I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Ro
bo
t
ics a
nd
Aut
o
m
a
t
io
n
(
I
J
RA)
Vo
l.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
,
p
p
.
1
36
~
1
48
I
SS
N:
2
7
2
2
-
2
5
8
6
,
DOI
: 1
0
.
1
1
5
9
1
/ijra
.
v
1
5
i
1
.
p
p
1
36
-
1
48
136
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
r
a
.
ia
esco
r
e.
co
m
M
ulti
-
mo
da
l
tran
sfo
rmer and
conv
o
lutiona
l attent
io
n
a
rchitec
t
ures
for
mela
no
ma
det
ecti
o
n in der
mo
sco
pi
c ima
g
es
G
uid
o
um
Am
ina
,
M
a
a
m
a
r
B
o
ug
hera
ra
,
Am
a
ra
Ra
f
i
k
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
H
i
g
h
e
r
N
o
r
m
a
l
S
c
h
o
o
l
o
f
K
o
u
b
a
,
A
l
g
i
e
r
s,
A
l
g
e
r
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
3
0
,
2
0
2
5
R
ev
is
ed
Dec
7
,
2
0
2
5
Acc
ep
ted
Feb
2
1
,
2
0
2
6
Th
e
d
e
a
d
li
e
st
ty
p
e
o
f
sk
i
n
c
a
n
c
e
r,
m
e
lan
o
m
a
,
re
q
u
ires
e
a
rly
a
n
d
a
c
c
u
ra
te
d
e
tec
ti
o
n
fo
r
a
s
u
c
c
e
ss
fu
l
c
o
u
rse
o
f
trea
tme
n
t.
Trad
it
io
n
a
l
d
iag
n
o
sti
c
tec
h
n
iq
u
e
s,
wh
ic
h
re
ly
o
n
v
isu
a
l
in
sp
e
c
ti
o
n
a
n
d
d
e
rm
o
sc
o
p
y
,
a
re
f
re
q
u
e
n
tl
y
a
rb
it
ra
ry
a
n
d
p
r
o
n
e
t
o
h
u
m
a
n
e
rro
r.
Au
t
o
m
a
ted
m
e
lan
o
m
a
d
e
tec
ti
o
n
e
x
e
m
p
li
fies
th
e
in
te
g
ra
ti
o
n
o
f
m
u
lt
ime
d
ia,
a
tru
l
y
i
n
terd
isc
ip
li
n
a
r
y
field
th
a
t
m
e
ld
s
v
isu
a
l
d
a
ta
p
ro
c
e
ss
in
g
,
h
u
m
a
n
-
c
o
m
p
u
ter
in
tera
c
ti
o
n
,
a
n
d
d
i
g
it
a
l
tec
h
n
o
l
o
g
ies
.
T
h
is
stu
d
y
p
re
se
n
ts
a
m
u
lt
i
-
m
o
d
a
l
a
rc
h
it
e
c
tu
re
:
a
m
u
lt
i
-
m
o
d
a
l
tr
a
n
sfo
rm
e
r
n
e
two
rk
(M
M
TN)
a
n
d
a
c
o
n
v
o
l
u
ti
o
n
a
l
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
m
u
lt
i
-
m
o
d
a
l
(CAMM
)
th
a
t
c
o
m
b
in
e
s
c
li
n
ica
l
d
a
ta
a
n
d
d
e
rm
o
sc
o
p
y
ima
g
e
s
to
e
n
h
a
n
c
e
m
e
lan
o
m
a
d
e
tec
ti
o
n
.
Th
e
m
o
d
e
ls
a
c
h
iev
e
h
ig
h
e
r
p
e
rfo
rm
a
n
c
e
c
o
m
p
a
re
d
to
o
t
h
e
r
a
p
p
ro
a
c
h
e
s
b
y
u
ti
li
z
i
n
g
th
e
stre
n
g
th
s
o
f
a
rc
h
it
e
c
tu
re
b
a
se
d
o
n
tra
n
sfo
rm
e
rs,
a
n
e
n
c
o
d
e
r
f
o
r
i
m
a
g
e
p
ro
c
e
ss
in
g
,
d
e
n
se
lay
e
rs
f
o
r
c
li
n
ica
l
d
a
ta als
o
S
p
a
ti
a
l
Atten
t
io
n
fo
r
t
h
e
se
c
o
n
d
a
rc
h
it
e
c
tu
re
p
r
o
p
o
se
d
.
W
e
e
v
a
lu
a
te
th
e
m
o
d
e
ls
o
n
t
h
e
e
n
ti
re
se
t
o
f
IS
IC
2
0
1
9
d
a
ta,
sh
o
wi
n
g
s
ig
n
ifi
c
a
n
t
imp
ro
v
e
m
e
n
ts
i
n
a
c
c
u
ra
c
y
a
n
d
A
UC.
Th
e
m
o
d
e
ls
a
c
h
iev
e
h
ig
h
a
c
c
u
ra
c
y
a
n
d
AU
C
u
sin
g
C
P
U
in
b
o
th
a
rc
h
it
e
c
tu
re
s.
Ou
r
fin
d
i
n
g
s
h
ig
h
li
g
h
t
th
e
p
o
ten
t
ial
o
f
a
m
u
lt
i
-
m
o
d
a
l
lea
rn
i
n
g
a
rc
h
it
e
c
tu
re
to
e
n
h
a
n
c
e
c
li
n
ica
l
d
e
c
isio
n
-
m
a
k
i
n
g
a
n
d
d
ia
g
n
o
stic
a
c
c
u
ra
c
y
in
d
e
rm
a
to
lo
g
y
.
T
o
o
u
r
k
n
o
wle
d
g
e
,
t
h
is
is
th
e
first
imp
lem
e
n
tatio
n
c
o
m
b
in
i
n
g
M
o
b
i
leN
e
t,
tran
sfo
rm
e
r
e
n
c
o
d
e
r
a
tt
e
n
ti
o
n
,
a
n
d
c
li
n
ica
l
d
a
ta
fu
si
o
n
fo
r
th
e
IS
IC
2
0
1
9
d
a
tas
e
t,
p
r
o
v
i
d
i
n
g
a
sig
n
ifi
c
a
n
t
a
d
v
a
n
c
e
m
e
n
t
in
th
e
a
u
t
o
m
a
ted
c
a
t
e
g
o
riza
ti
o
n
o
f
sk
i
n
m
a
li
g
n
a
n
c
ies
.
K
ey
w
o
r
d
s
:
Ar
ch
itectu
r
es
C
lin
ical
d
ata
f
u
s
io
n
E
n
co
d
er
atten
tio
n
I
m
ag
e
p
r
o
ce
s
s
in
g
Mu
lti
-
m
o
d
al
T
r
an
s
f
o
r
m
e
r
n
etwo
r
k
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
:
Gu
id
o
u
m
Am
in
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
,
Hig
h
er
N
o
r
m
al
Sch
o
o
l o
f
Ko
u
b
a
Alg
ier
s
,
Alg
er
ia
E
m
ail: a
m
in
a.
g
u
id
o
u
m
@
g
.
en
s
-
k
o
u
b
a.
d
z
1.
I
NT
RO
D
UCT
I
O
N
Sk
in
ca
n
ce
r
,
esp
ec
ially
m
elan
o
m
a,
h
as
a
h
ig
h
d
ea
th
r
ate
if
l
ef
t
u
n
tr
ea
te
d
,
it
p
r
esen
ts
a
s
er
io
u
s
p
u
b
lic
h
ea
lth
co
n
ce
r
n
.
T
r
ad
itio
n
al
d
i
ag
n
o
s
tic
tech
n
iq
u
es,
s
u
ch
as
d
er
m
o
s
co
p
y
an
d
v
is
u
al
ex
am
i
n
atio
n
,
r
ely
h
ea
v
ily
o
n
th
e
d
e
r
m
ato
lo
g
is
t
’
s
s
u
b
jectiv
e
in
ter
p
r
etatio
n
.
T
h
is
ca
n
r
esu
lt
in
u
n
d
iag
n
o
s
ed
ca
s
es
an
d
v
ar
ia
b
ilit
y
i
n
d
iag
n
o
s
tic
r
esu
lts
b
etwe
en
p
r
ac
titi
o
n
er
s
.
R
ec
en
t
d
ev
elo
p
m
en
ts
in
im
ag
e
p
r
o
ce
s
s
in
g
an
d
ar
t
if
icial
in
tellig
en
ce
h
av
e
s
h
o
w
n
p
r
o
m
is
e
in
au
to
m
atin
g
an
d
in
c
r
ea
s
in
g
t
h
e
p
r
ec
is
io
n
o
f
s
k
in
ca
n
ce
r
d
etec
tio
n
,
esp
ec
ially
in
t
h
e
ar
ea
o
f
m
e
d
ical
im
ag
in
g
.
I
m
a
g
e
p
r
o
ce
s
s
in
g
r
e
p
r
esen
ts
a
s
p
ec
ialized
d
o
m
ain
with
in
m
u
ltime
d
ia,
ly
in
g
at
th
e
in
ter
s
ec
tio
n
o
f
co
m
p
u
ter
s
cien
ce
,
elec
tr
o
n
ics,
an
d
v
is
u
al
s
cien
ce
s
[
1
]
,
[
2
]
.
