I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
,
p
p
.
4
9
4
3
~
4
9
5
6
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
6
.
p
p
4
9
4
3
-
4
9
5
6
4943
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
M
ela
no
ma
clas
sifica
tion usin
g
ense
mble deep
t
ra
ns
fe
r learning
So
um
y
a
G
a
da
g
1
,
P
a
nd
ura
ng
a
Ra
o
M
a
lo
de
Vis
hwa
na
t
ha
2
,
Virupa
x
i Ba
la
cha
nd
ra
Da
la
l
3
1
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
m
mu
n
i
c
a
t
i
o
n
,
Ja
i
n
C
o
l
l
e
g
e
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
R
e
se
a
r
c
h
,
K
a
r
n
a
t
a
k
a
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
En
g
i
n
e
e
r
i
n
g
,
F
ET
-
JA
I
N
D
e
e
me
d
t
o
b
e
U
n
i
v
e
r
s
i
t
y
,
B
a
n
g
a
l
o
r
e
,
I
n
d
i
a
3
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
m
mu
n
i
c
a
t
i
o
n
,
Ja
i
n
C
o
l
l
e
g
e
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
R
e
se
a
r
c
h
,
B
e
l
a
g
a
v
i
,
K
a
r
n
a
t
a
k
a
,
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
Ap
r
3
0
,
2
0
2
5
R
ev
is
ed
Au
g
1
1
,
2
0
2
5
Acc
ep
ted
Sep
7
,
2
0
2
5
M
e
lan
o
m
a
,
a
t
y
p
e
o
f
sk
i
n
c
a
n
c
e
r,
p
o
se
s
si
g
n
ifi
c
a
n
t
c
h
a
ll
e
n
g
e
s
in
e
a
rly
d
e
tec
ti
o
n
a
n
d
d
ia
g
n
o
sis.
S
e
v
e
ra
l
m
e
th
o
d
s
fo
r
e
a
rly
m
e
lan
o
m
a
d
e
tec
ti
o
n
,
in
c
lu
d
in
g
v
isu
a
l
i
n
sp
e
c
ti
o
n
a
n
d
se
v
e
ra
l
m
a
c
h
in
e
lea
rn
in
g
m
o
d
e
ls,
fa
c
e
c
h
a
ll
e
n
g
e
s
with
a
c
c
u
ra
c
y
.
To
o
v
e
rc
o
m
e
th
e
se
issu
e
s,
d
e
e
p
lea
rn
in
g
h
a
s
b
e
e
n
wid
e
ly
a
d
o
p
ted
in
v
a
rio
u
s
b
io
m
e
d
ica
l
a
p
p
li
c
a
ti
o
n
s
.
I
n
t
h
is
wo
rk
,
we
e
m
p
lo
y
d
e
e
p
tran
sfe
r
lea
rn
in
g
m
e
th
o
d
s
to
c
las
sify
m
e
lan
o
m
a
.
F
irstl
y
,
we
c
o
ll
e
c
t
p
u
b
li
c
ly
a
v
a
il
a
b
le
d
a
tas
e
ts
c
o
n
tai
n
in
g
m
e
lan
o
m
a
ima
g
e
s,
t
h
e
ir
c
o
rr
e
sp
o
n
d
in
g
g
ro
u
n
d
tr
u
th
fo
r
se
g
m
e
n
tatio
n
,
a
n
d
c
las
s
lab
e
ls.
S
u
b
se
q
u
e
n
tl
y
,
we
p
e
rfo
r
m
d
a
ta
p
re
p
r
o
c
e
ss
in
g
,
n
o
rm
a
li
z
a
ti
o
n
,
a
n
d
la
b
e
l
e
n
c
o
d
in
g
t
o
a
d
d
re
ss
issu
e
s
o
f
v
a
ried
il
l
u
m
in
a
ti
o
n
,
ima
g
e
n
o
is
e
,
a
n
d
d
a
ta
imb
a
lan
c
e
.
Ne
x
t,
w
e
c
o
n
d
u
c
t
fe
a
tu
re
e
x
trac
ti
o
n
u
ti
li
z
i
n
g
t
h
e
p
re
v
io
u
sly
train
e
d
d
e
e
p
lea
rn
i
n
g
m
o
d
e
ls,
VG
G
,
Re
sN
e
t,
In
c
e
p
ti
o
n
Re
sN
e
t,
a
n
d
M
o
b
i
leN
e
t.
Th
e
c
h
a
ra
c
teristic
v
e
c
to
rs
o
b
tai
n
e
d
fr
o
m
e
a
c
h
m
o
d
e
l
a
re
fu
se
d
to
p
ro
d
u
c
e
a
c
o
m
p
re
h
e
n
siv
e
d
e
p
ictio
n
a
m
o
n
g
th
e
p
ro
v
i
d
e
d
p
ictu
re
s.
In
th
e
c
las
sifica
ti
o
n
sta
g
e
,
we
e
m
p
lo
y
e
n
se
m
b
le
lea
rn
in
g
u
si
n
g
tran
sfe
r
lea
rn
in
g
m
o
d
e
ls,
i
n
c
lu
d
in
g
Ef
f
icie
n
tNe
t,
Xc
e
p
ti
o
n
,
a
n
d
De
n
se
Ne
t.
Th
e
se
m
o
d
e
ls
a
re
train
e
d
o
n
th
e
fin
a
l
fe
a
tu
re
v
e
c
to
r
to
c
las
sify
m
e
lan
o
m
a
ima
g
e
s
e
ffe
c
ti
v
e
ly
.
Th
e
e
ffe
c
ti
v
e
n
e
ss
o
f
th
e
su
g
g
e
ste
d
m
e
th
o
d
is
v
e
rifi
e
d
u
sin
g
p
u
b
l
icly
a
v
a
il
a
b
le
IS
IC
2
0
1
7
–
2
0
2
0
d
a
tas
e
ts,
th
e
se
m
o
d
e
l
re
p
o
rts
a
v
e
ra
g
e
a
c
c
u
ra
c
y
sc
o
re
s
o
f
9
6
.
1
0
%
,
9
7
.
2
3
%
,
9
7
.
5
0
%
,
9
8
.
3
3
%
,
a
n
d
9
8
.
6
0
%
,
in
th
a
t
o
r
d
e
r.
K
ey
w
o
r
d
s
:
C
las
s
if
icatio
n
Dee
p
lear
n
in
g
I
m
ag
e
p
r
o
ce
s
s
in
g
Me
lan
o
m
a
T
r
an
s
f
er
lear
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
:
So
u
m
y
a
Gad
ag
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
,
J
ain
C
o
lleg
e
o
f
E
n
g
in
ee
r
in
g
a
n
d
R
esear
ch
Kar
n
atak
a,
I
n
d
ia
E
m
ail:
s
o
u
m
y
a.
g
a
d
ag
5
@
g
m
ai
l.c
o
m
1.
I
NT
RO
D
UCT
I
O
N
Sk
in
ca
n
ce
r
is
th
e
m
o
s
t
p
r
ev
alen
t
f
o
r
m
o
f
ca
n
ce
r
,
s
u
r
p
ass
in
g
all
o
th
er
ty
p
es
co
m
b
in
e
d
in
ter
m
s
o
f
d
iag
n
o
s
is
r
ates
an
n
u
ally
.
I
n
t
h
e
Un
ited
States
alo
n
e,
th
er
e
ar
e
ap
p
r
o
x
im
ately
9
,
5
0
0
f
r
e
s
h
ca
s
es
id
en
tifie
d
d
aily
,
as
r
ep
o
r
ted
b
y
th
e
Sk
in
C
an
ce
r
Fo
u
n
d
atio
n
in
2
0
1
7
[
1
]
.
B
y
2
0
4
0
,
th
e
r
e
will
lik
ely
b
e
clo
s
e
to
h
alf
a
m
illi
o
n
o
cc
u
r
r
en
ce
s
o
f
s
k
in
ca
n
ce
r
,
with
m
elan
o
m
a
b
ein
g
th
e
d
ea
d
lies
t
ty
p
e,
m
ar
k
i
n
g
a
s
tag
g
er
in
g
6
2
%
s
u
r
g
e
s
in
ce
2
0
1
8
.
T
h
e
s
ev
er
ity
o
f
t
h
e
s
itu
atio
n
is
em
p
h
asized
b
y
th
e
alar
m
in
g
f
ac
t
th
at
o
n
e
p
er
s
o
n
lo
s
es
th
eir
life
to
s
k
in
ca
n
ce
r
ev
er
y
f
o
u
r
m
in
u
t
es,
p
r
o
m
p
tin
g
d
er
m
at
o
lo
g
is
ts
wo
r
ld
wid
e
to
class
if
y
its
r
is
in
g
in
cid
en
ce
as
a
g
lo
b
al
ep
id
e
m
ic,
as n
o
ted
b
y
Me
lan
o
m
a
UK
in
2
0
2
0
[
2
]
.
E
ar
ly
d
etec
tio
n
a
n
d
i
n
ter
v
en
ti
o
n
,
esp
ec
ially
f
o
r
m
elan
o
m
a,
em
er
g
e
as
p
iv
o
tal
f
ac
to
r
s
in
i
m
p
r
o
v
i
n
g
th
e
s
u
r
v
iv
al
r
ates
am
id
th
e
m
o
u
n
tin
g
ca
s
es
[
3
]
.
E
x
ce
s
s
iv
e
c
o
n
tact
to
u
ltra
v
io
let
(
UV)
r
ad
i
atio
n
s
tan
d
s
o
u
t
as
th
e
p
r
im
ar
y
id
e
n
tifia
b
le
cu
lp
r
i
t
b
eh
in
d
s
k
in
ca
n
ce
r
[
4
]
.
Natu
r
al
s
u
n
s
h
in
e
o
r
o
th
e
r
UV
s
o
u
r
ce
s
,
s
u
ch
as
in
d
o
o
r
tan
n
in
g
m
ac
h
in
es,
m
ay
b
e
th
e
s
o
u
r
ce
o
f
th
is
ex
p
o
s
u
r
e
(
C
an
ce
r
R
esear
ch
UK)
[
5
]
.
Sig
n
if
ican
tly
,
r
ed
u
ce
d
o
zo
n
e
lev
els
r
aise
th
e
r
is
k
o
f
ex
p
o
s
u
r
e
to
n
atu
r
al
s
u
n
s
h
in
e
b
y
in
c
r
ea
s
in
g
g
r
o
u
n
d
-
l
ev
el
UV
r
ad
iatio
n
(
Dep
ar
tm
en
t
f
o
r
E
n
v
ir
o
n
m
en
t
Fo
o
d
&
R
u
r
al
Af
f
air
s
,
2
0
2
0
)
[
6
]
.
Fu
r
th
er
m
o
r
e,
r
eg
io
n
s
s
it
u
ated
clo
s
er
to
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
9
4
3
-
4
9
5
6
4944
eq
u
ato
r
witn
ess
an
u
p
tick
i
n
n
o
n
-
m
elan
o
m
a
s
k
in
ca
n
ce
r
ca
s
es
d
u
e
to
elev
ated
U
V
r
ad
iatio
n
lev
els.
Fu
r
th
er
m
o
r
e
,
life
s
ty
le
f
ac
to
r
s
s
u
ch
as p
o
o
r
d
ietar
y
ch
o
ices,
a
lco
h
o
l c
o
n
s
u
m
p
tio
n
,
an
d
s
m
o
k
in
g
also
co
n
tr
ib
u
te
to
th
e
m
o
d
i
f
iab
le
r
is
k
p
r
o
f
ile
ass
o
ciate
d
with
s
k
in
ca
n
ce
r
.
T
h
er
ef
o
r
e
,
ea
r
l
y
d
etec
tio
n
a
n
d
p
r
ev
en
tio
n
o
f
t
h
ese
ca
n
ce
r
s
is
s
tu
d
ied
wid
ely
.
Sev
er
al
m
eth
o
d
s
h
av
e
b
ee
n
i
n
tr
o
d
u
ce
d
f
o
r
m
elan
o
m
a
d
ete
ctio
n
;
h
o
wev
er
,
im
ag
in
g
-
b
ase
d
m
eth
o
d
s
h
av
e
b
ee
n
wid
ely
ad
o
p
ted
in
v
ar
io
u
s
b
io
m
ed
ical
ap
p
licatio
n
s
.
Der
m
ato
lo
g
is
ts
ca
n
d
iag
n
o
s
e
m
alig
n
an
t
lesi
o
n
s
th
r
o
u
g
h
d
er
m
o
s
co
p
ic
v
is
u
al
e
x
am
in
atio
n
s
.
T
h
e
d
iv
er
s
e
tex
t
u
r
es
an
d
wo
u
n
d
s
p
r
esen
t
o
n
th
e
s
k
in
s
u
r
f
ac
e
ca
n
m
ak
e
d
etec
tin
g
s
k
in
ca
n
ce
r
c
h
allen
g
in
g
u
s
in
g
d
e
r
m
o
s
co
p
y
.
