3
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
lo
s
s
Plo
ts
Fig
u
r
e
4
.
R
OC
cu
r
v
e
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
T
ab
le
3
.
C
lass
if
icatio
n
r
ep
o
r
t
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
n
t
h
e
B
r
aT
S 2
0
2
0
test
d
ataset
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
S
u
p
p
o
r
t
N
o
t
u
m
o
r
0
.
9
1
0
.
9
3
0
.
9
2
7
3
7
Tu
m
o
r
0
.
9
3
0
.
9
1
0
.
9
2
7
3
9
A
c
c
u
r
a
c
y
0
.
9
2
1
4
7
6
M
a
c
r
o
a
v
e
r
a
g
e
0
.
9
2
0
.
9
2
0
.
9
2
1
4
7
6
W
e
i
g
h
t
e
d
a
v
e
r
a
g
e
0
.
9
2
0
.
9
2
0
.
9
2
1
4
7
6
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
Ma
tr
ix
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Dee
p
r
ein
fo
r
ce
men
t le
a
r
n
in
g
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
fr
a
mewo
r
k
u
s
in
g
o
p
tu
n
a
fo
r
…
(
A
a
s
h
u
t
o
s
h
K
h
a
r
b
)
1359
3
.
4
.
Co
m
pa
riso
n wit
h ba
s
eline
m
o
del
T
o
ev
alu
ate
t
h
e
ef
f
ec
tiv
en
e
s
s
o
f
th
e
p
r
o
p
o
s
ed
d
ee
p
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
(
DR
L
)
-
in
s
p
ir
ed
o
p
tim
izatio
n
f
r
am
ewo
r
k
,
we
c
o
m
p
ar
ed
th
e
p
er
f
o
r
m
an
ce
o
f
t
h
e
Op
tu
n
a
-
o
p
tim
ized
E
f
f
icien
tNet
m
o
d
el
ag
ain
s
t
s
ev
er
al
wid
ely
u
s
ed
b
aselin
e
m
o
d
els
in
b
r
ain
tu
m
o
r
d
ete
ctio
n
.
T
h
ese
in
cl
u
d
e
class
ical
m
ac
h
in
e
lear
n
in
g
m
eth
o
d
s
,
s
tan
d
ar
d
d
ee
p
lear
n
i
n
g
ar
ch
itec
tu
r
es,
an
d
m
an
u
ally
tu
n
ed
v
a
r
ian
ts
o
f
E
f
f
icien
tN
et.
T
h
e
co
m
p
ar
ati
v
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
a
n
d
th
e
b
aselin
e
m
o
d
els is
s
u
m
m
ar
ized
in
T
ab
le
4
.
T
ab
le
4
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
with
b
aselin
e
m
o
d
els o
n
th
e
B
r
aT
S 2
0
2
0
d
ataset
M
o
d
e
l
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
R
O
C
-
AUC
S
V
M
(
R
B
F
K
e
r
n
e
l
)
[
1
9
]
8
1
.
2
0
%
0
.
8
0
.
8
2
0
.
8
1
0
.
8
4
C
a
p
su
l
e
n
e
t
w
o
r
k
[
2
0
]
8
5
.
7
0
%
0
.
8
4
0
.
8
6
0
.
8
5
0
.
8
8
V
G
G
1
6
[
2
1
]
8
4
.
7
%
0
.
8
3
0
.
8
4
0
.
8
3
0
.
8
7
M
o
b
i
l
e
N
e
t
V
2
[
2
2
]
8
6
.
1
%
0
.
8
5
0
.
8
6
0
.
8
5
0
.
8
8
R
e
sN
e
t
5
0
[
2
3
]
8
8
.
4
%
0
.
8
7
0
.
8
8
0
.
8
7
0
.
9
0
D
e
n
seN
e
t
1
2
1
[
2
4
]
8
9
.
6
%
0
.
8
9
0
.
8
9
0
.
8
9
0
.
9
1
Ef
f
i
c
i
e
n
t
N
e
t
(
ma
n
u
a
l
t
u
n
i
n
g
)
[
1
]
8
7
.
4
0
%
0
.
8
6
0
.
8
7
0
.
8
6
0
.
9
Ef
f
i
c
i
e
n
t
N
e
t
+
O
p
t
u
n
a
(
p
r
o
p
o
se
d
)
9
2
.
0
0
%
0
.
9
1
0
.
9
2
0
.
9
1
0
.
9
6
T
h
e
class
ical
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
with
an
R
B
F
Ker
n
el
[
1
9
]
s
h
o
wed
lim
ited
p
er
f
o
r
m
a
n
ce
d
u
e
to
its
in
ab
ilit
y
to
ca
p
tu
r
e
d
ee
p
s
p
atial
h
ier
ar
ch
ies
in
MRI
d
ata.