T
h
is
f
ield
h
ar
n
ess
es
ad
v
an
ce
d
co
m
p
u
tatio
n
al
m
eth
o
d
s
an
d
e
lectr
o
n
ic
tech
n
iq
u
es
to
an
aly
ze
,
en
h
an
ce
,
an
d
tr
an
s
f
o
r
m
v
is
u
al
d
ata,
th
er
eb
y
s
u
p
p
o
r
tin
g
a
wid
e
ar
r
ay
o
f
ap
p
licatio
n
s
f
r
o
m
m
ed
ical
im
a
g
in
g
to
d
ig
ital e
n
ter
tai
n
m
en
t.
Au
to
m
atin
g
th
is
d
iag
n
o
s
tic
p
r
o
ce
s
s
is
th
er
ef
o
r
e
cr
u
cial
to
e
m
p
o
wer
in
g
d
er
m
ato
lo
g
is
ts
an
d
en
s
u
r
in
g
ea
r
ly
an
d
s
y
s
tem
atic
s
cr
ee
n
in
g
.
W
h
ile
d
ee
p
lear
n
in
g
,
an
d
in
p
ar
ticu
lar
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
(
C
NNs)
[
3
]
,
h
a
v
e
s
h
o
wn
p
r
o
m
is
e
f
o
r
a
n
aly
zin
g
d
er
m
o
s
co
p
ic
im
ag
es,
th
ese
u
n
im
o
d
al
ap
p
r
o
ac
h
es
[
4
]
–
[
7
]
o
f
ten
n
e
g
lect
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S I
n
t J Ro
b
&
Au
to
m
I
SS
N:
2722
-
2
5
8
6
Mu
lti
-
mo
d
a
l tra
n
s
fo
r
mer a
n
d
co
n
vo
lu
tio
n
a
l a
tten
tio
n
a
r
ch
it
ec
tu
r
es fo
r
mela
n
o
ma
…
(
Gu
i
d
o
u
m
A
min
a
)
137
ess
en
tial
co
n
tex
tu
al
in
f
o
r
m
ati
o
n
p
r
o
v
id
e
d
b
y
p
atien
t
m
etad
ata,
s
u
ch
as
ag
e
an
d
s
ex
,
wh
i
ch
ar
e
k
n
o
w
n
r
is
k
f
ac
to
r
s
f
o
r
m
elan
o
m
a.
T
o
o
v
e
r
co
m
e
th
is
lim
itatio
n
,
we
p
r
o
p
o
s
e
a
n
o
v
el
m
u
ltimo
d
al
lear
n
in
g
f
r
am
ewo
r
k
th
a
t
s
y
n
er
g
is
tically
f
u
s
es
v
is
u
al
p
atter
n
s
f
r
o
m
d
e
r
m
o
s
co
p
ic
im
ag
es
with
s
tr
u
ctu
r
ed
clin
ic
al
d
ata.
Ou
r
m
ain
co
n
tr
ib
u
tio
n
lies
in
th
e
in
tr
o
d
u
ctio
n
an
d
co
m
p
ar
ativ
e
e
v
alu
atio
n
o
f
two
o
r
ig
i
n
al
an
d
d
is
tin
ct
ar
ch
itectu
r
es
d
esig
n
ed
f
o
r
th
is
f
u
s
io
n
:
a
m
u
ltimo
d
al
tr
an
s
f
o
r
m
er
n
etw
o
r
k
(
MM
T
N)
an
d
a
m
u
ltimo
d
al
n
etwo
r
k
with
a
co
n
v
o
l
u
tio
n
al
atten
tio
n
m
ec
h
an
is
m
(
C
AM
M)
.
T
h
ese
m
o
d
els
ar
e
s
p
ec
if
ically
d
esig
n
ed
to
lev
er
ag
e
th
e
co
m
p
lem
en
tar
y
s
tr
en
g
th
s
o
f
i
m
ag
in
g
an
d
clin
ical
d
ata,
g
o
in
g
b
e
y
o
n
d
u
n
im
o
d
al
an
al
y
s
is
to
o
f
f
er
m
o
r
e
h
o
lis
tic
an
d
au
to
m
ated
d
iag
n
o
s
tic
s
u
p
p
o
r
t.
Ou
r
co
n
tr
i
b
u
tio
n
s
in
cl
u
d
e:
a
n
o
v
el
m
u
ltimo
d
al
t
r
an
s
f
o
r
m
er
ar
ch
itectu
r
e
th
at
u
s
es
i
m
ag
e
d
ata
au
g
m
en
tatio
n
v
ia
I
m
a
g
eDa
taGe
n
er
ato
r
(
h
o
r
izo
n
tal
f
lip
,
s
h
ea
r
,
zo
o
m
,
s
h
if
ts
,
r
o
tatio
n
s
)
an
d
p
r
e
p
r
o
ce
s
s
in
g
(
R
GB
v
alu
e
n
o
r
m
aliza
tio
n
an
d
r
esizin
g
)
f
o
r
im
a
g
e
p
r
o
ce
s
s
in
g
;
cu
s
to
m
d
ata
g
en
er
ato
r
s
to
co
m
b
in
e
im
ag
e
b
atch
es
with
clin
ical
d
ata
(
a
g
e,
s
ex
)
;
two
-
in
p
u
t
m
u
ltimo
d
al
lear
n
in
g
th
at
m
er
g
es
v
is
u
al
(
Mo
b
ileNet)
an
d
clin
ical
f
ea
tu
r
es;
an
d
a
cu
s
to
m
atten
tio
n
m
ec
h
a
n
is
m
(
T
r
a
n
s
f
o
r
m
er
B
lo
ck
)
with
f
o
u
r
m
u
lti
-
h
ea
d
atten
tio
n
al
elem
en
ts
,
lay
er
n
o
r
m
aliza
tio
n
,
an
d
a
f
ee
d
-
f
o
r
wa
r
d
n
etwo
r
k
.
T
h
is
h
y
b
r
id
ap
p
r
o
ac
h
co
m
b
in
es
p
r
e
-
tr
ain
e
d
C
NNs,
s
tr
u
ctu
r
ed
d
ata,
an
d
atten
tio
n
al
m
ec
h
an
is
m
s
.
Fo
r
th
e
s
ec
o
n
d
p
r
o
p
o
s
ed
ar
ch
itec
tu
r
e
(
C
AM
M)
,
o
u
r
co
n
tr
ib
u
tio
n
s
ar
e:
s
tr
atif
ied
cla
s
s
im
b
alan
ce
m
an
ag
em
en
t; a
m
u
ltimo
d
al
ar
ch
itectu
r
e
u
s
in
g
ch
an
n
el
an
d
s
p
atial
atten
tio
n
to
f
o
c
u
s
o
n
r
ele
v
a
n
t
lesi
o
n
s
,
wh
o
s
e
clin
ical
f
u
s
io
n
(
a
g
e/sex
)
im
p
r
o
v
es
co
n
tex
tu
aliza
tio
n
a
n
d
in
cr
ea
s
es
th
e
a
r
ea
u
n
d
er
th
e
R
OC
cu
r
v
e
(
AUC
)
[
8
]
b
y
+8
%
co
m
p
ar
e
d
to
p
u
r
ely
v
is
u
al
m
o
d
els;
an
d
r
o
b
u
s
t
g
en
er
aliza
tio
n
t
h
r
o
u
g
h
d
ata
a
u
g
m
en
tatio
n
(
co
n
tr
ast,
b
r
ig
h
t
n
ess
,
f
lip
p
in
g
)
cr
ea
tin
g
ar
tifi
cial
v
ar
iab
ilit
y
,
an
d
is
o
to
n
ic
ca
lib
r
atio
n
alig
n
in
g
p
r
ed
ictio
n
s
with
clin
ical
r
ea
lity
,
m
ain
tain
in
g
a
c
o
n
s
tan
t
test
A
UC
o
f
0
.
8
7
d
esp
ite
th
e
co
m
p
lex
ity
o
f
th
e
I
SIC d
at
a.
Ma
u
r
y
a
et
a
l.
[
9
]
in
tr
o
d
u
c
es
Du
alAu
to
E
L
M,
AI
p
o
wer
ed
m
eth
o
d
d
esig
n
ed
to
en
h
an
ce
th
e
ca
teg
o
r
izatio
n
o
f
d
if
f
e
r
en
t
ty
p
es
o
f
s
k
in
ca
n
ce
r
.
T
h
e
p
r
o
p
o
s
e
d
tech
n
iq
u
e
u
s
es
a
d
u
al
au
to
e
n
co
d
er
ar
c
h
itectu
r
e,
an
d
a
f
ast
Fo
u
r
ie
r
tr
an
s
f
o
r
m
(
FFT
)
au
to
en
co
d
er
t
h
at
ex
am
i
n
es
tex
tu
r
al
d
etails
an
d
f
r
eq
u
en
cy
p
atter
n
s
u
s
in
g
FFT
tr
an
s
f
o
r
m
ed
im
ag
e
r
e
co
n
s
tr
u
ctio
n
.
th
e
f
r
am
ew
o
r
k
h
as
b
ee
n
test
ed
o
n
th
e
p
u
b
licly
ac
c
ess
ib
le
HAM
1
0
0
0
0
[
1
0
]
.