Yet,
ac
cu
r
ately
d
ia
g
n
o
s
in
g
s
k
in
ca
n
ce
r
th
r
o
u
g
h
m
an
u
al
ex
a
m
in
atio
n
o
f
d
er
m
o
s
co
p
ic
im
ag
es
is
d
if
f
icu
lt
.
T
h
e
ac
c
u
r
ac
y
o
f
lesi
o
n
d
i
ag
n
o
s
is
is
h
ea
v
ily
in
f
lu
en
ce
d
b
y
th
e
d
er
m
ato
lo
g
is
t
'
s
ex
p
er
ien
ce
.
T
h
e
o
n
ly
alter
n
ativ
e
tech
n
iq
u
es
f
o
r
d
etec
ti
n
g
s
k
in
ca
n
ce
r
th
at
ar
e
n
o
w
av
ailab
le
ar
e
d
er
m
o
s
co
p
y
,
b
io
p
s
y
,
an
d
m
ac
r
o
s
co
p
i
c
in
s
p
ec
tio
n
.
B
ec
au
s
e
s
k
in
le
s
io
n
s
ar
e
co
m
p
lex
,
th
ey
r
eq
u
ir
e
m
o
r
e
tim
e
an
d
c
ar
e.
T
h
e
d
er
m
ato
l
o
g
is
ts
p
er
f
o
r
m
th
e
o
b
s
er
v
atio
n
with
n
a
k
e
d
ey
es,
d
er
m
o
s
co
p
y
m
ec
h
an
is
m
s
an
d
b
io
p
s
y
.
T
h
e
r
ef
o
r
e,
th
e
ac
cu
r
ac
y
o
f
th
ese
s
y
s
tem
s
r
elies
o
n
clin
ician
’
s
s
k
ill.
A
s
ig
n
if
ican
t
am
o
u
n
t
o
f
r
esear
ch
h
as
b
ee
n
d
ed
icate
d
to
d
ev
elo
p
in
g
co
m
p
u
ter
-
b
ased
im
ag
e
an
aly
s
is
alg
o
r
ith
m
s
f
o
r
t
h
e
ea
r
l
y
an
d
r
ap
id
d
iag
n
o
s
is
o
f
s
k
in
ca
n
ce
r
,
aim
i
n
g
to
o
v
er
c
o
m
e
t
h
e
p
r
ev
io
u
s
ly
m
en
tio
n
ed
ch
allen
g
es,
Prim
ar
ily
,
th
ese
alg
o
r
ith
m
s
h
av
e
b
ee
n
p
a
r
am
e
tr
ic,
r
ely
in
g
o
n
n
o
r
m
ally
d
is
tr
ib
u
ted
d
ata
f
o
r
o
p
er
atio
n
.
Ho
wev
er
,
d
u
e
to
th
e
u
n
co
n
tr
o
llab
le
n
atu
r
e
o
f
t
h
e
d
a
ta,
th
ese
m
eth
o
d
s
o
f
te
n
f
ail
to
p
r
o
v
id
e
ac
cu
r
ate
d
iag
n
o
s
es.
Nu
m
er
o
u
s
m
ed
ical
im
ag
in
g
r
esear
ch
er
s
h
av
e
in
tr
o
d
u
ce
d
co
m
p
u
ter
-
aid
e
d
d
esig
n
(
C
AD)
tech
n
iq
u
es.
T
h
is
f
o
u
r
-
s
tep
C
AD
p
r
o
ce
s
s
en
co
m
p
ass
es
im
ag
e
p
r
e
-
p
r
o
ce
s
s
in
g
,
id
en
tific
atio
n
o
f
af
f
ec
ted
ar
ea
s
,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
class
if
icatio
n
.
Sev
er
al
m
eth
o
d
s
h
a
v
e
b
ee
n
in
tr
o
d
u
ce
d
u
s
in
g
co
m
p
u
ter
v
is
io
n
ap
p
r
o
ac
h
s
u
ch
as
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
(
AN
N)
,
d
ec
is
io
n
tr
ee
s
(
DT
)
,
an
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es
(
SVM)
.
T
h
e
r
esear
ch
in
[7
]
,
[
8
]
p
r
o
v
i
d
es
a
th
o
r
o
u
g
h
an
aly
s
is
o
f
v
ar
io
u
s
tech
n
iq
u
es.
Nev
er
th
eless
,
th
er
e
ar
e
a
n
u
m
b
e
r
o
f
d
ata
p
r
o
ce
s
s
in
g
lim
itatio
n
s
with
m
ac
h
in
e
lear
n
in
g
tech
n
iq
u
es,
i
n
clu
d
in
g
th
e
n
ee
d
f
o
r
b
etter
co
n
tr
ast,
n
o
is
e
-
f
r
ee
,
an
d
clea
n
ed
im
a
g
es.
Mo
r
eo
v
e
r
,
a
n
u
m
b
er
o
f
c
r
iter
ia,
in
clu
d
in
g
s
tr
u
ctu
r
al
tr
aits
,
co
lo
r
attr
ib
u
tes,
an
d
tex
tu
r
e
attr
ib
u
tes,
ar
e
u
s
ed
to
class
if
y
s
k
in
[
9
]
,
[
1
0
]
.
Ho
wev
er
,
cl
ass
if
icatio
n
b
ased
o
n
in
ad
eq
u
ate
f
ea
tu
r
e
s
ets
ca
n
r
esu
lt
in
er
r
o
n
e
o
u
s
o
u
tco
m
es
d
u
e
to
th
e
h
ig
h
in
ter
-
class
h
o
m
o
g
en
eity
an
d
in
tr
a
-
class
h
et
er
o
g
en
eity
o
f
s
k
in
lesi
o
n
s
[
1
1
]
.
C
o
n
v
en
tio
n
al
p
a
r
am
etr
ic
m
eth
o
d
s
r
e
q
u
ir
e
tr
ain
in
g
d
ata
to
b
e
n
o
r
m
ally
d
is
tr
ib
u
ted
,
wh
ic
h
is
n
o
t
th
e
ca
s
e
f
o
r
u
n
co
n
tr
o
lled
s
k
in
ca
n
ce
r
d
ata.
Sin
ce
ev
er
y
lesi
o
n
h
as
a
d
if
f
er
e
n
t
p
atter
n
,
th
ese
tr
ad
itio
n
al
ap
p
r
o
ac
h
es
ar
e
in
s
u
f
f
icien
t.
As
a
r
esu
lt,
d
ee
p
lear
n
in
g
m
eth
o
d
s
h
av
e
s
h
o
wn
to
b
e
q
u
ite
s
u
cc
ess
f
u
l
at
class
if
y
in
g
s
k
in
,
h
elp
i
n
g
d
er
m
ato
lo
g
is
ts
d
iag
n
o
s
e
lesi
o
n
s
with
h
ig
h
p
r
ec
is
io
n
.
T
h
e
a
p
p
licatio
n
o
f
d
ee
p
lear
n
in
g
in
m
ed
icin
e
h
as
b
ee
n
wid
ely
ex
p
lo
r
ed
t
h
r
o
u
g
h
v
a
r
io
u
s
d
etailed
s
u
r
v
ey
s
.
T
h
e
m
o
s
t
r
ec
en
t
r
esear
ch
o
n
d
ee
p
lear
n
in
g
-
b
ased
m
eth
o
d
s
f
o
r
m
elan
o
m
a
d
etec
tio
n
an
d
cla
s
s
if
icatio
n
is
co
v
er
ed
in
s
ec
tio
n
2
.
Dee
p
lear
n
in
g
tec
h
n
iq
u
es
b
ased
o
n
tr
a
n
s
f
er
lear
n
in
g
h
a
v
e
g
a
in
ed
wid
esp
r
ea
d
tr
ac
tio
n
at
th
e
m
o
m
e
n
t.
W
ith
th
is
m
eth
o
d
,
lar
g
e
-
s
ca
le
d
atasets
f
r
o
m
o
n
e
d
o
m
ain
(
lik
e
n
atu
r
al
p
h
o
to
g
r
ap
h
s
)
a
r
e
u
s
ed
to
tr
ain
d
ee
p
lear
n
in
g
m
o
d
els,
wh
ich
ar
e
th
en
u
s
ed
to
tr
an
s
f
er
th
eir
lear
n
ed
r
ep
r
esen
tatio
n
s
to
a
tar
g
et
d
o
m
ain
(
b
io
m
ed
ical
im
ag
es).
B
y
f
in
e
-
tu
n
in
g
th
ese
p
r
e
-
tr
ain
ed
m
o
d
els
o
n
s
m
aller
b
io
m
ed
ical
d
atasets
,
r
esear
ch
er
s
ca
n
ef
f
ec
tiv
ely
class
if
y
b
io
m
ed
ical
im
ag
es
with
h
ig
h
ac
cu
r
ac
y
.
T
r
a
n
s
f
er
le
ar
n
in
g
in
b
io
m
ed
ical
im
ag
e
c
lass
if
icatio
n
o
f
f
er
s
s
ev
er
al
ad
v
an
tag
es.
Firstl
y
,
it
allo
ws
r
esear
ch
er
s
to
o
v
er
co
m
e
th
e
ch
allen
g
e
o
f
lim
ited
an
n
o
tated
d
ata
in
th
e
b
io
m
ed
ical
d
o
m
ain
b
y
lev
er
a
g
in
g
k
n
o
wled
g
e
lear
n
ed
f
r
o
m
lar
g
er
d
atasets
in
r
elate
d
d
o
m
ain
s
.
Seco
n
d
ly
,
p
r
e
-
tr
ain
ed
m
o
d
els
ca
p
tu
r
e
g
en
er
ic
im
ag
e
f
ea
tu
r
es
lik
e
te
x
tu
r
es,
s
h
ap
es,
an
d
ed
g
es,
wh
ich
ca
n
t
y
p
ically
tr
an
s
f
er
ab
le
ac
r
o
s
s
d
if
f
er
en
t
i
m
ag
e
d
o
m
ain
s
.
T
h
is
allo
ws
th
e
m
o
d
el
to
lear
n
im
p
o
r
ta
n
t
f
ea
tu
r
es
f
o
r
b
io
m
ed
ical
im
ag
e
class
if
icatio
n
task
s
wi
th
o
u
t
r
e
q
u
ir
in
g
ex
ten
s
iv
e
r
et
r
ain
in
g
f
r
o
m
s
cr
atch
.
Sev
er
al
m
o
d
els
h
av
e
b
ee
n
p
r
esen
ted
u
s
in
g
th
is
c
o
n
ce
p
t
o
f
tr
an
s
f
er
lear
n
in
g
s
u
c
h
as
Z
u
n
air
an
d
Ham
za
[
1
2
]
p
r
esen
ted
a
m
o
d
el
co
m
p
o
s
ed
o
f
two
s
tag
es,
co
m
b
in
i
n
g
ad
v
e
r
s
ar
ial
tr
ain
in
g
with
tr
an
s
f
er
le
ar
n
in
g
,
Q
u
r
esh
i
et
a
l.
[
1
3
]
u
s
ed
Go
o
g
le
Xce
p
tio
n
m
o
d
el
to
d
ev
elo
p
th
e
tr
an
s
f
er
lear
n
in
g
ar
ch
itectu
r
e
,
Ho
s
n
y
et
a
l.
[
1
4
]
u
s
ed
Alex
n
et.
Ho
wev
er
,
th
e
tr
an
s
f
e
r
lear
n
in
g
-
b
ased
m
o
d
els
s
u
f
f
e
r
f
r
o
m
d
if
f
er
en
t
is
s
u
es
s
u
ch
as
d
o
m
ain
tr
an
s
f
er
b
etwe
en
s
o
u
r
c
e
an
d
tar
g
et
d
o
m
ai
n
wh
er
e
r
eso
lu
tio
n
,
n
o
is
e
lev
el,
an
d
tis
s
u
e
v
a
r
iab
ilit
y
a
f
f
ec
ts
t
h
e
tr
an
s
f
er
ab
ilit
y
o
f
lea
r
n
ed
at
tr
ib
u
tes.
As
a
r
esu
lt,
th
e
p
r
e
-
tr
ain
e
d
m
o
d
el
is
u
n
ab
l
e
to
ca
p
tu
r
e
t
h
e
p
er
tin
e
n
t
ch
ar
ac
ter
is
tics
.
Fu
r
th
er
m
o
r
e,
a
s
ig
n
if
ican
t
q
u
a
n
tity
o
f
d
ata
is
n
ee
d
ed
to
r
ef
in
e
t
h
e
p
r
ev
io
u
s
ly
tr
ain
ed
m
o
d
els.
I
n
t
h
is
r
eser
ac
h
,
o
u
r
m
ain
o
b
jectiv
e
is
to
d
e
v
elo
p
a
d
ee
p
lear
n
in
g
-
b
ased
m
elan
o
m
a
ca
teg
o
r
izatio
n
t
h
r
o
u
g
h
th
e
u
s
e
o
f
tr
an
s
f
er
lear
n
in
g
s
tr
ateg
y
.