T
r
ad
itio
n
al
C
NN
m
o
d
els
[
2
0
]
,
alth
o
u
g
h
m
o
r
e
ex
p
r
ess
iv
e,
r
eq
u
ir
ed
m
a
n
u
al
ar
ch
itectu
r
e
d
esig
n
an
d
s
u
f
f
er
ed
f
r
o
m
s
u
b
o
p
t
im
al
h
y
p
e
r
p
ar
am
eter
s
ettin
g
s
.
T
h
e
p
r
et
r
ain
ed
R
esNet5
0
m
o
d
el
[
1
2
]
an
d
Den
s
eNe
t1
2
1
s
h
o
ws
h
ig
h
er
tr
ai
n
in
g
ac
cu
r
ac
y
b
u
t
h
av
e
in
co
n
s
is
ten
t
v
alid
atio
n
p
er
f
o
r
m
a
n
ce
.
Fu
r
th
er
,
Den
s
eNe
t1
2
1
r
eq
u
ir
es
m
o
r
e
co
m
p
u
tatio
n
tim
e
p
er
ep
o
c
h
th
an
E
f
f
icien
tNet.
T
h
e
m
ajo
r
r
ea
s
o
n
s
f
o
r
in
f
er
io
r
p
er
f
o
r
m
a
n
ce
o
f
o
th
e
r
m
o
d
el
s
ar
e
as
f
o
llo
ws.
Firstl
y
,
th
e
m
o
d
els
lik
e
VGG1
6
an
d
Mo
b
ileNetV2
h
av
e
lim
ited
f
ea
tu
r
e
r
e
p
r
esen
tatio
n
la
y
er
s
d
u
e
to
wh
ic
h
th
e
y
ar
e
less
e
f
f
ec
tiv
e
in
ca
p
tu
r
in
g
th
e
tu
m
o
r
b
o
u
n
d
ar
y
.
Ma
n
u
al
tu
n
in
g
o
f
E
f
f
icien
tNet
p
er
f
o
r
m
ed
s
lig
h
tly
b
etter
,
b
u
t
lac
k
ed
th
e
s
y
s
tem
atic
o
p
tim
izatio
n
th
at
Op
tu
n
a
p
r
o
v
id
es.
I
n
co
n
tr
ast,
t
h
e
p
r
o
p
o
s
ed
f
r
am
ew
o
r
k
,
wh
ich
in
teg
r
at
es
Op
tu
n
a
in
to
th
e
ar
ch
itectu
r
e
an
d
h
y
p
er
p
ar
a
m
eter
s
elec
tio
n
p
r
o
ce
s
s
,
o
u
tp
e
r
f
o
r
m
e
d
all
b
aselin
es
ac
r
o
s
s
ev
er
y
m
etr
ic.
T
h
is
clea
r
ly
d
em
o
n
s
tr
ates
th
e
ef
f
ec
tiv
en
ess
o
f
u
s
in
g
an
au
to
m
ate
d
,
DR
L
-
in
s
p
ir
ed
o
p
tim
izatio
n
ap
p
r
o
ac
h
to
r
e
d
u
ce
h
u
m
an
e
f
f
o
r
t a
n
d
im
p
r
o
v
e
d
ia
g
n
o
s
tic
ac
cu
r
ac
y
in
m
ed
ical
i
m
ag
in
g
.
Fu
r
th
er
,
th
e
s
tu
d
ies
in
[
3
]
,
[
5
]
,
[
7
]
,
[
9
]
,
[
1
2
]
h
av
e
s
h
o
wn
h
ig
h
er
p
er
f
o
r
m
an
ce
with
ac
cu
r
ac
y
r
an
g
in
g
f
r
o
m
9
8
.
5
to
9
9
.
3
%.
Ho
wev
er
th
ese
s
tu
d
ies
ar
e
b
ased
o
n
th
e
d
ataset
f
r
o
m
Ka
g
g
le
o
r
co
n
tain
cu
r
ated
o
r
b
alan
ce
d
d
ata
s
am
p
les
o
r
u
s
in
g
d
ataset
-
s
p
ec
if
ic
f
in
e
tu
n
in
g
,
th
u
s
r
esu
ltin
g
in
h
ig
h
er
ac
cu
r
a
cies.
B
u
t
o
u
r
wo
r
k
u
s
es
th
e
B
r
aT
S
2
0
2
0
d
a
taset,
wh
ich
is
m
o
r
e
c
o
m
p
lex
,
m
u
lti
-
m
o
d
al
a
n
d
h
eter
o
g
e
n
eo
u
s
.