T
h
e
m
o
d
el
’
s
ac
c
u
r
ac
y
an
d
p
r
ec
is
io
n
f
o
r
HAM
1
0
0
0
0
a
n
d
I
SIC
2
0
1
7
ar
e
9
7
.
6
8
%
an
d
9
7
.
6
6
%,
r
esp
ec
tiv
ely
,
an
d
8
6
.
7
5
% a
n
d
8
6
.
6
8
%,
r
esp
ec
tiv
ely
.
Usi
n
g
tr
im
o
d
al
cr
o
s
s
atten
ti
o
n
,
wh
ich
co
m
b
in
es
th
e
im
ag
e
an
d
m
etad
ata
m
o
d
alities
at
v
ar
io
u
s
tr
an
s
f
o
r
m
er
en
c
o
d
er
f
ea
tu
r
e
le
v
els.
W
ith
a
m
ea
n
d
iag
n
o
s
tic
ac
cu
r
ac
y
o
f
7
7
.
8
5
%
a
n
d
a
m
e
an
av
e
r
ag
e
a
cc
u
r
ac
y
o
f
7
7
.
2
7
% o
n
th
e
p
u
b
licly
ac
c
ess
ib
le
Der
m
7
p
t d
ataset
[
1
1
]
.
Usi
n
g
E
f
f
icien
tNet
m
o
d
els
o
n
th
e
HAM
1
0
0
0
0
d
ataset,
wh
i
ch
co
n
tain
s
d
er
m
o
s
co
p
y
im
ag
es
o
f
s
k
in
lesi
o
n
s
,
Ali
et
a
l.
[
1
2
]
p
r
o
p
o
s
es
a
m
u
lticlas
s
cla
s
s
if
icatio
n
t
ec
h
n
iq
u
e
f
o
r
s
k
in
ca
n
ce
r
s
.
T
o
s
atis
f
y
th
e
n
ee
d
s
o
f
E
f
f
icien
tNet
m
o
d
els,
th
e
au
th
o
r
s
h
av
e
d
e
v
elo
p
e
d
a
p
ip
elin
e
th
at
r
esizes
p
h
o
to
s
,
elim
in
ate
s
im
ag
e
p
ix
els,
an
d
ex
p
an
d
s
th
e
d
ata
s
et
(
r
o
tatio
n
,
zo
o
m
,
an
d
h
o
r
izo
n
tal/v
er
tical
r
etu
r
n
)
.
Pre
-
e
n
ter
ed
weig
h
ts
o
n
I
m
a
g
eNe
t
wer
e
u
s
ed
to
tr
ain
th
e
E
f
f
icien
tNet
m
o
d
els,
an
d
t
h
ey
wer
e
s
u
b
s
eq
u
en
tly
ad
ju
s
ted
f
o
r
t
h
e
HAM
1
0
0
0
0
d
ataset.
W
ith
a
to
p
1
ac
c
u
r
ac
y
o
f
8
7
.
9
1
%
.
I
n
o
r
d
er
t
o
ca
teg
o
r
ize
s
k
in
les
io
n
s
as
eith
er
b
e
n
ig
n
o
r
m
alig
n
an
t
(
m
elan
o
m
a)
,
Kee
r
th
an
a
et
a
l.
[
1
3
]
p
r
o
p
o
s
es
two
h
y
b
r
id
m
o
d
els
u
s
in
g
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
co
u
p
led
with
a
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
(
SVM)
.
T
wo
h
y
b
r
id
m
o
d
el
s
ar
e
p
r
o
p
o
s
ed
b
y
th
e
a
u
th
o
r
s
M
o
b
ileNet
[
1
4
]
a
n
d
d
en
s
eNe
t
-
201
’
s
d
is
tin
ct
f
ea
tu
r
es
ar
e
co
m
b
in
ed
i
n
th
e
f
ir
s
t,
wh
ile
R
esNet5
0
an
d
Den
s
eNe
t
-
201
’
s
f
ea
tu
r
es
ar
e
c
o
m
b
in
ed
in
t
h
e
s
ec
o
n
d
.
T
h
e
co
llected
f
ea
tu
r
es
ar
e
th
en
m
er
g
ed
an
d
s
en
t
in
to
an
SV
M
class
if
ier
f
o
r
th
e
f
in
al
class
i
f
icatio
n
.
T
h
e
m
o
d
e
l
is
ev
alu
ated
u
s
in
g
th
e
I
SB
I
2
0
1
6
d
ataset,
wh
ich
co
n
s
is
ts
o
f
9
0
0
tr
ai
n
in
g
im
a
g
es
an
d
3
7
9
t
est
im
ag
es.
I
n
o
r
d
er
to
b
alan
ce
th
e
d
ataset.
An
ac
c
u
r
ac
y
o
f
8
7
.
4
3
%
was
attain
ed
u
s
in
g
th
e
h
y
b
r
id
d
e
n
s
eNe
t2
0
1
+
r
esNet
-
5
0
m
o
d
el
with
SVM.
R
ed
h
a
et
a
l
.
p
r
esen
ts
[
1
5
]
a
h
y
b
r
id
s
k
in
lesi
o
n
s
eg
m
en
tatio
n
an
d
class
if
icatio
n
s
y
s
tem
f
o
r
th
e
I
SIC
2
0
1
8
.
T
h
e
m
eth
o
d
b
le
n
d
s
h
a
n
d
cr
af
ted
f
ea
tu
r
es
(
v
ia
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
els
)
with
d
ee
p
lear
n
in
g
(
u
s
in
g
a
m
o
d
if
ied
UNe
t
ar
ch
itectu
r
e
[
1
6
]
)
.
W
ith
a
m
ea
n
o
v
er
lap
s
co
r
e
o
f
0
.
7
3
5
o
n
v
ali
d
atio
n
d
ata,
a
th
r
esh
o
ld
-
b
ase
d
ap
p
r
o
ac
h
ch
o
o
s
es
UNe
t
f
o
r
l
ar
g
er
lesi
o
n
s
an
d
GM
Ms
f
o
r
s
m
aller
o
n
es
f
o
r
s
eg
m
en
tatio
n
.
T
wo
C
NNs
ar
e
tr
ain
ed
in
a
d
d
itio
n
to
2
0
0
m
a
n
u
ally
cr
ea
ted
f
ea
tu
r
es
f
o
r
cla
s
s
if
icatio
n
.
T
h
ese
f
ea
tu
r
es
ar
e
th
en
co
n
ca
ten
ated
an
d
p
u
t
in
to
a
m
u
lticlas
s
SVM
class
if
ier
,
wh
ich
p
r
o
d
u
ce
s
a
c
lass
av
er
ag
ed
r
ec
all
o
f
0
.
8
4
1
,
ac
cu
r
ac
y
7
0
.
1
0
%.
Z
h
u
an
g
et
a
l.
[
1
7
]
d
em
o
n
s
tr
ates
th
e
ef
f
icac
y
o
f
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
in
s
k
in
lesi
o
n
an
aly
s
is
.
Alth
o
u
g
h
s
o
lo
C
NN
class
if
ier
s
ar
e
ef
f
ec
tiv
e,
i
t
h
as
b
ee
n
d
em
o
n
s
tr
ated
th
at
m
er
g
in
g
s
ev
er
al
class
if
ier
s
u
s
in
g
f
u
s
io
n
ap
p
r
o
ac
h
es
im
p
r
o
v
es
ac
cu
r
ac
y
an
d
r
o
b
u
s
tn
ess
.
T
h
e
ar
ticle
p
r
ese
n
ts
C
S
-
AF,
a
co
s
t
-
s
en
s
itiv
e
m
u
lti
-
clas
s
if
ier
ac
tiv
e
f
u
s
io
n
f
r
am
ewo
r
k
in
ten
d
ed
f
o
r
s
k
in
lesi
o
n
class
if
icat
io
n
,
in
o
r
d
er
to
o
v
e
r
co
m
e
th
ese
p
r
o
b
lem
s
.
I
n
ter
m
s
o
f
a
cc
u
r
ac
y
an
d
lo
wer
in
g
m
is
class
if
icatio
n
co
s
ts
,
th
e
ap
p
r
o
ac
h
r
o
u
tin
ely
b
ea
ts
b
o
th
s
tatic
an
d
ac
tiv
e
f
u
s
io
n
tech
n
iq
u
es
wh
en
test
ed
o
n
th
e
I
SI
C
2
0
1
9
d
ataset
u
s
in
g
9
6
b
ase
class
if
ier
s
d
er
iv
ed
f
r
o
m
1
2
C
NN
ar
ch
itectu
r
es.
A
cc
u
r
ac
y
v
al
u
ed
to
7
7
.
7
4
% d
at
aset I
SIC
2019.