A
s
d
i
s
c
u
s
s
e
d
b
e
f
o
r
e
,
s
k
i
n
c
a
n
ce
r
o
c
c
u
r
s
m
o
r
e
c
o
m
m
o
n
l
y
t
h
a
n
a
n
y
o
t
h
e
r
f
o
r
m
o
f
c
a
n
c
e
r
,
w
i
t
h
d
i
a
g
n
o
s
i
s
r
a
t
e
s
s
u
r
p
a
s
s
i
n
g
a
ll
o
t
h
e
r
t
y
p
es
c
o
m
b
i
n
e
d
,
a
n
d
a
p
p
r
o
x
i
m
a
t
e
ly
9
,
5
0
0
n
e
w
c
a
s
e
s
a
r
e
d
i
a
g
n
o
s
e
d
e
a
c
h
d
a
y
i
n
t
h
e
U
n
i
t
e
d
S
ta
t
es
a
l
o
n
e
.
B
y
2
0
4
0
,
s
k
i
n
c
a
n
c
e
r
c
a
s
e
s
a
r
e
p
r
o
j
e
c
te
d
t
o
a
p
p
r
o
a
c
h
h
a
l
f
a
m
i
l
l
i
o
n
,
w
i
t
h
m
e
l
a
n
o
m
a
b
e
i
n
g
t
h
e
d
e
a
d
l
i
es
t
t
y
p
e
,
m
a
r
k
i
n
g
a
s
t
a
g
g
e
r
i
n
g
6
2
%
i
n
c
r
e
as
e
s
i
n
c
e
2
0
1
8
.
S
k
i
n
c
a
n
c
e
r
i
s
p
r
i
m
a
r
i
l
y
c
a
u
s
e
d
b
y
o
v
e
r
e
x
p
o
s
u
r
e
t
o
UV
r
a
d
i
at
i
o
n
f
r
o
m
s
u
n
l
i
g
h
t
o
r
a
r
t
i
f
ic
i
a
l
s
o
u
r
c
e
s
l
i
k
e
t
a
n
n
i
n
g
m
a
c
h
i
n
es
,
c
o
m
p
o
u
n
d
e
d
b
y
f
a
c
t
o
r
s
s
u
c
h
a
s
r
e
d
u
c
e
d
o
z
o
n
e
l
e
v
e
l
s
,
g
e
o
g
r
a
p
h
i
c
a
l
p
r
o
x
i
m
i
t
y
t
o
t
h
e
e
q
u
a
t
o
r
,
p
o
o
r
d
i
e
t
a
r
y
h
a
b
i
t
s
,
a
l
c
o
h
o
l
c
o
n
s
u
m
p
t
i
o
n
,
a
n
d
s
m
o
k
i
n
g
.
D
e
s
p
i
t
e
t
h
e
u
r
g
e
n
c
y
o
f
t
h
e
s
i
t
u
a
t
i
o
n
,
c
u
r
r
e
n
t
d
i
a
g
n
o
s
t
i
c
m
e
t
h
o
d
s
li
k
e
d
e
r
m
o
s
c
o
p
y
,
b
i
o
p
s
y
,
a
n
d
m
a
c
r
o
s
c
o
p
i
c
i
n
s
p
e
c
ti
o
n
a
r
e
h
i
g
h
l
y
d
e
p
e
n
d
e
n
t
o
n
t
h
e
c
l
i
n
i
ci
a
n
'
s
s
k
il
l
,
m
a
k
i
n
g
a
c
c
u
r
a
t
e
d
ia
g
n
o
s
is
c
h
a
l
le
n
g
i
n
g
a
n
d
t
i
m
e
-
c
o
n
s
u
m
i
n
g
.
T
r
a
d
it
i
o
n
a
l
c
o
m
p
u
t
e
r
i
m
a
g
e
a
n
a
l
y
s
i
s
a
l
g
o
r
i
t
h
m
s
a
n
d
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
t
e
c
h
n
i
q
u
e
s
h
a
v
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mela
n
o
ma
cla
s
s
ifica
tio
n
u
s
in
g
en
s
emb
le
d
ee
p
tr
a
n
s
fer lea
r
n
in
g
(
S
o
u
mya
G
a
d
a
g
)
4945
a
t
t
e
m
p
t
e
d
t
o
as
s
i
s
t
i
n
e
a
r
l
y
d
i
ag
n
o
s
i
s
,
b
u
t
t
h
es
e
m
e
t
h
o
d
s
o
f
t
en
f
a
l
l
s
h
o
r
t
d
u
e
t
o
t
h
e
u
n
c
o
n
t
r
o
l
l
e
d
n
a
t
u
r
e
o
f
s
k
i
n
c
a
n
c
e
r
d
a
t
a
a
n
d
i
s
s
u
e
s
li
k
e
i
n
ad
e
q
u
a
t
e
f
e
a
t
u
r
e
s
et
s
,
h
i
g
h
i
n
t
e
r
-
c
l
a
s
s
h
o
m
o
g
e
n
e
i
t
y
,
a
n
d
i
n
t
r
a
-
c
l
a
s
s
h
e
t
e
r
o
g
e
n
e
it
y
.
D
e
e
p
l
e
a
r
n
i
n
g
,
e
s
p
e
c
ia
l
l
y
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
,
o
f
f
e
r
s
a
p
r
o
m
i
s
i
n
g
s
o
l
u
t
i
o
n
b
y
l
e
v
e
r
a
g
i
n
g
l
a
r
g
e
-
s
c
a
l
e
d
a
ta
s
et
s
f
r
o
m
r
e
l
a
t
e
d
d
o
m
a
i
n
s
t
o
e
n
h
a
n
c
e
c
la
s
s
i
f
i
c
at
i
o
n
a
c
c
u
r
a
c
y
i
n
b
i
o
m
ed
i
c
a
l
i
m
a
g
es
.
H
o
w
e
v
e
r
,
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
m
o
d
e
l
s
f
a
c
e
c
h
a
l
l
e
n
g
e
s
s
u
c
h
a
s
d
o
m
a
in
t
r
a
n
s
f
e
r
i
s
s
u
es
a
n
d
t
h
e
n
e
e
d
f
o
r
s
i
g
n
i
f
i
c
a
n
t
a
m
o
u
n
t
s
o
f
d
a
t
a
t
o
f
i
n
e
-
t
u
n
e
p
r
e
-
t
r
a
i
n
e
d
m
o
d
e
l
s
e
f
f
e
c
ti
v
e
l
y
.
T
h
i
s
w
o
r
k
f
o
c
u
s
es
o
n
d
e
v
e
l
o
p
in
g
a
r
o
b
u
s
t
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
f
o
r
m
e
l
a
n
o
m
a
c
l
a
s
s
i
f
ic
a
t
i
o
n
t
h
r
o
u
g
h
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
,
a
i
m
i
n
g
t
o
o
v
e
r
c
o
m
e
t
h
e
l
i
m
i
t
a
ti
o
n
s
o
f
e
x
i
s
t
i
n
g
a
p
p
r
o
a
c
h
e
s
a
n
d
i
m
p
r
o
v
e
t
h
e
a
c
c
u
r
a
c
y
a
n
d
r
e
l
i
a
b
i
l
i
t
y
o
f
e
a
r
l
y
s
k
i
n
c
a
n
c
e
r
d
et
e
c
ti
o
n
.
T
h
e
k
e
y
co
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
ar
e
as
f
o
llo
ws:
T
h
e
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
p
er
f
o
r
m
s
s
e
v
er
al
task
s
s
u
ch
as
im
p
r
o
v
in
g
th
e
im
ag
e
q
u
ality
,
h
air
r
em
o
v
al
an
d
lab
el
en
co
d
in
g
to
p
r
o
ce
s
s
th
e
ef
f
icien
tly
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
es
tr
an
s
f
er
lear
n
in
g
ap
p
r
o
ac
h
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
wh
er
e
VGG,
I
n
ce
p
t
io
n
R
esNet,
R
e
s
Net
an
d
M
o
b
ileNet
m
o
d
u
les
ar
e
u
s
ed
to
o
b
tain
th
e
s
ig
n
if
ican
t
at
tr
ib
u
tes.
T
h
e
o
b
tain
ed
f
ea
tu
r
es
ar
e
f
u
s
ed
to
g
eth
e
r
to
attain
th
e
f
in
al
f
ea
t
u
r
e
v
ec
to
r
.
I
n
n
e
x
t
s
tep
,
d
ee
p
e
n
s
em
b
le
m
o
d
el
is
c
o
n
s
tr
u
cted
b
y
u
s
in
g
th
e
co
n
ce
p
t
o
f
tr
an
s
f
er
lear
n
in
g
wh
er
e
E
f
f
icien
tNet,
Xce
p
tio
n
,
an
d
Den
s
eNe
t
t
r
an
s
f
er
lear
n
in
g
m
o
d
els
ar
e
u
s
ed
to
o
b
tain
th
e
p
r
o
b
a
b
ilit
y
v
ec
to
r
f
o
r
p
r
e
d
ictin
g
th
e
s
k
in
ca
n
ce
r
.
Fin
ally
,
FC
an
d
s
ig
m
o
id
lay
er
s
a
r
e
u
s
ed
to
o
b
tain
th
e
co
n
clu
d
in
g
class
if
icatio
n
.
T
h
is
wo
r
k
in
tr
o
d
u
ce
s
s
ev
er
al
k
ey
in
n
o
v
atio
n
s
th
at
s
ig
n
if
ican
tly
en
h
an
ce
th
e
p
er
f
o
r
m
an
ce
o
f
m
elan
o
m
a
im
ag
e
class
if
icatio
n
s
y
s
tem
s
.
A
co
m
p
r
eh
en
s
iv
e
m
u
lti
-
s
tag
e
p
r
e
-
p
r
o
ce
s
s
in
g
p
ip
elin
e
is
im
p
lem
en
ted
to
ad
d
r
ess
co
m
m
o
n
d
er
m
at
o
lo
g
ical
im
ag
in
g
ch
allen
g
es.
T
h
is
p
ip
elin
e
im
p
r
o
v
es
co
n
tr
ast
to
h
ig
h
lig
h
t
c
r
itical
f
ea
tu
r
es
an
d
ap
p
lies
n
o
is
e
r
e
d
u
ctio
n
to
e
n
s
u
r
e
clea
n
er
a
n
d
m
o
r
e
r
eliab
l
e
in
p
u
t
f
o
r
f
u
r
t
h
er
an
aly
s
is
.
T
o
c
o
u
n
ter
ac
t
d
at
aset
im
b
alan
ce
—
a
f
r
e
q
u
en
t
is
s
u
e
in
m
ela
n
o
m
a
d
etec
t
io
n
.
Var
io
u
s
d
ata
au
g
m
en
tatio
n
tec
h
n
iq
u
es
ar
e
em
p
lo
y
ed
t
o
s
y
n
th
etica
lly
ex
p
an
d
th
e
m
in
o
r
ity
class
es,
en
h
an
cin
g
th
e
m
o
d
el’
s
r
o
b
u
s
tn
ess
an
d
ab
ilit
y
to
g
en
er
alize
ef
f
ec
tiv
ely
.
T
h
e
a
p
p
r
o
ac
h
also
lev
er
ag
es p
r
e
-
tr
ain
ed
d
e
ep
lear
n
in
g
m
o
d
els
lik
e
VGG,
R
e
s
Net,
Mo
b
ileN
et
an
d
I
n
ce
p
tio
n
R
esNet,
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
h
ese
m
o
d
els
ca
p
tu
r
e
r
ich
h
ier
ar
ch
ical
r
ep
r
esen
tatio
n
s
,
an
d
b
y
f
u
s
in
g
th
eir
f
ea
tu
r
e
v
ec
to
r
s
,
th
e
s
y
s
tem
ac
h
iev
es
a
m
o
r
e
co
m
p
r
eh
en
s
iv
e
an
d
d
is
cr
im
in
ativ
e
u
n
d
e
r
s
tan
d
in
g
o
f
t
h
e
in
p
u
t
im
ag
es.
Fin
all
y
,
en
s
em
b
le
class
if
icatio
n
u
s
in
g
tr
an
s
f
er
lear
n
in
g
is
em
p
lo
y
ed
,
co
m
b
in
in
g
th
e
s
tr
en
g
th
s
o
f
m
u
ltip
le
ar
ch
itect
u
r
es.
T
h
is
en
s
em
b
le
ap
p
r
o
ac
h
ca
p
tu
r
es
co
m
p
le
x
p
atter
n
s
an
d
s
u
b
tle
d
is
tin
ctio
n
s
with
in
m
elan
o
m
a
im
ag
es,
r
esu
ltin
g
in
im
p
r
o
v
ed
class
if
ic
atio
n
ac
cu
r
ac
y
an
d
o
v
er
all
s
y
s
tem
p
er
f
o
r
m
an
ce
.
T
h
e
s
u
b
s
eq
u
e
n
t
s
ec
tio
n
s
o
f
th
e
ar
ticle
ar
e
s
tr
u
ctu
r
ed
as
f
o
ll
o
ws:
s
ec
tio
n
2
p
r
o
v
id
es
a
co
m
p
r
eh
en
s
iv
e
liter
atu
r
e
r
ev
iew
o
f
cu
r
r
en
t
m
elan
o
m
a
class
if
icatio
n
tech
n
iq
u
es
.