T
h
er
ef
o
r
e
,
ac
h
iev
i
n
g
h
ig
h
er
ac
c
u
r
ac
ies is
m
o
r
e
ch
al
len
g
in
g
a
n
d
r
esu
lts
ar
e
ty
p
ically
lo
wer
b
u
t m
o
r
e
r
ea
lis
tic
.
3
.
5
.
Cro
s
s
da
t
a
s
et
e
v
a
lua
t
io
n
I
n
o
r
d
er
to
s
h
o
w
th
at
th
e
p
r
o
p
o
s
ed
o
p
tim
izatio
n
f
r
am
ew
o
r
k
is
r
o
b
u
s
t
an
d
g
en
er
alize
,
th
is
wo
r
k
f
u
r
th
er
ev
alu
ates
th
e
m
o
d
el
o
n
two
m
o
r
e
wid
ely
u
s
ed
b
r
ain
t
u
m
o
r
d
atasets
alo
n
g
with
B
r
a
T
S
2
0
2
0
.
T
h
ese
ar
e
Fig
s
h
ar
e
Data
s
et
th
at
h
as
im
a
g
es
class
if
ied
in
to
th
r
ee
c
ateg
o
r
ies,
th
at
is
,
g
lio
m
a,
m
en
in
g
io
m
a
an
d
p
itu
itar
y
an
d
Kag
g
le
B
r
ain
MRI
d
ataset.
B
o
th
th
ese
d
atasets
v
ar
y
f
r
o
m
B
r
aT
S
d
ataset
in
ter
m
s
o
f
ac
q
u
is
itio
n
,
class
d
is
tr
ib
u
tio
n
,
s
lice
q
u
ality
etc.
T
h
is
allo
ws
to
test
th
e
v
er
s
ati
lity
o
f
th
e
p
r
o
p
o
s
ed
wo
r
k
.
T
a
b
le
5
an
d
T
a
b
le
6
r
ep
r
esen
ts
th
e
v
alu
e
o
f
t
h
e
o
p
tim
ized
h
y
p
er
p
ar
am
eter
s
o
b
tain
ed
u
s
in
g
Op
tu
n
a
f
r
am
ewo
r
k
an
d
th
e
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
ac
r
o
s
s
d
if
f
er
en
t d
atasets
.
T
ab
le
5
.
Op
tim
ized
h
y
p
e
r
p
ar
a
m
eter
s
o
b
tain
ed
f
o
r
Fig
s
h
ar
e
a
n
d
Kag
g
le
d
atase
ts
u
s
in
g
Op
tu
n
a
H
y
p
e
r
p
a
r
a
me
t
e
r
s
F
i
g
s
h
a
r
e
d
a
t
a
se
t
K
a
g
g
l
e
d
a
t
a
se
t
Tr
i
a
l
35
24
S
t
a
t
u
s
o
f
t
r
i
a
l
F
i
n
i
sh
e
d
F
i
n
i
sh
e
d
O
b
j
e
c
t
i
v
e
f
u
n
c
t
i
o
n
0
.
9
7
0
.
9
8
N
u
mb
e
r
o
f
l
a
y
e
r
s
5
7
U
n
i
t
s
2
5
6
1
2
8
A
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
_
h
i
d
d
e
n
l
a
y
e
r
R
e
LU
R
e
LU
D
r
o
p
o
u
t
0
.
4
0
.
2
Le
a
r
n
i
n
g
r
a
t
e
1
.
0
9
E
-
05
2
.
0
1
E
-
05
B
a
t
c
h
si
z
e
32
32
O
p
t
i
mi
z
e
r
A
d
a
m
A
d
a
m
Ep
o
c
h
s
40
37
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
1
3
5
2
-
1
3
6
3
1360
T
ab
le
6
.
C
r
o
s
s
-
d
ataset
p
er
f
o
r
m
an
ce
o
f
t
h
e
p
r
o
p
o
s
ed
f
r
am
e
wo
r
k
D
a
t
a
s
e
t
Ta
sk
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
AUC
B
r
a
TS
D
e
t
e
c
t
i
o
n
9
2
%
0
.
9
1
0
.
9
2
0
.
9
2
0
.
9
6
F
i
g
s
h
a
r
e
3
-
c
l
a
ss
c
l
a
ssi
f
i
c
a
t
i
o
n
9
5
.
4
%
0
.
9
5
0
.
9
5
0
.
9
5
0
.
9
7
K
a
g
g
l
e
D
e
t
e
c
t
i
o
n
9
8
.
1
%
0
.
9
8
0
.