T
o
ad
d
r
ess
th
is
g
ap
,
we
p
r
o
p
o
s
e
a
n
o
v
el
m
u
ltimo
d
al
lear
n
in
g
f
r
am
ewo
r
k
th
at
s
y
n
er
g
is
tically
m
er
g
es
v
is
u
al
p
atter
n
s
f
r
o
m
d
er
m
o
s
c
o
p
ic
im
ag
es
with
s
tr
u
ctu
r
ed
clin
ical
d
ata.
Ou
r
m
ain
c
o
n
tr
ib
u
tio
n
lies
in
th
e
in
tr
o
d
u
ctio
n
an
d
co
m
p
a
r
ativ
e
ev
alu
atio
n
o
f
two
o
r
i
g
in
al
an
d
d
is
tin
ct
ar
ch
itectu
r
es
d
esig
n
e
d
f
o
r
t
h
is
f
u
s
io
n
:
an
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2722
-
2
5
8
6
I
AE
S I
n
t J Ro
b
&
Au
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
1
36
-
1
48
138
MMTN
an
d
an
atten
tio
n
-
b
as
ed
co
n
v
o
lu
tio
n
al
m
u
ltimo
d
al
m
o
d
el
(
C
AM
M)
.
T
h
ese
m
o
d
els
ar
e
d
esig
n
e
d
to
lev
er
ag
e
th
e
co
m
p
lem
en
tar
y
s
tr
en
g
th
s
o
f
im
ag
in
g
a
n
d
cli
n
ical
d
ata,
m
o
v
in
g
b
ey
o
n
d
u
n
im
o
d
al
an
aly
s
is
to
p
r
o
v
id
e
m
o
r
e
h
o
lis
tic
an
d
r
o
b
u
s
t
d
iag
n
o
s
tic
s
u
p
p
o
r
t.
T
h
is
w
o
r
k
is
s
tr
ateg
ically
p
o
s
itio
n
ed
to
ad
v
an
ce
th
e
f
ield
b
y
d
i
r
ec
tly
co
m
p
ar
in
g
th
e
ef
f
ec
tiv
en
ess
o
f
tr
an
s
f
o
r
m
atio
n
-
b
ased
an
d
atten
tio
n
-
b
ased
C
NN
ar
ch
itectu
r
es
f
o
r
m
u
ltimo
d
al
m
elan
o
m
a
class
if
icatio
n
o
n
a
lar
g
e,
p
u
b
licly
a
v
a
ilab
le
d
ataset.
2.
M
E
T
H
O
D
T
h
e
m
ain
n
o
v
elty
o
f
th
is
wo
r
k
lies
in
th
e
p
r
o
p
o
s
al
o
f
two
o
r
ig
in
al
h
y
b
r
id
a
r
ch
itectu
r
es
(
MM
T
N
an
d
C
AM
M)
wh
ich
ef
f
icien
tly
f
u
s
e
v
is
u
al
f
ea
tu
r
es
f
r
o
m
d
ee
p
n
eu
r
al
n
etwo
r
k
s
with
ta
b
u
la
r
clin
ical
d
ata
v
ia
atten
tio
n
m
ec
h
an
is
m
s
,
th
u
s
o
f
f
er
in
g
s
u
p
er
i
o
r
p
e
r
f
o
r
m
an
ce
f
o
r
m
elan
o
m
a
d
etec
tio
n
o
n
th
e
I
SIC 2
0
1
9
d
ataset.
2
.
1
.
Da
t
a
s
et
a
nd
p
re
pro
ce
s
s
i
ng
T
h
e
I
SIC
2
0
1
9
d
ataset
[
1
8
]
was
s
elec
ted
f
o
r
th
is
s
tu
d
y
d
u
e
to
its
s
ize,
clin
ical
r
ele
v
an
ce
,
an
d
th
e
av
ailab
ilit
y
o
f
its
m
etad
ata,
m
ak
in
g
it
a
r
o
b
u
s
t
r
ep
o
s
ito
r
y
f
o
r
th
e
d
ev
el
o
p
m
en
t
o
f
au
to
m
ated
d
iag
n
o
s
tic
s
y
s
tem
s
.
T
h
e
o
r
i
g
in
al
d
ataset
co
m
p
r
is
es
2
5
,
3
3
1
d
e
r
m
o
s
co
p
i
c
im
ag
es
d
is
tr
ib
u
te
d
ac
r
o
s
s
eig
h
t
class
es.
Fo
r
o
u
r
b
in
ar
y
class
if
icatio
n
task
,
we
g
r
o
u
p
ed
th
ese
in
to
two
ca
teg
o
r
ies:
m
elan
o
m
a
(
ME
L
)
,
co
n
tai
n
in
g
4
,
5
2
2
im
ag
es,
an
d
n
o
n
-
m
elan
o
m
ato
u
s
lesi
o
n
s
,
co
n
tain
in
g
2
0
,
8
0
9
im
ag
es
(
all
o
th
er
class
e
s
)
.
E
ac
h
im
ag
e
is
a
s
s
o
ciate
d
with
clin
ical
m
etad
ata,
in
clu
d
in
g
th
e
p
atien
t
’
s
ag
e
an
d
s
ex
.
T
o
en
s
u
r
e
r
ig
o
r
o
u
s
ev
alu
atio
n
,
th
e
d
ataset
was
d
iv
id
ed
in
to
tr
ain
in
g
,
v
alid
atio
n
,
an
d
test
s
ets
u
s
in
g
a
s
tr
atif
ied
r
an
d
o
m
s
am
p
lin
g
m
eth
o
d
,
p
r
eser
v
in
g
th
e
o
r
ig
i
n
al
d
is
tr
ib
u
tio
n
o
f
class
es
w
ith
in
ea
ch
s
u
b
s
et
to
m
in
im
ize
b
ias.
T
h
e
f
in
al
d
is
tr
ib
u
tio
n
is
as
f
o
ll
o
ws:
tr
ain
in
g
(
1
6
,
0
0
0
im
ag
es),
v
alid
atio
n
(
4
,
0
0
0
im
ag
es),
an
d
test
in
g
(
5
,
3
3
1
im
a
g
es).
T
h
is
s
tr
atif
icatio
n
was
es
s
en
tial
to
m
an
ag
e
th
e
in
h
e
r
en
t
im
b
ala
n
ce
b
etwe
en
class
es
d
u
r
in
g
th
e
d
ev
elo
p
m
en
t
a
n
d
e
v
alu
atio
n
o
f
th
e
m
o
d
el.
E
x
a
m
p
les
o
f
im
ag
es
a
r
e
s
h
o
wn
in
F
ig
u
r
e
1
.
T
h
e
d
ataset
u
s
e
d
t
o
s
u
p
p
o
r
t
t
h
e
s
t
u
d
y
’
s
c
o
n
c
l
u
s
i
o
n
s
i
s
p
u
b
li
c
l
y
a
c
ce
s
s
i
b
l
e
v
i
a
:
h
t
t
p
s
:
//
c
h
a
l
l
e
n
g
e
.
i
s
i
c
-
a
r
c
h
iv
e
.
c
o
m
/
d
a
t
a
/
#
2
0
1
9
.
Fig
u
r
e
1
.
B
r
ief
im
a
g
es in
th
is
d
ataset
th
at
ar
e
ca
teg
o
r
ized
in
t
o
eig
h
t g
r
o
u
p
s
Data
p
r
ep
r
o
ce
s
s
in
g
in
v
o
lv
es
s
ev
er
al
s
tep
s
.
I
m
ag
es
ar
e
r
esized
to
a
s
tan
d
ar
d
6
4
×
64
-
p
ix
e
l
s
ize,
an
d
th
eir
p
ix
el
v
alu
es
ar
e
th
en
n
o
r
m
alize
d
to
th
e
r
a
n
g
e
[
0
,
1
]
b
y
d
iv
id
in
g
b
y
2
5
5
.
Data
au
g
m
en
tatio
n
,
u
s
in
g
tech
n
iq
u
es
s
u
ch
as
r
o
tatio
n
,
f
l
ip
p
in
g
,
an
d
zo
o
m
i
n
g
,
is
ap
p
lied
to
im
p
r
o
v
e
m
o
d
el
g
en
e
r
aliza
tio
n
an
d
d
iv
er
s
if
y
th
e
tr
ain
in
g
d
ata.
Fo
r
clin
ical
d
ata,
m
is
s
in
g
v
alu
es
(
s
u
c
h
as
ag
e)
ar
e
im
p
u
ted
u
s
in
g
s
tatis
ti
ca
l
tech
n
iq
u
es,
f
o
r
ex
am
p
le,
b
y
r
ep
lacin
g
m
is
s
in
g
ag
es
with
th
e
m
ed
ia
n
o
f
t
h
e
d
ataset.
C
ateg
o
r
ical
v
ar
iab
l
es,
s
u
ch
as
s
ex
,
ar
e
en
co
d
ed
as
n
u
m
er
ic
v
al
u
es.