Sectio
n
3
p
r
esen
ts
an
in
-
d
ep
th
o
v
e
r
v
iew
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
.
Sectio
n
4
o
u
tlin
es
th
e
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
alo
n
g
with
a
co
m
p
ar
ativ
e
an
aly
s
is
with
cu
r
r
en
t te
ch
n
iq
u
es
.
L
astl
y
,
s
ec
tio
n
5
o
f
f
er
s
clo
s
in
g
t
h
o
u
g
h
ts
an
d
f
u
tu
r
e
d
ir
ec
tio
n
s
f
o
r
th
e
r
esear
ch
.
2.
L
I
T
E
R
AT
U
RE
SU
RVE
Y
T
h
is
s
ec
tio
n
g
iv
es
in
d
etailed
an
aly
s
is
o
f
th
e
cu
r
r
en
t
ap
p
r
o
a
ch
es
f
o
r
class
if
y
in
g
m
elan
o
m
a
with
th
e
h
elp
o
f
d
ee
p
lear
n
i
n
g
an
d
tr
an
s
f
er
lear
n
in
g
m
eth
o
d
s
.
L
u
a
n
d
Z
ad
eh
[
1
5
]
p
r
esen
ted
an
a
u
to
m
ated
m
eth
o
d
f
o
r
u
tili
zin
g
d
er
m
o
s
co
p
y
im
ag
es
to
d
iag
n
o
s
e
s
k
in
ca
n
ce
r
.
T
o
in
cr
ea
s
e
class
if
icatio
n
ac
cu
r
ac
y
,
th
e
m
o
d
el
u
s
e
d
d
ep
th
-
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
an
d
th
e
s
wis
h
ac
tiv
atio
n
f
u
n
ctio
n
,
with
Xce
p
tio
n
Net
s
er
v
in
g
as
th
e
b
ase
n
etwo
r
k
.
J
ain
et
a
l.
[
1
6
]
p
r
es
en
ted
a
d
ee
p
tr
an
s
f
er
lear
n
in
g
m
o
d
el
was
in
tr
o
d
u
ce
d
.
T
o
a
d
d
r
ess
th
e
is
s
u
e
o
f
d
ata
im
b
alan
ce
,
th
e
m
eth
o
d
in
clu
d
ed
im
ag
e
d
ata
a
u
g
m
en
tatio
n
.
M
o
r
eo
v
e
r
,
f
ea
t
u
r
e
ex
tr
ac
tio
n
an
d
class
if
icatio
n
task
s
wer
e
p
er
f
o
r
m
ed
u
s
in
g
tr
an
s
f
er
lear
n
in
g
tech
n
iq
u
es
s
u
ch
as
VGG1
9
,
I
n
ce
p
tio
n
V3
,
a
n
d
Mo
b
ileNet,
am
o
n
g
o
th
er
s
.
Fo
llo
win
g
th
ese
ar
ch
itectu
r
es,
th
e
task
was
co
m
p
leted
b
y
in
co
r
p
o
r
atin
g
m
a
x
p
o
o
lin
g
,
f
latten
in
g
,
a
d
en
s
e
la
y
er
,
an
d
th
e
s
ig
m
o
i
d
f
u
n
ctio
n
.
Ali
et
a
l.
[
1
7
]
d
is
cu
s
s
ed
th
e
ch
allen
g
es
ass
o
ciate
d
with
th
e
c
u
r
r
en
t
co
m
p
u
ter
ized
s
k
in
lesi
o
n
m
alig
n
an
cy
d
etec
tio
n
s
y
s
tem
d
u
e
to
v
ar
io
u
s
v
ar
ia
b
les,
s
u
ch
as
u
n
ev
en
lesi
o
n
s
izes
an
d
s
h
a
p
es,
d
if
f
er
e
n
t
co
l
o
r
illu
m
in
atio
n
s
,
an
d
v
ar
y
i
n
g
lig
h
t
co
n
d
itio
n
s
.
I
n
th
e
p
r
ep
r
o
ce
s
s
in
g
s
tag
e,
f
ilter
s
o
r
k
er
n
e
ls
ar
e
em
p
lo
y
ed
to
elim
in
ate
ar
tifa
cts
an
d
n
o
is
e,
f
o
llo
win
g
wh
ich
r
o
b
u
s
t
f
ea
t
u
r
es
ar
e
ex
tr
ac
ted
u
s
in
g
th
e
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
d
u
r
e.
Fin
ally
,
d
ata
a
u
g
m
en
tatio
n
is
u
s
ed
t
o
in
cr
ea
s
e
th
e
s
ize
o
f
th
e
im
a
g
e
co
ll
ec
tio
n
an
d
im
p
r
o
v
e
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
is
m
eth
o
d
is
co
m
p
ar
ed
with
th
at
o
f
o
th
er
tr
an
s
f
er
lear
n
in
g
m
o
d
els,
in
clu
d
in
g
Alex
Net
an
d
R
eNe
t
.
B
alah
a
an
d
Hass
an
[
1
8
]
p
r
esen
ted
an
au
to
m
ate
d
m
ela
n
o
m
a
class
if
icatio
n
a
n
d
s
eg
m
en
tatio
n
ap
p
r
o
ac
h
b
y
u
s
in
g
th
e
s
p
ar
r
o
w
s
ea
r
ch
alg
o
r
ith
m
(
Sp
aSA)
an
d
m
eta
-
h
e
u
r
is
tic
o
p
tim
izatio
n
to
o
l.
T
h
e
s
eg
m
en
tatio
n
m
eth
o
d
in
v
o
lv
es
em
p
lo
y
in
g
5
d
if
f
er
en
t
U
-
Net
m
o
d
els
—
U
-
Net,
U
-
Net+
+,
atten
tio
n
-
b
ased
UNe
t,
an
d
s
ev
er
al
o
th
e
r
s
—
ea
ch
with
a
u
n
iq
u
e
co
n
f
ig
u
r
atio
n
.
Ad
d
itio
n
ally
,
eig
h
t
p
r
e
-
tr
ain
ed
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
m
o
d
els,
in
clu
d
in
g
Mo
b
ileNet
wi
th
VGG
(
s
m
all,
b
i
g
)
,
ar
e
u
t
ilized
to
o
p
tim
ize
h
y
p
er
p
ar
am
eter
s
u
s
in
g
th
e
m
eta
-
h
eu
r
is
tic
Sp
aSA
o
p
tim
izer
.
Me
s
wal
et
a
l.
[
1
9
]
in
tr
o
d
u
ce
d
an
e
n
s
em
b
le
s
tr
ateg
y
f
o
r
class
if
y
in
g
s
k
in
lesi
o
n
s
u
s
in
g
weig
h
ted
av
er
a
g
es.
T
h
ey
o
b
tain
ed
th
e
weig
h
ted
s
u
m
o
f
tr
a
n
s
f
er
lear
n
in
g
m
o
d
els in
clu
d
in
g
I
n
c
ep
tio
n
V3
,
VGG1
6
,
Xce
p
tio
n
,
an
d
R
esNet5
0
to
cr
ea
te
th
e
e
n
s
em
b
le
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
9
4
3
-
4
9
5
6
4946
Sad
ik
et
a
l.
[
2
0
]
p
r
o
p
o
s
ed
a
C
NN
-
b
ased
ar
ch
itectu
r
e
th
at
co
m
b
in
es
Xce
p
tio
n
an
d
Mo
b
ileNet.
Sp
ec
if
ically
,
th
ey
u
tili
ze
d
C
NN
ar
ch
itectu
r
es
co
m
m
o
n
ly
ap
p
lied
in
c
o
m
p
u
ter
v
is
io
n
a
p
p
licatio
n
s
,
n
am
ely
Mo
b
ileNet
an
d
Xce
p
tio
n
,
t
o
d
ev
elo
p
a
s
y
s
tem
to
d
etec
t
s
k
in
ca
n
ce
r
.
Kar
r
i
et
a
l.
[
2
1
]
in
tr
o
d
u
ce
d
a
n
o
v
el
s
tr
ateg
y
to
a
d
d
r
ess
th
e
is
s
u
e
w
ith
cu
r
r
e
n
t
d
e
ep
lear
n
in
g
tech
n
iq
u
es,
n
a
m
ely
t
h
e
ch
alle
n
g
e
o
f
g
e
n
er
alizin
g
d
ata
f
r
o
m
v
ar
io
u
s
s
o
u
r
ce
s
,
lead
in
g
to
d
o
m
ain
s
h
if
t
e
v
en
in
well
-
tr
ain
ed
d
ee
p
lea
r
n
in
g
m
o
d
els.
T
o
o
v
er
c
o
m
e
th
is
,
th
ey
p
r
o
p
o
s
ed
a
tr
a
n
s
f
er
lear
n
in
g
m
eth
o
d
o
l
o
g
y
wh
ich
i
n
v
o
lv
es
two
-
p
h
ase
cr
o
s
s
-
d
o
m
ain
.
T
h
is
m
eth
o
d
u
tili
ze
s
th
e
I
m
ag
eNe
t
an
d
Mo
leM
ap
d
atasets
to
co
n
s
tr
u
ct
an
d
en
h
an
ce
d
ata
-
lev
el
an
d
m
o
d
el
-
lev
el
tr
an
s
f
er
lear
n
in
g
m
o
d
els.
Ad
d
itio
n
ally
,
t
h
ey
in
tr
o
d
u
ce
d
Sk
n
R
SUNet
f
o
r
s
eg
m
en
tatio
n
,
wh
ich
i
n
co
r
p
o
r
ate
s
s
p
atial
atten
tio
n
f
ea
tu
r
es m
er
g
in
g
.
Sh
ek
ar
an
d
Hailu
[
2
2
]
p
r
ese
n
ted
a
d
ee
p
t
r
an
s
f
er
lear
n
in
g
ap
p
r
o
ac
h
t
h
at
co
m
b
in
es
s
ix
s
p
ec
ially
d
esig
n
ed
alg
o
r
ith
m
s
with
th
e
Den
s
eNe
t
-
1
6
9
m
o
d
el
t
o
g
at
h
e
r
m
o
r
e
d
etailed
an
d
r
ich
e
r
f
ea
t
u
r
es.
Su
b
s
eq
u
e
n
tly
,
th
e
class
if
icatio
n
task
is
p
er
f
o
r
m
ed
u
s
in
g
th
e
g
r
a
d
ien
t
b
o
o
s
tin
g
m
ac
h
in
e
(
GB
M)
cla
s
s
if
icatio
n
m
o
d
el.
Den
g
[
2
3
]
h
ig
h
lig
h
te
d
th
e
s
ig
n
if
ican
ce
o
f
ch
allen
g
i
n
g
s
am
p
les,
s
u
g
g
esti
n
g
th
at
th
ey
co
n
tain
cr
u
cial
in
f
o
r
m
atio
n
.
I
n
th
eir
p
a
p
er
,
th
ey
in
tr
o
d
u
ce
d
a
n
o
v
el
m
eth
o
d
ca
lled
lim
ited
ex
am
p
les
n
etw
o
r
k
(
L
SNet)
aim
ed
at
r
ec
o
g
n
izin
g
an
d
e
n
h
an
ci
n
g
th
e
lear
n
in
g
o
f
s
u
c
h
d
if
f
i
cu
lt
ex
am
p
les.
L
SNet
u
tili
ze
s
a
p
s
eu
d
o
-
in
v
e
r
s
e
lear
n
in
g
au
to
e
n
co
d
e
r
with
a
p
atch
-
b
ased
s
tr
u
ctu
r
e
d
in
p
u
t
to
q
u
ick
ly
d
eter
m
i
n
e
p
o
s
itio
n
-
s
en
s
itiv
e
lo
s
s
.
B
y
ef
f
icien
tly
id
en
tify
in
g
p
o
s
itio
n
-
s
en
s
itiv
e
lo
s
s
,
L
SNet
ca
n
r
ec
o
g
n
ize
ch
allen
g
in
g
s
a
m
p
les
ef
f
ec
tiv
ely
.
Mo
r
eo
v
er
,
wh
en
d
ea
lin
g
with
s
k
in
lesi
o
n
d
atasets
with
f
ew
s
am
p
les,
d
ata
au
g
m
en
tatio
n
is
em
p
lo
y
ed
in
co
n
ju
n
ctio
n
with
tr
an
s
f
er
lear
n
in
g
to
im
p
r
o
v
e
th
e
ac
c
u
r
ac
y
o
f
class
if
icatio
n
in
d
ee
p
lear
n
in
g
m
o
d
els.
R
em
y
a
et
a
l
.
[
2
4
]
p
r
o
p
o
s
ed
a
d
ee
p
lear
n
in
g
-
b
ased
ar
c
h
itectu
r
e
in
teg
r
atin
g
v
is
io
n
tr
an
s
f
o
r
m
er
,
wh
ich
co
m
b
in
es
ch
a
n
n
el
atten
tio
n
a
n
d
tr
a
n
s
f
er
lear
n
in
g
tech
n
iq
u
es
to
d
eliv
er
ac
cu
r
ate
r
e
g
io
n
o
f
in
ter
est
(
R
OI
)
s
eg
m
en
tatio
n
an
d
class
if
icatio
n
.