9
8
0
.
9
8
0
.
9
9
I
t
ca
n
b
e
o
b
s
er
v
ed
f
r
o
m
th
e
r
esu
lts
th
at
th
e
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
g
en
e
r
alize
s
well
ac
r
o
s
s
m
u
ltip
le
d
atasets
with
v
ar
y
in
g
co
m
p
le
x
ity
an
d
class
s
tr
u
ctu
r
es.
T
h
e
B
r
aT
S
2
0
2
0
d
ataset
s
h
o
ws
th
e
r
ea
lis
tic
ac
cu
r
ac
y
o
f
9
2
%
as
it
p
r
esen
ts
h
eter
o
g
en
eity
an
d
n
o
is
e.
On
th
e
o
th
er
h
an
d
Fig
s
h
ar
e
d
ataset
is
m
o
r
e
b
alan
ce
d
an
d
clea
r
ac
h
iev
es
an
ac
cu
r
a
cy
o
f
9
5
.
4
%
an
d
Kag
g
le
d
ataset
ac
h
iev
es
an
ac
cu
r
ac
y
o
f
9
8
.
1
%
s
h
o
ws
th
at
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
ca
n
ac
h
ie
v
e
s
tate
o
f
th
e
ar
t p
e
r
f
o
r
m
an
ce
wh
e
n
th
e
d
ataset
d
if
f
icu
lty
is
lo
wer
.
3
.
6
.
E
v
a
lua
t
i
o
n wit
h
im
ba
la
nced
da
t
a
s
et
Alth
o
u
g
h
th
e
p
r
im
ar
y
ex
p
e
r
im
en
ts
wer
e
co
n
d
u
cted
o
n
a
b
alan
ce
d
d
ataset,
r
ea
l
-
wo
r
ld
m
ed
ical
im
ag
in
g
d
atasets
ar
e
o
f
ten
in
h
er
en
tly
im
b
alan
ce
d
,
with
f
ewe
r
p
ath
o
lo
g
ical
s
am
p
les
th
an
n
o
r
m
al
ca
s
es.
T
o
ass
ess
th
e
r
o
b
u
s
tn
ess
o
f
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
u
n
d
e
r
s
u
ch
co
n
d
itio
n
s
,
ad
d
itio
n
al
ex
p
er
im
en
ts
wer
e
p
er
f
o
r
m
ed
u
s
in
g
ar
tific
ially
i
m
b
alan
ce
d
v
er
s
io
n
s
o
f
t
h
e
d
at
aset.
C
las
s
im
b
alan
ce
was
in
tr
o
d
u
ce
d
b
y
r
ed
u
ci
n
g
th
e
n
u
m
b
er
o
f
s
am
p
les
in
o
n
e
class
,
r
esu
ltin
g
in
im
b
alan
ce
r
atio
s
o
f
7
0
:3
0
a
n
d
8
0
:2
0
.
T
h
e
p
r
o
p
o
s
ed
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
-
o
p
tim
ized
E
f
f
ic
ien
tNet
m
o
d
el
was
ev
alu
ated
u
s
in
g
im
b
alan
ce
-
s
en
s
itiv
e
m
etr
ics,
in
clu
d
in
g
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e
,
an
d
R
OC
–
AU
C
,
with
p
ar
ticu
lar
e
m
p
h
asis
o
n
m
i
n
o
r
i
ty
-
class
(
tu
m
o
r
)
p
er
f
o
r
m
an
ce
.
I
t
ca
n
b
e
o
b
s
er
v
ed
f
r
o
m
T
a
b
le
7
,
th
at
as
th
e
class
im
b
alan
ce
in
cr
ea
s
es
th
e
ac
cu
r
ac
y
h
as
b
ee
n
d
ec
r
ea
s
ed
a
litt
l
e
b
u
t
th
e
r
e
ca
ll
an
d
F1
-
s
co
r
e
r
em
ain
s
s
tab
le.
T
h
is
p
er
f
o
r
m
an
ce
ca
n
b
e
attr
ib
u
ted
to
th
e
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
–
i
n
s
p
ir
ed
Op
tu
n
a
o
p
tim
izatio
n
,
wh
ic
h
d
y
n
am
ically
a
d
ap
ts
class
if
ier
ar
ch
itectu
r
e
an
d
tr
ain
in
g
h
y
p
e
r
p
ar
am
eter
s
.