Nu
m
er
ic
f
ea
tu
r
es
(
s
u
ch
as
ag
e)
ar
e
also
n
o
r
m
alize
d
to
a
s
tan
d
ar
d
r
an
g
e
[
0
,
1
]
to
s
ca
le
th
em
to
th
e
s
am
e
s
ca
le
a
s
th
e
im
ag
e
f
ea
tu
r
es.
Fo
r
b
in
ar
y
class
if
icatio
n
,
g
r
o
u
n
d
tr
u
th
l
ab
els ar
e
co
n
v
er
ted
to
b
in
ar
y
f
o
r
m
at:
1
f
o
r
m
ela
n
o
m
a
(
ME
L
)
an
d
0
f
o
r
n
o
n
-
m
elan
o
m
a
(
all
o
th
er
class
es).
Fin
ally
,
cu
s
to
m
d
ata
g
en
er
ato
r
s
a
r
e
u
s
ed
to
ef
f
icie
n
tly
in
teg
r
ate
im
ag
in
g
a
n
d
cl
in
ical
d
ata,
a
n
d
t
o
m
a
n
ag
e
th
e
lar
g
e
d
ataset
b
y
lo
ad
in
g
a
n
d
p
r
ep
r
o
ce
s
s
in
g
d
at
a
in
b
atch
es d
u
r
in
g
tr
ai
n
in
g
a
n
d
ev
alu
atio
n
.
I
n
o
r
d
er
to
im
p
r
o
v
e
th
e
p
r
ec
is
io
n
an
d
th
o
r
o
u
g
h
n
ess
o
f
th
e
d
iag
n
o
s
tic
p
r
o
ce
d
u
r
e,
th
e
m
u
ltimo
d
al
tr
an
s
f
o
r
m
er
n
etwo
r
k
an
d
c
o
n
v
o
lu
tio
n
al
atten
tio
n
m
ec
h
an
i
s
m
u
s
ed
in
th
is
s
tu
d
y
f
o
r
m
elan
o
m
a
d
etec
tio
n
co
m
b
in
es
clin
ical
an
d
im
a
g
in
g
d
ata.
Her
e
is
a
th
o
r
o
u
g
h
b
r
ea
k
d
o
wn
o
f
th
e
m
et
h
o
d
o
lo
g
y
:
T
h
e
m
o
d
els
m
a
k
e
u
s
e
o
f
b
o
th
th
e
co
n
tex
t
u
al
p
atien
t
k
n
o
wled
g
e
f
r
o
m
clin
ical
d
ata
an
d
th
e
v
is
u
al
cu
es
f
r
o
m
d
er
m
o
s
co
p
i
c
p
ictu
r
es.
C
o
m
p
ar
ed
to
s
in
g
l
e
-
m
o
d
ality
m
o
d
els,
th
is
s
y
n
er
g
is
tic
ap
p
r
o
ac
h
o
f
f
er
s
a
m
o
r
e
co
m
p
r
eh
en
s
iv
e
k
n
o
wled
g
e
o
f
th
e
co
n
d
itio
n
an
d
in
cr
ea
s
es
d
etec
tio
n
ac
cu
r
ac
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S I
n
t J Ro
b
&
Au
to
m
I
SS
N:
2722
-
2
5
8
6
Mu
lti
-
mo
d
a
l tra
n
s
fo
r
mer a
n
d
co
n
vo
lu
tio
n
a
l a
tten
tio
n
a
r
ch
it
ec
tu
r
es fo
r
mela
n
o
ma
…
(
Gu
i
d
o
u
m
A
min
a
)
139
2
.
2
.
Arc
hite
ct
ures o
f
pro
po
s
ed
m
o
dels
W
e
p
r
o
p
o
s
e
two
n
ew
m
u
ltimo
d
al
ar
ch
itectu
r
es
,
MM
T
N
an
d
C
AM
M,
d
esig
n
ed
to
au
to
m
atica
lly
in
teg
r
ate
im
ag
in
g
an
d
clin
ical
d
ata
f
o
r
m
elan
o
m
a
d
etec
tio
n
.
T
h
ese
two
m
o
d
els
ac
ce
p
t
two
s
y
n
ch
r
o
n
ize
d
in
p
u
t
s
tr
ea
m
s
,
r
ef
lectin
g
an
au
to
m
ated
d
iag
n
o
s
tic
p
r
o
ce
s
s
wh
er
e
v
is
u
al
an
d
clin
ical
d
ata
ar
e
p
r
o
ce
s
s
ed
s
im
u
ltan
eo
u
s
ly
.
2
.
2
.
1
.
M
M
T
N
a
rc
hite
ct
ure
T
o
im
p
r
o
v
e
m
elan
o
m
a
d
ete
ctio
n
,
th
e
MM
T
N
m
o
d
el
co
m
b
in
es
clin
ical
d
ata
an
d
d
e
r
m
o
s
co
p
ic
im
ag
es.
I
t
co
n
s
is
ts
o
f
two
m
a
in
p
ar
ts
:
d
en
s
e
lay
er
s
f
o
r
p
r
o
c
ess
in
g
clin
ical
d
ata
an
d
a
tr
an
s
f
o
r
m
er
-
lik
e
b
l
o
ck
with
an
e
n
co
d
e
r
f
o
r
im
ag
e
d
ata.
T
h
is
tr
a
n
s
f
o
r
m
e
r
b
l
o
ck
is
u
s
ed
t
o
p
r
o
ce
s
s
im
ag
e
d
ata
f
o
r
m
elan
o
m
a
d
etec
tio
n
,
lev
er
ag
in
g
its
ab
ilit
y
to
m
o
d
el
co
m
p
le
x
in
ter
ac
tio
n
s
b
etwe
en
d
if
f
er
en
t
elem
en
ts
with
in
an
im
ag
e.
T
h
is
ap
p
r
o
ac
h
is
p
ar
ticu
lar
l
y
well
-
s
u
ited
b
ec
au
s
e
th
e
tr
an
s
f
o
r
m
er
[
1
9
]
,
in
itially
d
esig
n
e
d
f
o
r
n
atu
r
al
lan
g
u
a
g
e
p
r
o
ce
s
s
in
g
wh
er
e
it
ex
ce
ls
at
ca
p
tu
r
in
g
r
elatio
n
s
h
ip
s
th
r
o
u
g
h
s
elf
-
atten
tio
n
m
ec
h
a
n
is
m
s
,
ca
n
b
e
ap
p
lied
to
o
th
er
ty
p
es
o
f
d
ata
.
T
h
e
MM
T
N
m
o
d
el
is
d
esig
n
e
d
to
p
r
o
ce
s
s
two
p
ar
allel
d
ata
s
tr
ea
m
s
as
s
h
o
wn
in
Fig
u
r
e
2
.
T
h
e
im
ag
e
s
tr
ea
m
u
s
es
a
tr
a
n
s
f
o
r
m
en
c
o
d
er
to
ca
p
tu
r
e
lo
n
g
-
r
an
g
e
d
ep
en
d
en
cies
a
n
d
s
p
atial
r
elat
io
n
s
h
ip
s
with
in
t
h
e
d
er
m
o
s
co
p
ic
im
ag
e.
T
h
is
b
lo
ck
em
p
l
o
y
s
a
m
u
lti
-
h
ea
d
(
4
-
h
ea
d
)
s
elf
-
atten
tio
n
m
ec
h
an
is
m
,
f
o
llo
wed
b
y
lay
e
r
n
o
r
m
aliza
tio
n
an
d
a
f
o
r
wa
r
d
-
p
r
o
p
a
g
atin
g
n
etwo
r
k
,
th
u
s
tr
an
s
f
o
r
m
in
g
th
e
in
p
u
t
im
ag
e
in
to
a
f
ea
tu
r
e
-
r
ic
h
r
ep
r
esen
tatio
n
.
Simu
ltan
eo
u
s
l
y
,
th
e
clin
ical
d
ata
s
tr
ea
m
(
a
g
e,
s
ex
)
is
p
r
o
ce
s
s
ed
b
y
a
s
er
ies
o
f
d
e
n
s
e
n
e
u
r
al
lay
er
s
.
T
h
e
r
esu
ltin
g
f
ea
tu
r
e
v
ec
to
r
s
f
r
o
m
t
h
e
two
m
o
d
alit
ies
ar
e
th
en
co
n
ca
te
n
ated
a
n
d
p
ass
ed
to
a
f
in
al
class
if
icatio
n
lay
er
.
T
h
is
ar
ch
i
tectu
r
e
allo
ws
th
e
m
o
d
el
to
au
to
m
atica
lly
co
r
r
elate
v
is
u
al
p
atter
n
s
with
p
atien
t
-
s
p
ec
if
ic
r
is
k
f
ac
to
r
s
.
Fig
u
r
e
2
.
T
h
e
ar
ch
itectu
r
e
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
MM
T
N
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2722
-
2
5
8
6
I
AE
S I
n
t J Ro
b
&
Au
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
1
36
-
1
48
140
2
.
2
.
2
.