3.
P
RO
P
O
SE
D
M
O
D
E
L
T
h
is
s
ec
tio
n
g
iv
es
d
etail
ab
o
u
t
th
e
s
u
g
g
ested
d
ee
p
tr
a
n
s
f
er
lear
n
in
g
-
b
ased
s
tr
ateg
y
to
cla
s
s
if
y
s
k
in
ca
n
ce
r
.
Sev
er
al
s
tep
s
co
m
p
r
is
e
th
e
en
tire
m
o
d
el,
in
clu
d
in
g
f
ea
tu
r
e
ex
tr
ac
tio
n
,
au
g
m
en
tati
o
n
,
p
r
e
-
p
r
o
ce
s
s
in
g
,
an
d
class
if
icatio
n
.
Fig
u
r
e
1
illu
s
tr
ates
th
e
g
en
er
al
ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
tr
a
n
s
f
er
lear
n
in
g
-
b
ase
d
m
eth
o
d
o
l
o
g
y
f
o
r
ca
te
g
o
r
izin
g
s
k
in
ca
n
ce
r
.
‒
Step
1
:
c
o
llectio
n
a
n
d
lo
ad
in
g
th
e
d
ataset.
I
n
t
h
is
s
tep
,
w
e
co
n
s
id
er
th
e
m
elan
o
m
a
r
elate
d
p
u
b
lically
av
ailab
le
d
ataset
f
o
r
p
r
o
ce
s
s
in
g
.
T
h
e
d
ataset
co
n
s
is
ts
o
f
i
m
ag
es,
th
eir
co
r
r
esp
o
n
d
in
g
g
r
o
u
n
d
tr
u
th
f
o
r
s
eg
m
en
tatio
n
an
d
lab
els f
o
r
cl
ass
es.
‒
Step
2
:
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
,
n
o
r
m
aliza
tio
n
an
d
lab
el
en
c
o
d
i
n
g
.
Gen
er
ally
,
th
e
s
k
in
ca
n
ce
r
im
ag
es
h
av
e
v
ar
ied
illu
m
in
atio
n
wh
ich
af
f
e
cts
th
e
im
ag
e
an
aly
s
is
task
s
.
T
h
er
ef
o
r
e
,
we
ap
p
ly
c
o
n
tr
ast
en
h
an
ce
m
en
t
t
o
o
b
tain
th
e
r
ef
in
e
d
d
e
r
m
o
s
co
p
y
im
ag
e.
M
o
r
eo
v
e
r
,
d
u
r
in
g
ca
p
t
u
r
in
g
t
h
ese
im
ag
es,
th
e
q
u
alit
y
o
f
im
ag
es
is
d
eg
r
ad
e
d
d
u
e
to
n
o
is
e
f
ac
to
r
.
T
o
o
v
er
c
o
m
e
th
is
is
s
u
e,
we
ad
o
p
t
im
ag
e
f
ilter
in
g
m
o
d
elto
r
em
o
v
e
t
h
e
n
o
is
e.
T
h
e
m
ela
n
o
m
a
a
f
f
ec
ted
d
ata
is
im
b
alan
ce
d
d
ata
th
er
e
f
o
r
e
we
also
in
c
o
r
p
o
r
ate
d
ata
au
g
m
en
tatio
n
m
eth
o
d
s
to
ad
d
r
ess
th
e
d
ata
im
b
alan
ce
is
s
u
e.
Fin
ally
,
we
a
p
p
ly
lab
el
en
co
d
i
n
g
m
ec
h
an
is
m
to
m
ak
e
it
co
m
p
atib
le
with
d
ee
p
lear
n
in
g
p
r
o
ce
s
s
in
g
m
o
d
u
les.
‒
S
t
e
p
3
:
f
e
at
u
r
e
e
x
t
r
a
ct
i
o
n
a
n
d
f
e
a
t
u
r
e
v
e
c
t
o
r
.
O
n
c
e
t
h
e
i
m
a
g
e
d
a
t
a
i
s
p
r
e
-
p
r
o
c
e
s
s
e
d
a
n
d
l
a
b
e
ls
a
r
e
e
n
c
o
d
ed
a
p
p
r
o
p
r
i
a
t
e
l
y
,
w
e
p
e
r
f
o
r
m
f
e
atu
r
e
e
x
t
r
a
c
ti
o
n
t
as
k
b
y
u
s
i
n
g
p
r
e
-
t
r
a
i
n
e
d
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
el
s
.
I
n
t
h
is
w
o
r
k
,
w
e
h
a
v
e
u
s
e
d
V
G
G
,
R
es
N
et
,
I
n
c
e
p
t
i
o
n
R
es
N
et
,
a
n
d
M
o
b
il
e
N
e
t
f
o
r
f
e
a
t
u
r
e
e
x
t
r
ac
t
i
o
n
b
y
u
s
i
n
g
t
h
e
i
r
p
r
e
-
t
r
a
i
n
e
d
w
e
i
g
h
t
s
.
T
h
e
o
b
t
ai
n
e
d
f
e
a
t
u
r
e
v
e
c
t
o
r
s
o
f
e
a
c
h
m
o
d
e
l
s
a
r
e
f
u
s
e
d
t
o
g
e
t
h
e
r
t
o
f
o
r
m
u
l
a
t
e
t
h
e
f
i
n
a
l
v
e
c
t
o
r
.
‒
Step
4
:
en
s
em
b
le
u
s
in
g
tr
an
s
f
er
lear
n
in
g
m
o
d
els
f
o
r
class
if
icatio
n
.
I
n
th
is
s
tep
th
e
r
esu
ltan
t
f
ea
tu
r
e
v
ec
to
r
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
tr
ain
in
g
p
r
o
ce
s
s
wh
ich
u
s
es
E
f
f
icien
tNet,
Xce
p
tio
n
a
n
d
Den
s
eNe
t
m
o
ce
ls
ar
e
u
s
ed
to
class
if
y
th
e
im
ag
e
d
ata.
3
.
1
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
3
.
1
.
1
.
VG
G
T
h
e
V
G
G
m
o
d
e
l
i
s
a
d
e
e
p
C
N
N
.
T
h
e
V
G
G
a
r
c
h
i
t
e
c
t
u
r
e
i
s
c
o
n
s
t
r
u
c
t
e
d
b
y
u
s
i
n
g
p
o
o
l
i
n
g
l
a
y
e
r
,
c
o
n
v
o
l
u
t
i
o
n
l
a
y
e
r
,
a
n
d
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
.
C
o
n
s
i
d
e
r
a
s
t
h
e
i
n
p
u
t
i
m
a
g
e
,
a
n
d
(
)
a
s
t
h
e
f
e
a
t
u
r
e
s
e
x
t
r
a
c
t
e
d
f
r
o
m
t
h
e
V
G
G
m
o
d
e
l
.
T
h
e
c
o
n
v
o
l
u
t
i
o
n
o
p
e
r
a
t
i
o
n
o
f
t
h
i
s
m
o
d
e
l
c
a
n
b
e
a
r
t
i
c
u
l
a
t
e
d
a
s
s
h
o
w
n
i
n
(
1
)
a
n
d
(
2
)
:
[
]
=
(
[
−
1
]
,
[
]
,
[
]
)
(
1
)
[
]
=
(
[
]
)
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mela
n
o
ma
cla
s
s
ifica
tio
n
u
s
in
g
en
s
emb
le
d
ee
p
tr
a
n
s
fer lea
r
n
in
g
(
S
o
u
mya
G
a
d
a
g
)
4947
W
h
er
e
d
e
p
icts
th
e
in
d
ex
o
f
t
h
e
lay
er
,
[
−
1
]
r
ep
r
esen
ts
th
e
ac
tiv
atio
n
f
r
o
m
t
h
e
p
r
ev
io
u
s
lay
er
,
[
]
an
d
[
]
ar
e
th
e
weig
h
ts
an
d
b
iases
o
f
ℎ
co
n
v
o
lu
tio
n
al
la
y
er
,
r
esp
ec
tiv
ely
an
d
[
]
r
ep
r
esen
ts
th
e
o
u
tp
u
t
o
f
co
n
v
o
l
u
tio
n
o
p
er
atio
n
.
T
h
e
M
ax
Po
o
lin
g
o
p
er
atio
n
ca
n
b
e
p
e
r
f
o
r
m
e
d
as sh
o
wn
in
(
3
)
:
[
]
=
(
(
)
)
(
3
)
Fin
ally
,
in
(
4
)
to
(
6
)
s
h
o
ws th
e
o
p
er
atio
n
o
f
t
h
e
f
u
lly
c
o
n
n
ec
t
ed
lay
er
:
=
(
[
]
)
(
4
)
[
+
1
]
=
.
[
+
1
]
+
[
+
1
]
(
5
)
(
)
=
(
[
+
1
]
(
6
)
Fig
u
r
e
1
.
Ar
c
h
itectu
r
e
o
f
p
r
o
p
o
s
ed
m
o
d
el
3
.
1
.
2
.
ResNet
m
o
del
A
d
ee
p
C
NN
ar
ch
itectu
r
e
ca
lled
r
esid
u
al
n
etwo
r
k
(
R
esNet
)
was
cr
ea
ted
to
s
o
lv
e
t
h
e
is
s
u
e
o
f
v
an
is
h
in
g
g
r
ad
ien
ts
.
I
n
s
tead
o
f
teac
h
in
g
th
e
n
etwo
r
k
th
e
u
n
d
er
ly
i
n
g
m
ap
p
in
g
s
d
ir
ec
tly
,
it
in
tr
o
d
u
ce
s
s
k
ip
co
n
n
ec
tio
n
s
,
wh
ic
h
en
a
b
le
th
e
n
etwo
r
k
t
o
lear
n
r
esid
u
al
f
u
n
ctio
n
s
.
T
h
is
en
a
b
les
tr
ain
in
g
o
f
m
u
ch
d
ee
p
er
n
etwo
r
k
s
with
im
p
r
o
v
e
d
p
er
f
o
r
m
an
ce
.
I
n
o
r
d
e
r
to
p
er
f
o
r
m
th
e
f
ea
tu
r
e
ex
tr
ac
tio
n
task
f
o
r
m
elan
o
m
a
im
ag
es,
let
u
s
co
n
s
id
er
th
at
in
p
u
t
i
m
ag
e
is
d
en
o
ted
as
an
d
th
e
f
ea
tu
r
es
ex
tr
ac
ted
f
r
o
m
th
e
R
esNet
m
o
d
el
a
s
(
)
.
Fo
r
f
ea
tu
r
e
ex
tr
ac
ti
o
n
,
th
is
m
o
d
el
u
s
es
co
n
v
o
l
u
tio
n
,
r
es
id
u
al
b
l
o
ck
,
p
o
o
lin
g
lay
er
,
a
n
d
f
u
lly
co
n
n
ec
ted
la
y
er
.
T
h
e
(
7
)
an
d
(
8
)
ex
p
r
ess
R
esNet
's co
n
v
o
lu
ti
o
n
lay
er
ac
tio
n
as:
[
]
=
(
[
−
1
]
,
[
]
,
[
]
)
(
7
)
[
]
=
ℎ
(
(
[
]
)
)
(
8
)
W
h
er
e
r
ep
r
esen
ts
lay
er
’
s
in
d
ex
,
[
−
1
]
r
ep
r
esen
ts
th
e
ac
tiv
atio
n
f
r
o
m
t
h
e
p
r
e
v
io
u
s
lay
er
,
[
]
d
en
o
tes
ℎ
weig
h
t
co
n
v
o
lu
tio
n
lay
er
,
[
]
d
en
o
tes
b
iases
o
f
ℎ
co
n
v
o
lu
tio
n
al
lay
er
,
an
d
[
]
d
ep
icts
o
u
tp
u
t
o
f
co
n
v
o
l
u
tio
n
o
p
er
atio
n
.
I
n
n
ex
t
p
h
ase,
it p
er
f
o
r
m
s
r
esid
u
al
b
lo
ck
o
p
er
atio
n
wh
ic
h
ca
n
b
e
w
r
itten
as (
9
)
:
[
]
=
[
−
1
]
+
(
[
−
1
]
,
[
]
)
(
9
)
W
h
er
e
r
ep
r
esen
ts
th
e
r
esid
u
al
f
u
n
ctio
n
to
b
e
lear
n
ed
b
y
th
e
r
esid
u
al
b
lo
ck
,
[
−
1
]
r
ep
r
esen
t
s
th
e
in
p
u
t
ac
tiv
atio
n
to
th
e
r
esid
u
al
b
lo
c
k
.
T
h
e
ad
d
itio
n
o
p
er
atio
n
d
en
o
tes
th
e
s
k
ip
co
n
n
ec
tio
n
,
allo
win
g
th
e
n
etwo
r
k
to
lear
n
r
esid
u
al
f
u
n
ctio
n
s
.