T
ab
le
7
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
an
d
b
aselin
e
m
o
d
el
u
n
d
e
r
d
i
f
f
er
en
t
class
im
b
alan
ce
r
atio
s
(
tu
m
o
r
:
non
-
tu
m
o
r
)
I
mb
a
l
a
n
c
e
r
a
t
i
o
M
o
d
e
l
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
R
O
C
–
AUC
5
0
:
5
0
Ef
f
i
c
i
e
n
t
N
e
t
(
ma
n
u
a
l
t
u
n
i
n
g
)
8
7
.
4
%
0
.
8
6
0
.
8
7
0
.
8
7
0
.
9
0
P
r
o
p
o
se
d
Ef
f
i
c
i
e
n
t
N
e
t
+
O
p
t
u
n
a
9
2
%
0
.
9
1
0
.
9
2
0
.
9
2
0
.
9
6
7
0
:
3
0
Ef
f
i
c
i
e
n
t
N
e
t
(
ma
n
u
a
l
t
u
n
i
n
g
)
8
4
.
6
%
0
.
8
3
0
.
8
2
0
.
8
3
0
.
8
7
P
r
o
p
o
se
d
Ef
f
i
c
i
e
n
t
N
e
t
+
O
p
t
u
n
a
9
0
.
3
%
0
.
8
9
0
.
9
0
0
.
8
9
0
.
9
4
8
0
:
2
0
Ef
f
i
c
i
e
n
t
N
e
t
(
ma
n
u
a
l
t
u
n
i
n
g
)
8
1
.
2
%
0
.
7
9
0
.
7
8
0
.
7
8
0
.
8
4
P
r
o
p
o
se
d
Ef
f
i
c
i
e
n
t
N
e
t
+
O
p
t
u
n
a
8
8
.
1
%
0
.
8
7
0
.
8
8
0
.
8
7
0
.
9
2
3
.
7
.
Dis
cus
s
io
n o
n
perf
o
rm
a
nce
g
a
ins
a
nd
ex
pla
ina
bil
it
y
persp
ec
t
iv
e
T
h
e
p
r
o
p
o
s
ed
o
p
tim
izatio
n
f
r
am
ewo
r
k
,
wh
ich
is
b
ased
o
n
E
f
f
icien
tNet
an
d
in
s
p
ir
ed
b
y
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
s
h
o
ws
b
etter
p
er
f
o
r
m
a
n
ce
.
T
h
is
is
d
u
e
to
th
e
co
m
b
in
ed
ef
f
ec
t
o
f
th
e
ef
f
ec
tiv
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
,
a
u
to
m
atic
ar
ch
itectu
r
e
ad
ap
tatio
n
,
an
d
h
y
p
e
r
p
ar
am
eter
tu
n
in
g
.
Firstl
y
,
th
e
E
f
f
icien
tNet
b
ac
k
b
o
n
e
m
a
k
es
s
u
r
e
th
at
th
e
n
etwo
r
k
’
s
d
e
p
th
,
wid
th
,
a
n
d
r
eso
lu
ti
o
n
ar
e
all
in
th
e
r
ig
h
t
p
lace
.
T
h
is
p
r
o
v
id
es
d
etailed
s
p
atial
a
n
d
i
n
ten
s
ity
-
b
ased
f
ea
tu
r
es
f
r
o
m
b
r
ain
MRI
s
ca
n
s
wh
ich
is
v
er
y
im
p
o
r
tan
t
f
o
r
th
e
B
r
aT
S
2
0
2
0
d
ataset
b
ec
au
s
e
th
e
e
d
g
es
o
f
tu
m
o
r
r
eg
io
n
s
ar
e
o
f
te
n
n
o
t
clea
r
an
d
th
ey
l
o
o
k
d
if
f
e
r
en
t
f
r
o
m
ea
ch
o
th
er
.
I
n
co
m
p
ar
is
o
n
to
th
e
o
th
er
d
ee
p
lear
n
i
n
g
m
o
d
els,
E
f
f
icien
tNet
ca
n
f
in
d
p
atter
n
s
th
at
ar
e
s
p
ec
if
ic
to
t
u
m
o
r
s
wh
ile
u
s
in
g
less
co
m
p
u
ter
p
o
wer
.
T
h
e
VGG1
6
an
d
Mo
b
ileNetV2
h
av
e
lim
ited
f
ea
tu
r
e
r
ep
r
esen
tatio
n
ca
p
ab
ilit
ies.
B
u
t
d
ee
p
er
ar
ch
it
ec
tu
r
es
lik
e
R
esNet5
0
an
d
Den
s
eNe
t1
2
1
u
s
u
ally
n
ee
d
m
o
r
e
p
r
o
ce
s
s
in
g
p
o
we
r
a
n
d
d
o
n
’
t
alwa
y
s
wo
r
k
well
o
n
co
m
p
licated
m
e
d
ical
d
atasets
,
wh
ich
lim
its
th
eir
p
r
ac
tical
ef
ec
tiv
en
ess
.