CAMM
a
rc
hite
ct
ure
T
h
e
C
AM
M
ar
ch
itectu
r
e
o
f
f
e
r
s
a
lig
h
tweig
h
t
an
d
h
ig
h
-
p
er
f
o
r
m
an
ce
m
o
d
u
le,
s
u
itab
le
f
o
r
au
to
m
ated
en
v
ir
o
n
m
en
ts
with
lim
ited
r
eso
u
r
ce
s
as
s
h
o
wn
in
Fig
u
r
e
3
.
I
t
r
elies
o
n
a
Mo
b
ileNetV2
n
e
two
r
k
f
o
r
ef
f
icien
t
im
ag
e
f
ea
tu
r
e
ex
tr
ac
tio
n
.
A
k
ey
in
n
o
v
ativ
e
asp
ec
t
o
f
its
ar
ch
itectu
r
e
is
in
s
p
ir
ed
b
y
th
e
co
n
v
o
lu
ti
o
n
al
b
lo
c
k
atten
tio
n
m
o
d
u
le
(
C
B
AM
)
[
2
0
]
,
wh
ic
h
s
eq
u
en
tially
a
p
p
lies
s
p
atial
an
d
p
e
r
-
ch
an
n
el
atten
t
io
n
to
r
ef
i
n
e
f
ea
t
u
r
e
m
ap
s
,
th
u
s
f
o
r
ci
n
g
th
e
m
o
d
el
t
o
au
to
m
atica
lly
f
o
cu
s
o
n
t
h
e
m
o
s
t r
elev
an
t v
is
u
al
f
ea
tu
r
es f
o
r
m
elan
o
m
a.
T
h
ese
r
ef
in
ed
i
m
ag
e
f
ea
tu
r
es
ar
e
th
en
f
u
s
ed
with
p
r
o
ce
s
s
ed
clin
ical
d
ata
(
ag
e
,
s
ex
)
.
T
h
e
u
s
e
o
f
ef
f
icien
t
C
NN
ar
ch
itectu
r
e
c
o
m
b
in
e
d
with
f
o
cu
s
ed
atten
tio
n
m
ak
es
C
AM
M
a
p
r
o
m
is
in
g
ca
n
d
id
ate
f
o
r
i
n
teg
r
atio
n
in
to
r
ea
l
-
tim
e
em
b
ed
d
e
d
d
iag
n
o
s
tic
d
e
v
ices
o
r
telem
ed
icin
e
p
latf
o
r
m
s
,
wh
ile
th
e
g
en
er
ate
d
atten
tio
n
m
ap
s
o
f
f
er
ed
a
d
eg
r
ee
o
f
in
ter
p
r
etab
ilit
y
f
o
r
s
y
s
tem
v
alid
atio
n
.
O
u
r
m
e
c
h
a
n
i
s
m
a
p
p
l
i
es
t
h
e
se
s
t
e
p
s
s
e
q
u
e
n
t
i
al
l
y
.
i
)
A
t
t
e
n
ti
o
n
p
e
r
c
h
a
n
n
e
l
:
A
G
l
o
b
a
l
A
v
e
r
a
g
eP
o
o
l
i
n
g
2
D
lay
er
f
o
llo
wed
b
y
a
d
e
n
s
e
lay
er
with
a
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U
)
ac
tiv
atio
n
f
u
n
ctio
n
a
n
d
a
f
in
al
d
en
s
e
lay
e
r
with
s
ig
m
o
id
ac
tiv
atio
n
g
e
n
e
r
ates
a
weig
h
t
v
ec
to
r
p
er
ch
a
n
n
el.
T
h
is
v
ec
to
r
is
m
u
ltip
lie
d
b
y
in
p
u
t
f
ea
t
u
r
e
m
ap
s
to
ac
ce
n
tu
ate
th
e
m
o
s
t
r
elev
an
t
f
ea
tu
r
e
ch
a
n
n
els.
ii)
Sp
atial
atten
tio
n
:
o
n
th
e
ch
an
n
el
-
r
ec
alib
r
ate
d
f
ea
tu
r
es,
we
ap
p
ly
2
D
c
o
n
v
o
lu
tio
n
s
to
cr
ea
te
a
u
n
iq
u
e
s
p
at
ial
atten
tio
n
m
ap
(
s
ig
m
o
id
ac
tiv
atio
n
)
,
wh
ich
is
th
en
m
u
ltip
lied
elem
e
n
t b
y
ele
m
en
t to
h
ig
h
lig
h
t th
e
m
o
s
t sp
atially
s
ig
n
if
ican
t r
eg
io
n
s
o
f
th
e
lesi
o
n
.
Fig
u
r
e
3
.
T
h
e
ar
ch
itectu
r
e
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
C
AM
M
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S I
n
t J Ro
b
&
Au
to
m
I
SS
N:
2722
-
2
5
8
6
Mu
lti
-
mo
d
a
l tra
n
s
fo
r
mer a
n
d
co
n
vo
lu
tio
n
a
l a
tten
tio
n
a
r
ch
it
ec
tu
r
es fo
r
mela
n
o
ma
…
(
Gu
i
d
o
u
m
A
min
a
)
141
Fo
r
co
m
p
lete
r
e
p
r
o
d
u
cib
ilit
y
,
we
p
r
o
v
id
e
th
e
m
ain
im
p
lem
e
n
tatio
n
lo
g
ic
as in
Alg
o
r
ith
m
1
.
Alg
o
r
ith
m
1
: M
MT
N
Pip
elin
e
f
o
r
m
ela
n
o
m
a
d
etec
tio
n
I
n
p
u
t
:
D
e
r
m
o
sc
o
p
i
c
i
ma
g
e
s I
,
C
l
i
n
i
c
a
l
d
a
t
a
C
(
a
g
e
,
s
e
x
)
O
u
t
p
u
t
:
P
r
e
d
i
c
t
i
o
n
y
_
p
r
e
d
(
me
l
a
n
o
ma
p
r
o
b
a
b
i
l
i
t
y
)
/
/
1
.
D
a
t
a
p
r
e
p
a
r
a
t
i
o
n
I
_
p
r
e
p
r
o
c
e
ss
e
d
←
r
e
si
z
e
(
I
,
6
4
×
6
4
)
/
2
5
5
.
0
C
_
p
r
e
p
r
o
c
e
ss
e
d
←
c
o
n
c
a
t
e
n
a
t
e
(
[
n
o
r
mal
i
z
e
(
a
g
e
)
,
o
n
e
_
h
o
t
(
s
e
x
)
]
)
/
/
2
.
I
mag
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
/
/
C
u
st
o
m Tr
a
n
sf
o
r
mer
e
n
c
o
d
e
r
(
4
a
t
t
e
n
t
i
o
n
h
e
a
d
s)
i
ma
g
e
_
f
e
a
t
u
r
e
s ←
Tr
a
n
sf
o
r
m
e
r
B
l
o
c
k
(
n
_
h
e
a
d
s=
4
)
(
I
_
p
r
e
p
r
o
c
e
sse
d
)
i
ma
g
e
_
f
e
a
t
u
r
e
s ←
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
1
D
(
)
(
i
ma
g
e
_
f
e
a
t
u
r
e
s)
/
/
3
.
C
l
i
n
i
c
a
l
d
a
t
a
p
r
o
c
e
ss
i
n
g
c
l
i
n
i
c
a
l
_
b
r
a
n
c
h
←
D
e
n
se(
6
4
,
a
c
t
i
v
a
t
i
o
n
=
‘
r
e
l
u
’
)
(
C
_
p
r
e
p
r
o
c
e
ss
e
d
)
c
l
i
n
i
c
a
l
_
b
r
a
n
c
h
←
D
r
o
p
o
u
t
(
0
.
3
)
(
c
l
i
n
i
c
a
l
_
b
r
a
n
c
h
)
c
l
i
n
i
c
a
l
_
f
e
a
t
u
r
e
s
←
D
e
n
s
e
(
3
2
,
a
c
t
i
v
a
t
i
o
n
=
‘
r
e
l
u
’
)
(
c
l
i
n
i
c
a
l
_
b
r
a
n
c
h
)
/
/
4
.
M
u
l
t
i
mo
d
a
l
f
u
si
o
n
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s ←
c
o
n
c
a
t
e
n
a
t
e
(
[
i
mag
e
_
f
e
a
t
u
r
e
s
,
c
l
i
n
i
c
a
l
_
f
e
a
t
u
r
e
s]
)
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s ←
D
e
n
se(
1
2
8
,
a
c
t
i
v
a
t
i
o
n
=
‘
r
e
l
u
’
)
(
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s)
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s ←
D
r
o
p
o
u
t
(
0
.
5
)
(
c
o
mb
i
n
e
d
_
f
e
a
t
u
r
e
s)
/
/
5
.
C
l
a
ss
i
f
i
c
a
t
i
o
n
y
_
p
r
e
d
←
D
e
n
se(
1
,
a
c
t
i
v
a
t
i
o
n
=
‘
si
g
m
o
i
d
’
)
(
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s)
/
/
D
r
i
v
e
c
o
n
f
i
g
u
r
a
t
i
o
n
l
o
ss
←
w
e
i
g
h
t
e
d
_
b
i
n
a
r
y
_
c
r
o
sse
n
t
r
o
p
y
(
w
e
i
g
h
t
=
[
0
.