L
ater
,
p
o
o
lin
g
o
p
e
r
atio
n
is
p
er
f
o
r
m
e
d
as (
1
0
)
:
[
]
=
(
[
]
)
(
1
0
)
W
h
er
e
[
]
is
th
e
o
u
tco
m
e
o
f
m
ax
p
o
o
lin
g
o
p
er
atio
n
.
Fin
a
lly
,
th
e
f
u
lly
co
n
n
ec
ted
lay
e
r
o
p
er
atio
n
s
ar
e
p
er
f
o
r
m
ed
as (
1
1
)
to
(
1
3
)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
9
4
3
-
4
9
5
6
4948
=
(
[
]
)
(
1
1
)
[
+
1
]
=
.
[
+
1
]
+
[
+
1
]
(
1
2
)
(
)
=
(
[
+
1
]
)
(
1
3
)
3
.
1
.
3
.
I
ncept
io
n Re
s
Net
T
h
e
I
n
ce
p
tio
n
R
esNet
ar
ch
itectu
r
e
co
m
b
i
n
es
I
n
ce
p
tio
n
m
o
d
u
les
an
d
r
esid
u
al
co
n
n
e
ctio
n
s
f
r
o
m
R
esNet,
r
esu
ltin
g
ac
cu
r
ate
an
d
ef
f
icien
t
C
NN
.
I
t
in
ten
d
s
to
u
s
e
th
e
ad
v
a
n
tag
es
o
f
b
o
th
ar
ch
itectu
r
es
to
ac
h
iev
e
im
p
r
o
v
ed
p
er
f
o
r
m
an
c
e
o
n
v
a
r
io
u
s
co
m
p
u
ter
v
is
io
n
task
s
.
T
h
is
m
o
d
el
co
n
s
is
ts
o
f
in
ce
p
tio
n
m
o
d
u
le,
r
esid
u
al
co
n
n
ec
tio
n
,
p
o
o
lin
g
lay
er
,
an
d
f
u
lly
c
o
n
n
ec
te
d
la
y
er
.
T
h
e
I
n
ce
p
tio
n
m
o
d
u
le
h
as
v
ar
io
u
s
p
ar
allel
co
n
v
o
l
u
tio
n
al
b
r
a
n
ch
es
with
d
if
f
er
en
t
k
e
r
n
el
s
izes
an
d
p
o
o
lin
g
o
p
er
atio
n
s
.
E
ac
h
b
r
a
n
ch
ca
p
tu
r
es
f
ea
tu
r
es
at
d
if
f
er
en
t scale
s
an
d
r
eso
lu
tio
n
s
.
I
t c
an
b
e
r
e
p
r
esen
ted
as (
1
4
)
:
(
[
−
1
]
)
=
[
ℎ
1
,
ℎ
2
,
…
,
ℎ
]
(
1
4
)
W
h
er
e
[
−
1
]
is
ac
t
iv
atio
n
f
r
o
m
p
r
ev
io
u
s
lay
er
.
Similar
to
R
es
Net,
I
n
ce
p
tio
n
R
esNet
in
co
r
p
o
r
ates
r
esid
u
al
co
n
n
ec
tio
n
s
with
in
its
ar
ch
itectu
r
e
to
f
ac
ilit
ate
tr
ain
in
g
o
f
v
er
y
d
ee
p
n
etwo
r
k
s
.
T
h
is
o
p
er
atio
n
ca
n
b
e
ex
p
r
ess
ed
as (
1
5
)
:
=
[
−
1
]
+
(
[
−
1
]
)
(
1
5
)
W
h
er
e
r
ep
r
esen
ts
th
e
r
esid
u
al
f
u
n
ctio
n
to
b
e
lear
n
e
d
b
y
t
h
e
r
esid
u
al
co
n
n
ec
tio
n
.
I
n
n
ex
t
s
tep
,
we
ap
p
ly
p
o
o
lin
g
o
p
er
atio
n
s
im
ilar
to
R
esNet
an
d
VGG
m
o
d
el
as sh
o
wn
in
(
1
6
)
:
[
]
=
(
[
]
)
(
1
6
)
Fin
ally
,
it
u
s
es
f
u
lly
c
o
n
n
ec
te
d
lay
er
o
p
er
ati
o
n
s
as
d
is
cu
s
s
ed
in
VGG
an
d
R
esNet
m
o
d
els
wh
ich
is
ex
p
r
ess
ed
as (
1
7
)
to
(
1
9
)
:
=
(
[
]
)
(
1
7
)
[
+
1
]
=
.
[
+
1
]
+
[
+
1
]
(
1
8
)
(
)
=
(
[
+
1
]
)
(
1
9
)
3
.
1
.
4
.
M
o
bil
eNe
t
Mo
b
ileNet
is
a
lig
h
tweig
h
t
C
NN
ar
ch
itectu
r
e
th
at
u
tili
ze
s
d
ep
th
wis
e
s
ep
ar
a
b
le
co
n
v
o
lu
tio
n
s
.
I
t
h
elp
s
to
m
i
n
im
ize
th
e
r
eq
u
ir
ed
n
u
m
b
er
o
f
p
ar
am
eter
s
f
o
r
tr
ain
in
g
r
esu
ltin
g
i
n
r
ed
u
ce
d
co
m
p
u
tatio
n
tim
e
an
d
m
ain
tain
in
g
h
ig
h
ac
cu
r
a
cy
.
L
et'
s
d
en
o
te
th
e
in
p
u
t
im
ag
e
as
,
an
d
th
e
f
ea
tu
r
es
ex
tr
ac
ted
f
r
o
m
th
e
Mo
b
ileNet
m
o
d
el
as
(
)
.
T
h
e
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
in
clu
d
es
d
ep
th
wis
e
an
d
p
o
in
t
wis
e
co
n
v
o
l
u
tio
n
o
p
er
atio
n
s
.
T
h
e
d
ep
th
wis
e
co
n
v
o
l
u
tio
n
ca
n
b
e
ar
ticu
lated
as (
2
0
)
an
d
(
2
1
)
:
[
]
=
ℎ
(
[
−
1
]
,
[
]
,
[
]
)
(
2
0
)
[
]
=
ℎ
(
(
[
]
)
)
(
2
1
)
W
h
er
e
r
ep
r
esen
ts
lay
er
’
s
in
d
ex
,
[
−
1
]
r
ep
r
esen
ts
th
e
ac
tiv
atio
n
f
r
o
m
t
h
e
p
r
e
v
io
u
s
lay
er
,
[
]
d
en
o
tes
ℎ
weig
h
t
c
o
n
v
o
lu
tio
n
lay
er
,
[
]
d
en
o
tes
b
iases
o
f
ℎ
co
n
v
o
l
u
tio
n
al
lay
er
,
an
d
[
]
r
ep
r
esen
ts
th
e
o
u
t
p
u
t
o
f
d
ep
th
wis
e
co
n
v
o
l
u
tio
n
o
p
er
a
tio
n
.
T
h
e
n
e
x
t
s
tep
p
er
f
o
r
m
s
,
Po
in
twis
e
co
n
v
o
lu
tio
n
o
p
er
atio
n
wh
ic
h
is
ex
p
r
ess
ed
as (
2
2
)
an
d
(
2
3
)
:
[
+
1
]
=
(
[
]
,
[
+
1
]
,
[
+
1
]
)
(
2
2
)
(
)
=
(
[
+
1
]
)
(
2
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mela
n
o
ma
cla
s
s
ifica
tio
n
u
s
in
g
en
s
emb
le
d
ee
p
tr
a
n
s
fer lea
r
n
in
g
(
S
o
u
mya
G
a
d
a
g
)
4949
T
h
e
f
in
al
f
ea
tu
r
e
v
ec
to
r
ca
n
b
e
o
b
tain
ed
b
y
co
n
ca
ten
atin
g
t
h
ese
attr
ib
u
tes
an
d
n
o
r
m
alizin
g
th
e
c
o
n
ca
ten
ate
d
f
ea
tu
r
e
v
ec
to
r
.
I
t c
a
n
b
e
e
x
p
r
e
s
s
ed
as (
2
4
)
:
=
[
(
(
)
,
(
)
,
(
)
,
(
)
(
2
4
)
3
.
2
.
E
ns
em
ble
deep
t
ra
ns
f
er
lea
rning
m
o
del
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
e
n
s
em
b
le
m
o
d
el
o
f
d
ee
p
tr
a
n
s
f
er
lear
n
in
g
m
o
d
els
f
o
r
m
elan
o
m
a
class
if
icatio
n
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
es
s
tack
ed
en
s
em
b
le
o
f
E
f
f
icien
tNetB
0
,
Xce
p
tio
n
,
an
d
Den
s
eNe
t1
2
1
t
o
o
b
tain
h
e
class
if
icatio
n
p
r
o
b
a
b
ilit
y
v
ec
to
r
.
Fin
ally
,
th
e
f
u
lly
co
n
n
ec
ted
an
d
s
ig
m
o
id
la
y
er
s
ar
e
u
s
ed
to
o
b
tain
th
e
f
in
al
class
if
icatio
n
b
ased
o
n
th
e
co
n
ca
ten
atio
n
o
f
in
itial
p
r
o
b
ab
ilit
y
v
ec
to
r
.
I
n
F
ig
u
r
e
2
d
ep
icts
th
e
s
tack
ed
en
s
em
b
le
class
if
icatio
n
m
o
d
el
.
Fig
u
r
e
2
.
Pro
p
o
s
ed
d
ee
p
tr
an
s
f
er
lear
n
in
g
m
o
d
u
le
f
o
r
m
elan
o
m
a
class
if
icatio
n
E
f
f
icien
t
Net
is
a
f
am
ily
o
f
C
NN
ar
ch
itectu
r
es
th
at
h
av
e
b
ee
n
d
esig
n
e
d
to
ac
h
ie
v
e
th
e
im
p
r
o
v
e
d
p
er
f
o
r
m
an
ce
with
less
co
m
p
u
tatio
n
al
p
ar
am
eter
s
co
m
p
ar
e
d
to
o
th
er
ar
ch
itectu
r
es.
I
n
th
is
wo
r
k
,
we
h
av
e
u
s
ed
E
f
f
icien
tNetB
0
wh
ich
is
th
e
b
aselin
e
m
o
d
el
in
th
e
E
f
f
icie
n
tNet
f
am
ily
.
I
t
p
er
f
o
r
m
s
co
n
v
o
lu
tio
n
o
p
er
atio
n
s
,
d
ep
th
wis
e
s
ep
a
r
ab
le
c
o
n
v
o
l
u
tio
n
,
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
an
d
f
u
lly
c
o
n
n
e
cted
o
p
er
atio
n
.
T
h
e
f
ea
t
u
r
e
ex
tr
ac
tio
n
m
o
d
el
ca
n
b
e
ar
ticu
lated
as (
2
5
)
a
n
d
(
2
6
)
:
[
]
=
(
[
−
1
]
,
[
]
,
[
]
)
(
2
5
)
[
]
=
ℎ
(
ℎ
(
[
]
)
)
(
2
6
)
W
h
er
e
r
ep
r
esen
ts
lay
er
’
s
in
d
ex
,
[
−
1
]
r
ep
r
esen
ts
th
e
ac
tiv
atio
n
af
ter
ap
p
ly
in
g
b
atch
n
o
r
m
aliz
atio
n
an
d
th
e
s
wis
h
ac
tiv
atio
n
f
u
n
ctio
n
,
[
]
is
weig
h
ts
,
[
]
is
b
iases
o
f
ℎ
co
n
v
o
lu
tio
n
al
lay
er
,
an
d
[
]
d
e
n
o
tes
th
e
o
u
tp
u
t
o
f
co
n
v
o
lu
tio
n
o
p
e
r
atio
n
.
Ne
x
t step
,
it p
er
f
o
r
m
s
g
lo
b
al
av
e
r
a
g
e
p
o
o
li
n
g
(
GAP)
as in
(
2
7
)
:
=
(
[
]
)
(
2
7
)
Fin
ally
,
th
e
f
u
lly
c
o
n
n
ec
te
d
la
y
er
is
ap
p
lied
as (
2
8
)
a
n
d
(
2
9
)
:
]
+
1
]
=
.
[
+
1
]
+
[
+
1
]
(
2
8
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
9
4
3
-
4
9
5
6
4950
0
(
)
=
(
[
+
1
]
)
(
2
9
)
Her
e,
So
f
tm
ax
d
en
o
tes th
e
s
o
f
tm
ax
ac
tiv
atio
n
f
u
n
ctio
n
,
wh
ic
h
co
n
v
e
r
ts
th
e
r
aw
s
co
r
es in
to
class
p
r
o
b
ab
ilit
ies.
Similar
ly
,
Xce
p
tio
n
an
d
Den
s
en
et
m
o
d
els
also
u
s
ed
to
c
o
n
s
tr
u
ct
th
e
class
if
ier
wh
er
e
So
f
tm
ax
f
u
n
ctio
n
co
n
v
er
ts
th
e
r
aw
s
co
r
e
in
to
cla
s
s
p
r
o
b
ab
ilit
ies.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
A
d
etailed
s
tu
d
y
o
f
th
e
r
ec
o
m
m
en
d
ed
ap
p
r
o
ac
h
is
g
iv
e
n
i
n
th
is
s
ec
tio
n
.