Seco
n
d
ly
,
th
e
Op
tu
n
a
f
r
am
e
wo
r
k
wh
ich
is
in
s
p
ir
ed
b
y
r
e
in
f
o
r
ce
m
e
n
t
lear
n
in
g
p
lay
s
a
n
im
p
o
r
tan
t
r
o
le
in
th
e
p
e
r
f
o
r
m
an
ce
en
h
an
ce
m
en
t
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el,
as
it
au
to
m
atica
l
ly
o
p
ti
m
izes
th
e
tr
ain
in
g
h
y
p
er
p
ar
am
eter
s
an
d
th
e
class
if
ier
ar
ch
itectu
r
e.
T
h
e
s
u
g
g
e
s
ted
f
r
am
ewo
r
k
d
y
n
am
ically
s
ea
r
ch
es
th
e
s
ea
r
ch
s
p
ac
e
f
o
r
t
h
e
b
est
co
n
f
ig
u
r
atio
n
s
f
o
r
t
h
e
task
,
wh
ile
m
an
u
all
y
tu
n
e
d
m
o
d
els
r
ely
o
n
f
ix
ed
d
esig
n
ch
o
ices.
T
h
e
s
m
o
o
th
tr
ain
in
g
cu
r
v
es
an
d
b
alan
ce
d
p
r
ec
is
io
n
-
r
ec
all
v
alu
es
s
h
o
w
th
at
th
e
m
o
d
el
ca
n
r
ea
ch
s
tab
le
co
n
v
er
g
en
ce
,
r
ed
u
ce
o
v
er
f
itti
n
g
,
an
d
i
m
p
r
o
v
e
g
e
n
er
aliza
tio
n
d
u
e
to
th
is
two
-
lev
el
o
p
tim
izat
io
n
.
T
h
ir
d
,
th
e
s
u
g
g
ested
m
eth
o
d
s
h
o
ws
s
tr
o
n
g
g
en
er
aliza
tio
n
o
n
clin
ically
im
p
o
r
tan
t
an
d
v
ar
ied
d
ata.
T
h
e
B
r
aT
S
2
0
2
0
d
ataset
h
a
s
a
lo
t
o
f
d
if
f
er
en
ce
s
in
im
ag
in
g
m
o
d
alities
,
tu
m
o
r
ty
p
es,
an
d
ac
q
u
is
itio
n
p
r
o
to
co
ls
.
T
h
is
m
a
k
es
it
h
ar
d
e
r
to
u
s
e
th
a
n
well
-
o
r
g
an
ize
d
o
r
b
alan
ce
d
d
atasets
lik
e
Kag
g
l
e
o
r
Fig
s
h
ar
e.
T
h
e
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
ac
h
iev
e
s
r
ea
lis
tic
an
d
d
e
p
en
d
a
b
le
p
er
f
o
r
m
a
n
ce
b
y
d
ir
ec
tly
o
p
ti
m
izin
g
th
e
lear
n
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Dee
p
r
ein
fo
r
ce
men
t le
a
r
n
in
g
i
n
s
p
ir
ed
o
p
timiz
a
tio
n
fr
a
mewo
r
k
u
s
in
g
o
p
tu
n
a
fo
r
…
(
A
a
s
h
u
t
o
s
h
K
h
a
r
b
)
1361
p
r
o
ce
s
s
o
n
th
is
in
tr
icate
d
is
tr
ib
u
tio
n
.
T
h
is
is
wh
y
th
e
m
o
d
el
s
th
at
wer
e
m
an
u
ally
t
u
n
ed
a
n
d
th
e
o
n
es
th
at
wer
e
s
et
u
p
as a
b
aselin
e
im
p
r
o
v
ed
.
C
o
n
s
id
er
in
g
th
e
in
te
r
p
r
etab
i
lit
y
p
er
s
p
ec
tiv
e,
ex
p
lain
a
b
le
ar
t
if
icial
in
telig
en
ce
(
XAI
)
m
et
h
o
d
s
s
h
o
w
wh
ich
p
ar
ts
o
f
an
im
a
g
e
ar
e
m
o
s
t
im
p
o
r
tan
t
f
o
r
th
e
m
o
d
el
’
s
p
r
ed
ictio
n
s
.
T
h
ese
tech
n
iq
u
e
s
ar
e
v
er
y
u
s
ef
u
l
in
m
ed
ical
im
ag
in
g
b
ec
a
u
s
e
th
e
y
m
ak
e
s
u
r
e
th
at
th
e
m
o
d
el
o
n
ly
lo
o
k
s
at
tu
m
o
r
ar
ea
s
th
at
ar
e
im
p
o
r
tan
t
to
d
o
cto
r
s
an
d
n
o
t
b
ac
k
g
r
o
u
n
d
n
o
is
e.