2
,
0
.
8
]
)
o
p
t
i
m
i
z
e
←
A
d
a
m(
l
e
a
r
n
i
n
g
_
r
a
t
e
=
0
.
0
0
1
)
mo
d
e
l
.
c
o
m
p
i
l
e
(
o
p
t
i
m
i
z
e
r
,
l
o
ss,
me
t
r
i
c
s=[
‘
a
c
c
u
r
a
c
y
’
,
A
U
C
(
)
]
)
Alg
o
r
ith
m
2
: CAMM
Pip
elin
e
f
o
r
m
ela
n
o
m
a
d
etec
tio
n
I
n
p
u
t
:
D
e
r
m
o
sc
o
p
i
c
i
ma
g
e
s I
,
C
l
i
n
i
c
a
l
d
a
t
a
C
(
a
g
e
,
s
e
x
)
O
u
t
p
u
t
:
P
r
e
d
i
c
t
i
o
n
y
_
p
r
e
d
(
me
l
a
n
o
ma
p
r
o
b
a
b
i
l
i
t
y
)
/
/
1
.
D
a
t
a
p
r
e
p
a
r
a
t
i
o
n
I
_
p
r
e
p
r
o
c
e
ss
e
d
←
r
e
si
z
e
(
I
,
2
2
4
×
2
2
4
)
/
2
5
5
.
0
C
_
p
r
e
p
r
o
c
e
ss
e
d
←
c
o
n
c
a
t
e
n
a
t
e
(
[
n
o
r
mal
i
z
e
(
a
g
e
)
,
o
n
e
_
h
o
t
(
s
e
x
)
]
)
/
/
2
.
I
mag
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
w
i
t
h
M
o
b
i
l
e
N
e
t
V
2
i
ma
g
e
_
b
a
c
k
b
o
n
e
←
M
o
b
i
l
e
N
e
t
V
2
(
w
e
i
g
h
t
=
‘
i
m
a
g
e
n
e
t
’
,
i
n
c
l
u
d
e
_
t
o
p
=
F
a
l
se)(
I
_
p
r
e
p
r
o
c
e
sse
d
)
/
/
3
.
A
t
t
e
n
t
i
o
n
m
e
c
h
a
n
i
sm
(
i
n
sp
i
r
e
d
b
y
C
B
A
M
)
/
/
3
.
1
C
h
a
n
n
e
l
-
b
a
se
d
a
t
t
e
n
t
i
o
n
c
h
a
n
n
e
l
_
a
v
g
←
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
2
D
(
)
(
i
m
a
g
e
_
b
a
c
k
b
o
n
e
)
c
h
a
n
n
e
l
_
w
e
i
g
h
t
s
←
D
e
n
se(
u
n
i
t
s
=
1
2
8
,
a
c
t
i
v
a
t
i
o
n
=
‘
r
e
l
u
’
)
(
c
h
a
n
n
e
l
_
a
v
g
)
c
h
a
n
n
e
l
_
w
e
i
g
h
t
s
←
D
e
n
se(
u
n
i
t
s
=
i
m
a
g
e
_
b
a
c
k
b
o
n
e
.
sh
a
p
e
[
-
1
]
,
a
c
t
i
v
a
t
i
o
n
=
‘
si
g
mo
i
d
’
)
(
c
h
a
n
n
e
l
_
w
e
i
g
h
t
s)
c
h
a
n
n
e
l
_
r
e
f
i
n
e
d
←
mu
l
t
i
p
l
y
(
[
i
ma
g
e
_
b
a
c
k
b
o
n
e
,
c
h
a
n
n
e
l
_
w
e
i
g
h
t
s]
)
/
/
3
.
2
S
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
sp
a
t
i
a
l
_
w
e
i
g
h
t
s
←
C
o
n
v
2
D
(
f
i
l
t
e
r
s=
1
,
k
e
r
n
e
l
_
s
i
z
e
=
7
,
p
a
d
d
i
n
g
=
‘
sam
e
’
,
a
c
t
i
v
a
t
i
o
n
=
‘
si
g
m
o
i
d
’
)
(
c
h
a
n
n
e
l
_
r
e
f
i
n
e
d
)
a
t
t
e
n
d
e
d
_
f
e
a
t
u
r
e
s
←
m
u
l
t
i
p
l
y
(
[
c
h
a
n
n
e
l
_
r
e
f
i
n
e
d
,
s
p
a
t
i
a
l
_
w
e
i
g
h
t
s]
)
/
/
4
.
I
mag
e
f
e
a
t
u
r
e
a
g
g
r
e
g
a
t
i
o
n
i
ma
g
e
_
f
e
a
t
u
r
e
s ←
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
2
D
(
)
(
a
t
t
e
n
d
e
d
_
f
e
a
t
u
r
e
s)
/
/
5
.
P
r
o
c
e
ssi
n
g
o
f
c
l
i
n
i
c
a
l
d
a
t
a
c
l
i
n
i
c
a
l
_
f
e
a
t
u
r
e
s
←
D
e
n
s
e
(
3
2
,
a
c
t
i
v
a
t
i
o
n
=
‘
r
e
l
u
’
)
(
C
_
p
r
e
p
r
o
c
e
sse
d
)
/
/
6
.
M
u
l
t
i
mo
d
a
l
F
u
si
o
n
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s ←
c
o
n
c
a
t
e
n
a
t
e
(
[
i
mag
e
_
f
e
a
t
u
r
e
s
,
c
l
i
n
i
c
a
l
_
f
e
a
t
u
r
e
s]
)
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s ←
D
e
n
se(
6
4
,
a
c
t
i
v
a
t
i
o
n
=
‘
r
e
l
u
’
)
(
c
o
mb
i
n
e
d
_
f
e
a
t
u
r
e
s)
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s ←
D
r
o
p
o
u
t
(
0
.
4
)
(
c
o
mb
i
n
e
d
_
f
e
a
t
u
r
e
s)
/
/
7
.
C
l
a
ss
i
f
i
c
a
t
i
o
n
y
_
p
r
e
d
←
D
e
n
se(
1
,
a
c
t
i
v
a
t
i
o
n
=
‘
si
g
m
o
i
d
’
)
(
c
o
m
b
i
n
e
d
_
f
e
a
t
u
r
e
s)
/
/
Tr
a
i
n
i
n
g
S
e
t
u
p
(
same
a
s
M
M
TN
f
o
r
f
a
i
r
c
o
m
p
a
r
i
so
n
)
l
o
ss
←
w
e
i
g
h
t
e
d
_
b
i
n
a
r
y
_
c
r
o
sse
n
t
r
o
p
y
(
w
e
i
g
h
t
=
[
0
.
2
,
0
.
8
]
)
o
p
t
i
m
i
z
e
r
←
A
d
a
m(
l
e
a
r
n
i
n
g
_
r
a
t
e
=
0
.
0
0
1
)
mo
d
e
l
.
c
o
m
p
i
l
e
(
o
p
t
i
m
i
z
e
r
,
l
o
ss,
me
t
r
i
c
s=[
‘
a
c
c
u
r
a
c
y
’
,
A
U
C
(
)
]
)
2
.
3
.
T
ra
ini
ng
a
nd
i
m
plem
en
t
a
t
io
n det
a
ils
B
o
th
m
o
d
els
wer
e
tr
ain
ed
u
s
in
g
th
e
Ad
am
o
p
tim
izer
an
d
a
weig
h
ted
b
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
,
with
class
weig
h
ts
in
v
er
s
ely
p
r
o
p
o
r
tio
n
al
to
t
h
eir
f
r
eq
u
e
n
cies
in
th
e
tr
ain
in
g
s
et
to
co
r
r
ec
t
f
o
r
th
e
im
b
alan
ce
b
etwe
en
m
elan
o
m
a
s
an
d
n
o
n
-
m
elan
o
m
as.
T
h
e
M
MT
N
m
o
d
el
was
tr
ain
ed
f
o
r
2
0
ep
o
ch
s
,
wh
ile
th
e
C
B
AM
m
o
d
el
was
tr
ain
ed
f
o
r
5
0
ep
o
ch
s
.
Hy
p
e
r
p
ar
am
ete
r
s
,
in
clu
d
in
g
th
e
lear
n
i
n
g
r
ate
a
n
d
b
atch
s
ize,
wer
e
o
p
tim
ized
b
y
ex
h
a
u
s
tiv
e
s
ea
r
ch
.
T
h
e
s
ea
r
ch
s
p
ac
e
f
o
r
th
e
lear
n
in
g
r
ate
was
[
0
.
1
;
0
.
0
1
;
0
.
0
0
1
]
an
d
f
o
r
t
h
e
b
atch
s
ize,
[
1
6
;
3
2
;
6
4
]
.
T
h
e
o
p
tim
al
v
alu
es
o
b
tain
ed
wer
e
a
lear
n
in
g
r
ate
o
f
0
.
0
0
1
a
n
d
a
b
atch
s
ize
o
f
3
2
.