T
h
e
d
ataset
u
tili
s
ed
in
th
is
wo
r
k
is
d
escr
ib
ed
in
d
ep
th
in
th
e
f
ir
s
t
s
u
b
s
ec
tio
n
.
T
h
e
n
ex
t
s
u
b
s
ec
tio
n
ex
p
lain
s
th
e
m
etr
ics
th
at
wer
e
u
s
ed
to
ass
es
s
th
e
p
r
o
p
o
s
ed
wo
r
k
'
s
p
er
f
o
r
m
a
n
ce
.
Fin
ally
,
a
co
m
p
ar
is
o
n
an
aly
s
is
o
f
p
r
o
p
o
s
ed
wo
r
k
with
ex
is
tin
g
tech
n
iq
u
es is
d
o
n
e
.
4
.
1
.
Da
t
a
s
et
det
a
ils
R
esear
ch
o
n
th
e
class
if
icatio
n
o
f
m
elan
o
m
a
h
as
a
d
v
an
ce
d
s
ig
n
if
ican
tly
d
u
e
to
th
e
atte
n
tio
n
f
r
o
m
I
n
ter
n
atio
n
al
Sk
in
I
m
ag
in
g
C
o
llab
o
r
atio
n
(
I
SIC)
ch
allen
g
e
s
.
Hig
h
r
eso
lu
tio
n
s
k
in
lesi
o
n
p
ictu
r
e
co
llectio
n
s
with
ex
p
er
t
a
n
n
o
tatio
n
,
b
i
o
p
s
y
-
p
r
o
v
e
n
,
a
n
d
g
lo
b
al
in
f
o
r
m
atio
n
ar
e
m
ad
e
a
v
ailab
le
b
y
th
e
I
SIC.
T
h
e
o
r
g
an
is
atio
n
h
as
co
n
d
u
cted
a
n
n
u
al
s
k
in
lesi
o
n
c
h
allen
g
es
in
an
ef
f
o
r
t
to
b
o
o
s
t
r
esear
ch
er
en
g
ag
e
m
en
t
an
d
im
p
r
o
v
e
C
AD
alg
o
r
ith
m
d
etec
tio
n
.
T
ab
le
1
p
r
o
v
id
es a
n
o
v
er
v
iew
o
f
th
e
I
SIC d
atasets
f
r
o
m
2
0
1
6
to
2
0
2
0
.
‒
I
SIC
2
0
1
6
d
ataset:
th
er
e
ar
e
1
2
7
9
p
h
o
to
s
in
to
tal
in
th
is
d
at
aset,
9
0
0
s
am
p
les
ar
e
u
s
ed
f
o
r
tr
ain
in
g
an
d
3
7
9
ar
e
u
s
ed
f
o
r
test
in
g
.
B
o
th
th
e
tr
ain
in
g
a
n
d
test
in
g
s
ets h
a
v
e
ac
ce
s
s
to
th
e
g
r
o
u
n
d
tr
u
th
d
ata.
‒
I
SIC
2
0
1
7
d
ataset:
in
th
is
d
ata
s
et
2
6
0
0
im
ag
es
ar
e
av
ailab
le
o
f
wh
ich
2
0
0
0
ar
e
u
s
esd
f
o
r
n
etwo
r
k
test
in
g
an
d
6
0
0
f
o
r
tr
ai
n
in
g
.
T
h
is
d
at
aset
co
n
tain
s
th
e
g
r
o
u
n
d
tr
u
th
s
f
o
r
f
o
u
r
d
if
f
e
r
en
t
class
g
r
o
u
p
s
:
m
elan
o
m
a,
s
eb
o
r
r
h
o
eic
k
er
ato
s
is
,
n
ev
u
s
,
an
d
m
elan
o
m
a
n
e
v
u
s
‒
I
SIC
2
0
1
8
d
ataset:
th
is
lar
g
e
d
ataset
h
as
a
to
tal
o
f
1
1
,
5
2
7
p
h
o
to
s
,
o
f
wh
ich
1
0
,
0
1
5
wer
e
u
s
ed
f
o
r
n
etwo
r
k
tr
ain
i
n
g
an
d
th
e
r
e
m
a
in
in
g
1
5
1
2
f
o
r
n
etwo
r
k
test
in
g
an
d
p
e
r
f
o
r
m
an
ce
ev
al
u
atio
n
.
‒
I
SIC
2
0
1
9
d
ataset:
o
f
th
e
3
3
,
5
6
9
p
h
o
to
s
in
th
e
I
SIC
2
0
1
9
d
ataset,
th
er
e
ar
e
8
,
2
3
8
im
ag
es
in
th
e
test
in
g
s
et
an
d
2
5
,
3
3
1
im
ag
es
in
th
e
t
r
ain
in
g
s
et.
B
u
t
ju
s
t
th
e
lab
els
f
o
r
t
h
e
tr
ain
in
g
s
et
p
h
o
t
o
s
,
wh
ich
r
ep
r
esen
t
eig
h
t
class
es,
ar
e
in
clu
d
ed
in
th
is
d
ataset.
T
h
e
in
f
o
r
m
atio
n
ab
o
u
t
th
e
tr
ai
n
in
g
a
n
d
test
in
g
p
h
o
to
s
is
co
n
tain
ed
in
th
e
m
etad
ata.
T
h
e
tr
ain
in
g
m
etad
ata
co
n
tain
s
a
ll
p
er
tin
en
t
p
atien
t
in
f
o
r
m
atio
n
,
wh
er
ea
s
th
e
test
in
g
m
etad
ata
co
n
tain
s
th
e
a
g
e,
g
en
d
er
,
a
n
ato
m
ical
s
ite,
an
d
lesi
o
n
I
D
o
f
th
e
p
atien
t.
‒
I
SIC
2
0
2
0
d
ataset:
th
er
e
a
r
e
4
4
1
0
8
p
h
o
to
s
in
t
o
tal
in
th
is
d
a
taset;
3
3
1
2
6
ar
e
u
s
ed
f
o
r
tr
ain
in
g
an
d
1
0
9
8
2
ar
e
u
s
ed
f
o
r
test
in
g
.
T
h
e
g
r
o
u
n
d
tr
u
t
h
d
ata,
wh
ich
in
cl
u
d
es
l
esio
n
I
D,
g
e
n
d
er
,
ag
e,
p
atien
t
I
D,
d
iag
n
o
s
is
,
an
ato
m
ical
p
lace
,
an
d
b
e
n
ig
n
o
r
m
alig
n
an
t
s
tatu
s
,
is
s
u
p
p
lie
d
f
o
r
th
e
tr
ain
in
g
s
et,
j
u
s
t
lik
e
it
was
th
e
y
ea
r
b
ef
o
r
e.
‒
I
SIC
Kag
g
le:
in
t
h
is
wo
r
k
,
K
ag
g
le
d
ataset
o
f
s
k
in
lesi
o
n
s
wh
ich
is
p
u
b
licly
av
aila
b
le
f
r
o
m
th
e
I
SIC
lib
r
ar
y
,
is
u
s
ed
to
tr
ain
an
d
v
alid
ate
o
u
r
s
tack
in
g
e
n
s
em
b
le
m
o
d
el.
T
h
er
e
a
r
e
1
8
0
0
b
e
n
ig
n
a
n
d
1
4
9
7
m
alig
n
an
t
m
o
le
p
h
o
to
s
in
t
h
e
co
llectio
n
[
1
4
]
.
Up
o
n
cl
o
s
er
in
s
p
ec
tio
n
,
s
o
u
n
d
s
an
d
ar
tef
ac
ts
wer
e
d
is
co
v
er
ed
in
t
h
e
g
ath
er
e
d
s
k
i
n
lesi
o
n
p
h
o
t
o
s
.
W
e
u
s
ed
,
s
tan
d
ar
d
p
r
e
-
p
r
o
ce
s
s
in
g
m
eth
o
d
s
lik
e
s
ca
lin
g
,
n
o
r
m
alis
atio
n
,
n
o
is
e
r
ed
u
ctio
n
,
an
d
c
o
n
tr
ast en
h
a
n
ce
m
en
t,
to
ad
d
r
ess
th
is
.
T
h
e
p
ix
el
in
ten
s
ity
v
alu
e
r
a
n
g
e
o
f
ev
e
y
im
ag
e
is
n
o
r
m
ali
ze
d
s
et
to
[
0
,
1
]
.
T
h
e
im
ag
es
ar
e
th
en
u
n
if
o
r
m
ly
en
lar
g
e
d
to
2
2
4
b
y
2
2
4
d
im
en
s
io
n
s
.
T
h
e
s
am
p
le
im
ag
e
o
f
B
en
ig
n
an
d
Ma
lig
n
a
n
t
m
elan
o
m
a
ca
s
es
is
s
h
o
wn
in
Fig
u
r
es 3
(
a)
an
d
3
(
b
)
r
esp
ec
tiv
ely
,
th
ese
im
ag
es
ar
e
o
b
tain
ed
f
r
o
m
I
SIC c
h
alle
n
g
e
d
atasets
.
T
ab
le
1
.
Su
m
m
a
r
y
o
f
I
SIC 2
0
1
6
-
2
0
2
0
d
ataset
D
a
t
a
s
e
t
Tr
a
i
n
Te
st
To
t
a
l
I
S
I
C
2
0
1
6
9
0
0
3
7
9
1
2
7
9
I
S
I
C
2
0
1
7
2
0
0
0
6
0
0
2
6
0
0
I
S
I
C
2
0
1
8
1
0
0
1
5
1
5
1
2
1
1
5
2
7
I
S
I
C
2
0
1
9
2
5
3
3
1
8
2
3
8
3
3
5
6
9
I
S
I
C
2
0
2
0
3
3
1
2
6
1
0
9
8
2
4
4
1
0
8
4
.
2
.
P
er
f
o
r
m
a
nce
m
e
a
s
urem
ent
pa
ra
m
et
er
s
T
h
is
s
u
b
s
ec
tio
n
d
escr
ib
es
th
e
p
er
f
o
r
m
an
ce
m
ea
s
u
r
em
en
t
p
ar
am
eter
s
th
at
wer
e
u
s
ed
to
a
n
aly
ze
th
e
o
v
er
all
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
T
h
e
class
if
ier
's
p
er
f
o
r
m
an
ce
is
ev
alu
ated
u
s
in
g
th
e
co
n
f
u
s
io
n
m
atr
ix
,
wh
ich
d
is
p
lay
s
th
e
q
u
an
tity
o
f
tr
u
e
n
e
g
ativ
es
(
T
N)
,
an
d
f
alse
n
eg
ati
v
es
(
FN)
,
tr
u
e
p
o
s
itiv
es
(
T
P)
an
d
f
alse p
o
s
itiv
es (
FP
)
.
Fig
u
r
e
4
p
r
o
v
id
es a
b
asic illu
s
tr
atio
n
o
f
a
co
n
f
u
s
io
n
m
atr
ix
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mela
n
o
ma
cla
s
s
ifica
tio
n
u
s
in
g
en
s
emb
le
d
ee
p
tr
a
n
s
fer lea
r
n
in
g
(
S
o
u
mya
G
a
d
a
g
)
4951
(
a)
(
b
)
Fig
u
r
e
3
.
T
h
e
s
am
p
le
im
ag
e
o
f
m
elan
o
m
a
ca
s
es
f
o
r
(
a)
b
en
ig
n
s
am
p
les (
b
)
m
alig
n
an
t
s
am
p
l
es
Fig
u
r
e
4
.
R
ep
r
esen
tatio
n
o
f
c
o
n
f
u
s
io
n
m
atr
i
x
T
o
e
v
al
u
a
te
th
e
e
f
f
ec
ti
v
e
n
ess
o
f
o
u
r
p
r
o
p
o
s
e
d
m
et
h
o
d
,
w
e
e
m
p
lo
y
esta
b
l
is
h
e
d
m
et
r
ics
in
cl
u
d
in
g
ac
c
u
r
a
cy
,
F
-
m
ea
s
u
r
e
,
p
r
ec
is
i
o
n
,
f
alse
p
o
s
iti
v
e
r
ate
,
tr
u
e
p
o
s
i
t
iv
e
r
at
e.
T
h
ese
m
ea
s
u
r
es
d
ep
e
n
d
o
n
tr
u
e
p
o
s
itiv
e,
tr
u
e
n
eg
ativ
e,
f
alse
p
o
s
itiv
e,
a
n
d
f
alse
n
eg
ati
v
e
b
ein
g
d
is
tin
g
u
is
h
ed
.
I
n
th
e
c
o
n
tex
t
o
f
t
h
is
s
tu
d
y
,
T
P,
T
N,
F
P,
an
d
FN
r
e
p
r
ese
n
t
p
r
o
p
er
ly
r
ec
o
g
n
ize
d
m
a
li
g
n
a
n
t i
m
a
g
es,
c
o
r
r
e
ctl
y
c
lass
ed
b
en
ig
n
im
a
g
es
,
wr
o
n
g
l
y
ca
te
g
o
r
i
ze
d
b
e
n
i
g
n
im
ag
es,
a
n
d
in
c
o
r
r
e
ctl
y
cl
ass
i
f
ie
d
m
ali
g
n
a
n
t
i
m
a
g
es
,
r
esp
ec
t
iv
el
y
.