T
h
ey
also
h
ig
h
lig
h
t
th
e
s
p
ec
if
ic
ar
ea
s
wh
ich
co
n
tr
ib
u
te
d
to
th
e
r
esu
lts
an
d
h
elp
s
to
u
n
d
er
s
tan
d
th
e
r
esu
lts
m
o
r
e
ef
f
ec
tiv
ely
.
T
h
e
cu
r
r
en
t
s
tu
d
y
is
f
o
cu
s
in
g
o
n
th
e
a
u
to
m
ated
o
p
tim
izatio
n
f
o
r
b
r
ain
tu
m
o
r
d
etec
tio
n
.
T
h
e
in
teg
r
atio
n
o
f
XAI
-
b
ased
v
is
u
al
an
aly
s
is
,
o
n
th
e
o
th
er
h
an
d
,
is
a
b
ig
s
tep
f
o
r
war
d
f
o
r
th
e
f
u
tu
r
e.
Usi
n
g
th
ese
m
eth
o
d
s
wo
u
ld
m
ak
e
th
in
g
s
ev
en
clea
r
e
r
an
d
m
o
r
e
r
e
liab
le
in
th
e
clin
ic
b
y
lettin
g
y
o
u
c
o
m
p
ar
e
t
h
e
b
a
s
elin
e
m
o
d
els to
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
in
a
q
u
alitativ
e
way
.
4.
CO
NCLU
SI
O
N
AND
F
U
T
U
RE
SCO
P
E
I
n
th
is
wo
r
k
,
we
p
r
o
p
o
s
ed
a
n
au
to
m
ated
,
r
ein
f
o
r
ce
m
en
t
lea
r
n
in
g
–
i
n
s
p
ir
ed
o
p
tim
izatio
n
f
r
am
ewo
r
k
f
o
r
b
r
ain
tu
m
o
r
d
etec
tio
n
u
s
in
g
MRI
im
ag
es,
with
th
e
aim
o
f
r
ed
u
cin
g
t
h
e
r
elian
ce
o
n
m
an
u
al
m
o
d
el
d
esig
n
an
d
h
eu
r
is
tic
h
y
p
er
p
a
r
am
eter
tu
n
in
g
.
T
h
e
p
r
o
p
o
s
ed
o
p
tim
iz
atio
n
f
r
am
ewo
r
k
,
wh
ich
u
s
es
E
f
f
icien
tNetB
0
as
th
e
f
ea
tu
r
e
ex
tr
ac
to
r
with
th
e
Op
tu
n
a,
ca
n
au
to
m
atica
l
ly
co
m
p
u
te
b
o
th
th
e
ar
ch
itectu
r
e
o
f
class
if
ier
an
d
th
e
tr
ain
in
g
p
ar
a
m
eter
s
.
T
h
is
d
esig
n
h
elp
s
th
e
m
o
d
el
lear
n
b
ette
r
f
r
o
m
c
o
m
p
licated
m
ed
ical
i
m
ag
es
wh
ile
cu
ttin
g
d
o
wn
o
n
th
e
n
ee
d
f
o
r
a
lo
t
o
f
e
x
p
er
t h
elp
.
T
h
e
r
esu
lts
s
h
o
w
th
at
t
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
o
u
t
p
er
f
o
r
m
s
b
o
th
th
e
b
aselin
e
an
d
m
an
u
ally
tu
n
e
d
m
o
d
els.
T
h
e
f
r
am
ew
o
r
k
h
as
a
h
ig
h
class
if
icatio
n
ac
cu
r
ac
y
an
d
wo
r
k
s
well
in
a
v
a
r
iety
o
f
test
in
g
s
itu
atio
n
s
.
I
m
p
o
r
ta
n
tly
,
wh
e
n
ev
al
u
ated
u
n
d
er
im
b
alan
ce
d
d
ata
co
n
d
itio
n
s
,
th
e
m
o
d
el
m
ain
tain
s
r
elia
b
le
d
etec
tio
n
o
f
t
h
e
tu
m
o
r
class
,
wh
ich
is
cr
itical
in
clin
ical
ap
p
licatio
n
s
wh
er
e
m
is
s
in
g
a
p
o
s
itiv
e
ca
s
e
ca
n
h
av
e
s
er
io
u
s
co
n
s
eq
u
en
ce
s
.