T
h
e
r
atio
o
f
T
P
to
th
e
to
tal
n
u
m
b
er
o
f
p
ictu
r
es
class
if
ied
as
m
alig
n
an
t
is
u
s
ed
to
ca
lcu
late
p
r
ec
is
io
n
is
s
h
o
wn
in
(
3
0
)
.
=
+
(
3
0
)
Div
id
e
th
e
to
tal
n
u
m
b
er
o
f
h
a
r
m
f
u
l p
ictu
r
es b
y
th
e
tr
u
e
p
o
s
itiv
e
r
ate
(
T
PR
)
as in
(
3
1
)
:
=
+
(
3
1
)
T
h
e
f
alse p
o
s
itiv
e
r
ate
(
FP
R
)
is
ca
lcu
lated
as in
(
3
2
)
:
=
+
(
3
2
)
T
h
e
d
ef
in
itio
n
o
f
ac
c
u
r
ac
y
is
th
e
p
r
o
d
u
ct
o
f
T
N
an
d
T
P d
iv
i
d
ed
b
y
o
v
e
r
all
p
ictu
r
es,
as wr
i
tten
in
(
3
3
)
:
=
+
+
+
+
(
3
3
)
F
-
m
ea
s
u
r
e
in
(
3
4
)
r
e
p
r
esen
ts
th
e
h
ar
m
o
n
ic
m
ea
n
o
f
p
r
ec
is
io
n
an
d
r
ec
all
:
−
=
2
×
×
(
+
)
(
3
4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
4
,
No
.
6
,
Dec
em
b
er
2
0
2
5
:
4
9
4
3
-
4
9
5
6
4952
4.
3
.
Co
m
pa
ra
t
iv
e
perf
o
r
m
a
nce
a
na
ly
s
is
T
h
is
s
u
b
s
ec
tio
n
s
h
o
ws
r
esu
lt
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
alo
n
g
with
a
p
er
f
o
r
m
a
n
ce
co
m
p
ar
is
o
n
b
etwe
en
th
e
o
b
tain
ed
r
esu
lts
an
d
cu
r
r
en
t
s
y
s
tem
s
.
T
ab
le
2
d
is
p
la
y
s
th
e
co
m
p
ar
is
o
n
an
aly
s
is
u
s
in
g
s
ev
er
al
d
ee
p
lear
n
in
g
-
b
ased
tech
n
iq
u
es
an
d
d
atasets
.
T
ab
le
2
s
h
o
ws
h
o
ws
well
th
e
p
r
o
p
o
s
ed
m
o
d
el
p
er
f
o
r
m
s
alo
n
g
s
id
e
cu
ttin
g
-
ed
g
e
m
elan
o
m
a
class
if
ier
s
.
A
d
ee
p
lear
n
in
g
m
o
d
el
was
p
r
esen
ted
in
[
2
5
]
with
a
class
if
icatio
n
ac
cu
r
ac
y
o
f
0
.
9
2
.
A
n
ew
VG
G
-
1
3
m
o
d
el
f
o
r
s
k
in
ca
n
ce
r
d
iag
n
o
s
is
was
g
iv
en
b
y
Gilan
i
et
a
l
.
[
2
6
]
,
a
n
d
it
ac
h
iev
ed
an
8
9
.
5
7
%
d
etec
tio
n
ac
cu
r
ac
y
.
B
ased
o
n
I
n
ce
p
t
io
n
-
V3
,
th
e
C
o
n
v
Net
m
o
d
el
in
tr
o
d
u
ce
d
in
[
2
7
]
f
o
cu
s
es
o
n
b
in
ar
y
class
if
icatio
n
o
f
s
k
in
co
n
d
itio
n
s
an
d
s
u
cc
ess
f
u
lly
d
if
f
er
en
tiates
b
etwe
en
b
en
ig
n
an
d
m
alig
n
an
t
ty
p
es
o
f
s
k
in
ca
n
ce
r
.
Ma
lik
et
a
l.
[
2
8
]
s
h
o
wca
s
ed
th
e
m
u
lti
-
class
if
icatio
n
o
f
s
k
i
n
lesi
o
n
s
u
s
in
g
2
D
s
u
p
er
p
ix
els
with
R
esNet
-
5
0
,
ac
h
iev
in
g
an
ac
cu
r
ac
y
o
f
8
5
.
5
0
%.
L
in
g
et
a
l.
[
2
9
]
ac
h
iev
ed
a
p
r
ec
is
io
n
o
f
8
8
.
1
0
% in
th
e
m
u
lti
-
class
if
icatio
n
o
f
s
k
in
ca
n
ce
r
.
Z
h
o
u
et
a
l.
[
3
0
]
p
r
esen
ted
SC
DNe
t a
n
d
ac
h
iv
ed
ac
cu
r
ac
y
o
f
9
2
.
8
9
%
in
class
if
icatio
n
o
f
s
k
in
ca
n
ce
r
.
I
n
co
n
tr
ast
to
estab
lis
h
ed
well
k
n
o
wn
m
et
h
o
d
s
,
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
d
em
o
n
s
tr
ates a
n
im
p
r
o
v
e
d
ac
c
u
r
ac
y
f
o
r
I
SIC 2
0
1
6
-
2
0
2
0
d
ata
s
ets.
T
ab
le
2
.
Data
s
et
co
m
p
ar
ati
v
e
an
aly
s
is
with
d
if
f
er
en
t d
ataset
s
A
r
t
i
c
l
e
Y
e
a
r
M
o
d
e
l
D
a
t
a
s
e
t
A
c
c
u
r
a
c
y
R
e
c
a
l
l
P
r
e
c
i
s
i
o
n
F1
-
sc
o
r
e
[
2
5
]
2
0
2
3
C
N
N
I
S
I
C
-
2
0
1
7
9
2
.
0
1
9
1
.
9
1
9
1
.
6
6
9
1
.
9
9
[
2
6
]
2
0
2
3
VGG
-
13
I
S
I
S
-
2
0
1
9
8
9
.
5
8
9
0
.
6
9
8
9
.
6
5
8
9
.
6
3
[
2
7
]
2
0
2
1
C
o
n
v
N
e
t
I
S
I
C
-
2
0
1
8
8
6
.
8
9
8
6
.
1
5
8
7
.
5
0
-
[
2
8
]
2
0
2
2
R
C
N
N
+
2
D
s
u
p
e
r
p
i
x
e
l
HAM
-
1
0
0
0
0
8
5
.
4
9
8
3
.
3
9
8
4
.
4
9
8
5
.
3
0
[
2
9
]
2
0
2
1
R
e
se
t
X
t
1
0
1
I
S
I
C
-
2
0
1
9
8
8
.
4
9
8
7
.
3
9
8
8
.
1
0
8
8
.
3
0
[
3
0
]
2
0
2
2
S
C
D
N
e
t
I
S
I
C
-
2
0
1
9
9
2
.
8
9
9
2
.
2
0
9
2
.
1
9
9
2
.
2
0
P
r
o
p
o
se
d
I
S
I
C
2
0
1
6
9
6
.
1
0
9
5
.
1
5
9
6
.
2
5
9
5
.
2
0
I
S
I
C
2
0
1
7
9
7
.
2
3
9
6
.
2
0
9
6
.
3
0
9
5
.
5
5
I
S
I
C
2
0
1
8
9
7
.
5
0
9
7
.
8
8
9
7
.
5
0
9
8
.
2
0
I
S
I
C
2
0
1
9
9
8
.
3
3
9
8
.
5
0
9
8
.
3
0
9
8
.
1
5
I
S
I
C
2
0
2
0
9
8
.
6
0
9
8
.
9
0
9
8
.
5
0
9
7
.
3
0
I
n
th
e
f
o
llo
win
g
ex
p
e
r
im
en
t,
we
co
m
p
ar
ed
th
e
ac
cu
r
ac
y
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
to
th
e
m
o
s
t
ad
v
an
ce
d
d
ee
p
lear
n
in
g
,
m
a
ch
in
e
lear
n
in
g
,
an
d
tr
a
n
s
f
er
lear
n
in
g
tech
n
iq
u
es.
T
h
e
T
ab
le
3
,
s
h
o
ws
th
e
co
m
p
ar
ativ
e
a
n
aly
s
is
f
o
r
HAM
-
1
0
0
0
0
.
As
d
is
cu
s
s
ed
b
ef
o
r
e,
th
e
tr
an
s
f
er
lear
n
in
g
m
o
d
el
s
h
av
e
g
ain
ed
h
u
g
e
atten
tio
n
in
th
is
b
io
m
e
d
ical
im
ag
in
g
d
o
m
ain
th
e
r
ef
o
r
e
s
ev
er
al
tr
an
s
f
er
lear
n
in
g
-
b
ased
m
o
d
els
h
av
e
b
ee
n
in
tr
o
d
u
ce
d
.
T
o
ass
ess
th
e
ef
f
icien
cy
o
f
t
h
ese
tr
an
s
f
er
lear
n
in
g
m
o
d
els,
we
ev
alu
ated
th
e
p
er
f
o
r
m
an
ce
f
o
r
Kag
g
le
I
SIC
d
ataset.
T
ab
le
4
d
em
o
n
s
tr
ates
th
e
o
u
tc
o
m
e
o
f
m
o
s
tly
u
s
ed
tr
an
s
f
er
lear
n
in
g
m
o
d
els
f
o
r
im
a
g
e
class
if
icatio
n
task
s
.
T
ab
le
3
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
f
o
r
HAM
1
0
0
0
0
d
ataset
A
r
t
i
c
l
e
Y
e
a
r
M
o
d
e
l
D
a
t
a
s
e
t
A
c
c
u
r
a
c
y
(
%)
[
3
1
]
2
0
2
0
A
l
e
x
N
e
t
HAM
-
1
0
0
0
0
84
[
3
2
]
2
0
1
9
M
o
b
i
l
e
N
e
t
HAM
-
1
0
0
0
0
8
3
.
9
[
3
3
]
2
0
2
0
M
o
b
i
l
e
N
e
t
,
V
G
G
-
16
HAM
-
1
0
0
0
0
8
0
.
6
1
[
3
4
]
2
0
1
9
S
V
M
HAM
-
1
0
0
0
0
7
4
.
7
5
[
3
5
]
2
0
2
0
R
e
sN
e
t
HAM
-
1
0
0
0
0
78
2
0
2
0
X
c
e
p
t
i
o
n
82
2
0
2
0
D
e
n
seN
e
t
82
[
3
6
]
2
0
2
0
C
N
N
HAM
-
1
0
0
0
0
77
[
3
7
]
2
0
2
1
M
o
b
i
l
e
N
e
t
a
n
d
LST
M
HAM
-
1
0
0
0
0
85
[
3
8
]
2
0
2
1
I
n
c
e
p
t
i
o
n
-
V3
HAM
-
1
0
0
0
0
8
9
.
7
3
[
3
9
]
2
0
2
3
I
n
c
e
p
t
i
o
n
R
e
s
n
e
t
-
V2
HAM
-
1
0
0
0
0
9
1
.
2
6
Tr
a
n
sf
e
r
Le
a
r
n
i
n
g
HAM
-
1
0
0
0
0
9
8
.
5
5
T
ab
le
4
.
Ov
e
r
all
p
er
f
o
r
m
an
ce
an
aly
s
is
f
o
r
Kag
g
le
I
SIC d
atas
et
M
o
d
e
l
Ac
c
u
r
a
c
y
c
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
S
p
e
c
i
f
i
c
i
t
y
F1
-
sc
o
r
e
AUC
R
e
sN
e
t
5
0
8
8
.
7
8
9
3
.
3
3
8
5
.
5
6
9
2
.
6
7
8
9
.
2
8
0
.
8
9
1
VGG
-
16
9
0
.
9
1
9
5
.
6
8
8
6
.
1
1
9
5
.
3
3
9
0
.
6
4
0
.
9
0
7
X
c
e
p
t
i
o
n
9
2
.
4
2
9
3
.
3
0
9
2
.
7
8
9
2
.
0
0
9
3
.
0
4
0
.
9
2
4
D
e
n
seN
e
t
1
2
1
9
2
.
2
7
9
1
.
8
7
9
4
.
1
7
9
0
.
0
0
9
3
.
0
0
0
.
9
2
1
Ef
f
i
c
i
e
n
t
N
e
t
B
O
9
2
.
3
0
9
4
.
0
2
9
1
.
6
7
9
3
.
0
0
9
2
.
8
3
0
.
9
2
3
P
r
o
p
o
se
d
M
o
d
e
l
9
8
.
7
6
9
8
.
6
0
9
7
.
6
7
9
5
.
6
7
9
7
.
1
3
0
.
9
8
7
Evaluation Warning : The document was created with Spire.PDF for Python.