Fu
r
t
h
er
,
th
e
c
r
o
s
s
-
d
ataset
ev
alu
atio
n
s
also
s
h
o
w
th
at
th
e
p
r
o
p
o
s
ed
m
eth
o
d
im
p
r
o
v
es
g
en
er
aliza
tio
n
,
t
h
u
s
m
ak
in
g
it
less
s
en
s
itiv
e
to
th
e
d
ataset
-
s
p
ec
if
ic
ch
ar
ac
ter
is
tics
.
T
h
e
p
r
im
ar
y
c
o
n
tr
ib
u
tio
n
o
f
th
is
s
tu
d
y
is
d
em
o
n
s
tr
atin
g
th
at
th
e
s
im
u
ltan
eo
u
s
o
p
tim
izatio
n
o
f
class
if
ier
s
tr
u
ctu
r
e
an
d
tr
ain
i
n
g
h
y
p
er
p
ar
am
eter
s
in
an
au
t
o
m
ated
f
ash
io
n
ca
n
en
h
an
ce
th
e
r
o
b
u
s
tn
ess
an
d
r
eliab
ilit
y
o
f
b
r
ain
tu
m
o
r
d
e
tectio
n
s
y
s
tem
s
.
Un
lik
e
tr
ad
itio
n
al
tr
an
s
f
er
lear
n
in
g
m
eth
o
d
s
th
at
u
s
e
f
ix
ed
ar
ch
itectu
r
es,
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
c
h
an
g
es
b
ased
o
n
t
h
e
ch
ar
ac
te
r
is
tics
o
f
th
e
d
ata.
T
h
is
m
ak
es
it
w
o
r
k
b
etter
an
d
m
o
r
e
r
eliab
ly
in
to
u
g
h
s
itu
atio
n
s
.
T
h
ese
r
esu
lts
s
h
o
w
h
o
w
u
s
ef
u
l
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
-
in
s
p
ir
ed
o
p
tim
izatio
n
ca
n
b
e
f
o
r
an
aly
z
in
g
m
ed
ical
im
ag
es.
As
a
f
u
tu
r
e
s
co
p
e
th
is
wo
r
k
ca
n
b
e
ex
ten
d
ed
with
o
t
h
er
b
ac
k
b
o
n
e
a
r
ch
itectu
r
es
l
ik
e
d
ee
p
e
r
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
s
an
d
v
is
io
n
tr
an
s
f
o
r
m
er
s
.
T
h
e
ex
p
lai
n
ab
le
AI
lik
e
Gr
ad
-
C
am
,
L
I
ME
etc
ca
n
h
elp
m
a
k
e
m
o
d
el
p
r
e
d
ictio
n
s
m
o
r
e
u
n
d
er
s
tan
d
ab
le
b
y
g
iv
in
g
v
is
u
al
in
f
o
r
m
atio
n
a
b
o
u
t th
e
m
.
Mo
r
e
a
d
v
an
ce
d
m
et
h
o
d
s
f
o
r
d
ea
lin
g
with
s
ev
er
e
class
im
b
alan
ce
,
lik
e
co
s
t
-
s
en
s
itiv
e
lea
r
n
in
g
an
d
ad
a
p
tiv
e
s
am
p
lin
g
,
m
ay
also
im
p
r
o
v
e
p
er
f
o
r
m
an
ce
.
F
in
ally
,
test
in
g
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
o
n
lar
g
e
clin
ical
d
atasets
f
r
o
m
m
u
ltip
le
in
s
titu
tio
n
s
an
d
lo
o
k
in
g
in
to
h
o
w
it
co
u
l
d
b
e
u
s
ed
in
r
ea
l
tim
e
will
b
e
im
p
o
r
tan
t
s
tep
s
to
war
d
g
ettin
g
i
t
u
s
ed
in
r
ea
l
life
.
Ov
er
all,
th
is
s
tu
d
y
s
h
o
ws
th
at
au
to
m
ated
o
p
tim
izati
o
n
f
r
am
ewo
r
k
s
ar
e
a
p
r
o
m
is
in
g
way
t
o
m
ak
e
b
r
ain
tu
m
o
r
d
etec
tio
n
s
y
s
tem
s
th
at
ar
e
ac
c
u
r
ate,
s
tr
o
n
g
,
an
d
u
s
ef
u
l in
t
h
e
clin
ic.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
r
ec
eiv
ed
n
o
s
p
ec
if
ic
g
r
an
t f
r
o
m
a
n
y
f
u
n
d
in
g
ag
en
cy
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Aash
u
to
s
h
Kh
ar
b
✓
✓
✓
✓
✓
✓
✓
✓
✓
Pra
ch
i Ch
au
d
h
ar
y
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.