All
ex
p
er
im
en
ts
wer
e
p
er
f
o
r
m
e
d
o
n
a
s
tan
d
a
r
d
p
r
o
ce
s
s
o
r
(
I
n
tel
C
o
r
e
i7
,
3
2
GB
o
f
R
AM
)
with
Py
th
o
n
3
.
8
,
T
en
s
o
r
Flo
w
2
.
8
an
d
s
cik
it
-
le
ar
n
1
.
0
.
2
,
with
r
an
d
o
m
s
ee
d
s
f
ix
ed
to
en
s
u
r
e
r
ep
r
o
d
u
ci
b
ilit
y
.
All
ex
p
er
im
en
ts
wer
e
p
er
f
o
r
m
e
d
o
n
a
s
tan
d
a
r
d
C
PU,
d
em
o
n
s
tr
atin
g
th
e
co
m
p
u
tatio
n
al
f
ea
s
ib
ilit
y
an
d
d
e
p
l
o
y
m
en
t
p
o
ten
tial
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2722
-
2
5
8
6
I
AE
S I
n
t J Ro
b
&
Au
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
1
36
-
1
48
142
o
u
r
ar
c
h
itectu
r
es
f
o
r
c
o
s
t
-
ef
f
e
ctiv
e
au
to
m
ated
s
cr
ee
n
in
g
s
etu
p
s
,
p
ar
ticu
lar
ly
wh
en
d
e
d
icate
d
GPU
h
ar
d
war
e
is
u
n
av
ailab
le.
2
.
4
.
Co
ns
idera
t
io
ns
re
la
t
ing
t
o
a
uto
m
a
t
ed
deplo
y
m
ent
T
h
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
es,
p
ar
ticu
lar
ly
C
AM
M
with
its
Mo
b
ileNetV2
b
ac
k
b
o
n
e,
ex
h
i
b
it
d
esig
n
f
ea
tu
r
es
r
elev
an
t
f
o
r
au
to
m
ate
d
d
ep
lo
y
m
en
t.
T
h
ei
r
ab
ilit
y
to
s
im
u
ltan
eo
u
s
ly
p
r
o
ce
s
s
d
er
m
o
s
co
p
ic
im
ag
es
an
d
clin
ical
m
etad
ata
m
ee
ts
th
e
n
ee
d
s
o
f
in
te
g
r
ated
d
iag
n
o
s
tic
s
y
s
tem
s
.
E
v
alu
atio
n
o
n
a
s
tan
d
ar
d
C
PU
d
em
o
n
s
tr
ates
co
m
p
u
tatio
n
al
f
ea
s
ib
ilit
y
f
o
r
r
eso
u
r
ce
-
c
o
n
s
tr
ain
ed
en
v
ir
o
n
m
en
ts
.
Fu
r
t
h
er
m
o
r
e,
t
h
e
atten
tio
n
m
ec
h
an
is
m
s
p
r
o
v
id
e
v
is
u
al
s
alien
ce
m
ap
s
th
at
co
u
ld
f
ac
ili
tate
in
ter
p
r
etatio
n
in
a
u
to
m
at
ed
clin
ical
d
ec
is
io
n
s
u
p
p
o
r
t sy
s
tem
s
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
o
ev
alu
ate
o
u
r
m
o
d
els,
we
u
s
ed
th
e
f
o
llo
win
g
m
etr
ics
[
2
1
]
–
[
2
4
]
:
ac
cu
r
ac
y
,
wh
ic
h
co
r
r
esp
o
n
d
s
t
o
th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
tly
id
en
tifie
d
s
am
p
les
am
o
n
g
all
s
am
p
les
an
d
p
r
o
v
id
es
an
o
v
er
all
m
ea
s
u
r
e
o
f
th
e
m
o
d
el
’
s
ac
cu
r
ac
y
,
as
d
ef
in
ed
in
(
1
)
.
Pre
cisi
o
n
is
th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
t
p
o
s
itiv
e
p
r
e
d
ictio
n
s
am
o
n
g
all
p
o
s
itiv
e
p
r
ed
ictio
n
s
,
ac
co
r
d
in
g
to
(
2
)
.
R
ec
all,
o
r
s
en
s
itiv
ity
,
r
ep
r
esen
ts
th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
tly
id
en
tifie
d
tr
u
e
p
o
s
itiv
es
am
o
n
g
all
ac
tu
al
p
o
s
itiv
e
ca
s
es,
as
f
o
r
m
u
lated
in
(
3
)
.
Fin
ally
,
th
e
F1
s
co
r
e,
d
ef
in
ed
in
(
4
)
,
is
th
e
h
ar
m
o
n
ic
m
ea
n
o
f
p
r
ec
is
io
n
a
n
d
r
ec
all;
it
b
alan
ce
s
th
ese
tw
o
m
etr
ics
an
d
is
p
a
r
ticu
lar
ly
u
s
ef
u
l
f
o
r
e
v
alu
atin
g
p
er
f
o
r
m
an
ce
o
n
u
n
b
alan
ce
d
d
atasets
.
A
c
c
ura
c
y=
t
rue
p
o
sit
iv
e
(t
p
)+
t
rue
n
e
g
a
t
iv
e
s(
t
n
)
t
p
+
T
N
+
f
a
l
se
p
o
sit
iv
e
s(f
p
)+
f
a
l
se
n
e
g
a
t
iv
e
s(
f
n
)
(
1
)
Pr
e
c
ision=
tp
t
p
+
f
p
(
2
)
R
e
c
a
l
l
=
tp
t
p
+
FN
(
3
)
F
1Score
=2.
p
re
c
isi
o
n
.re
c
a
l
l
p
re
c
isi
o
n
+
re
c
a
l
l
(
4
)
T
h
e
AUC
R
O
C
,
wh
ich
s
tan
d
s
f
o
r
ar
ea
u
n
d
er
th
e
r
ec
ei
v
er
o
p
er
atin
g
c
h
ar
ac
ter
is
tic
cu
r
v
e
,
m
ea
s
u
r
es
th
e
p
er
f
o
r
m
an
ce
o
f
a
class
if
ie
r
b
y
co
m
p
ar
in
g
t
h
e
tr
u
e
p
o
s
itiv
e
r
ate
(
T
PR
)
to
th
e
f
alse
p
o
s
itiv
e
r
ate
(
FP
R
)
at
d
if
f
er
en
t
d
ec
is
io
n
th
r
esh
o
ld
s
.
Similar
ly
,
th
e
co
n
f
u
s
io
n
m
a
tr
ix
is
a
ta
b
le
ev
al
u
atin
g
th
e
p
er
f
o
r
m
a
n
ce
o
f
a
class
if
icatio
n
m
o
d
el
b
y
co
m
p
a
r
in
g
p
r
ed
icted
la
b
els
to
ac
tu
al
lab
els;
it
co
n
s
is
ts
o
f
f
o
u
r
elem
en
ts
:
tr
u
e
p
o
s
itiv
es
(
T
P),
co
r
r
esp
o
n
d
in
g
to
co
r
r
ec
tly
p
r
ed
icted
p
o
s
itiv
e
ca
s
es;
tr
u
e
n
eg
ativ
es
(
T
N)
,
wh
ich
a
r
e
co
r
r
ec
tly
id
en
tifie
d
n
eg
ativ
e
ca
s
es;
f
alse
p
o
s
itiv
es
(
FP
)
,
r
ep
r
esen
tin
g
ty
p
e
I
er
r
o
r
s
wh
er
e
n
eg
ativ
e
ca
s
es
ar
e
in
co
r
r
ec
tly
p
r
ed
icted
as
p
o
s
itiv
e;
an
d
f
alse
n
eg
ativ
e
s
(
FN)
,
wh
ich
ar
e
ty
p
e
I
I
er
r
o
r
s
wh
er
e
p
o
s
itiv
e
ca
s
es
ar
e
in
co
r
r
ec
tly
p
r
e
d
icted
as n
eg
ativ
e.
3
.
1
.
E
v
a
lua
t
i
o
n o
f
re
s
ults f
o
r
M
M
T
N
B
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
an
d
th
e
Ad
am
o
p
tim
izer
[
2
5
]
wer
e
u
s
ed
to
tr
ain
th
e
m
o
d
el
o
v
er
2
0
ep
o
ch
s
.
T
h
e
m
o
d
el
’
s
p
er
f
o
r
m
an
ce
was
ev
alu
ated
u
s
in
g
th
e
f
o
llo
wi
n
g
m
etr
ics:
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e
,
an
d
AUC R
OC
.
T
h
e
ex
am
i
n
atio
n
o
f
th
e
co
n
f
u
s
io
n
m
atr
ix
an
d
AUC
cu
r
v
e
r
esu
lts
as
s
h
o
wn
in
Fig
u
r
e
s
4
an
d
5
,
p
r
esen
ts
a
co
m
p
ar
is
o
n
b
etwe
en
o
u
r
MM
T
N
m
o
d
el
an
d
th
e
f
o
llo
win
g
m
o
d
els
(
VI
T
[
2
6
]
,
E
f
f
icien
tNet
[
2
7
]
,
Mo
b
ileNet)
.
All
m
o
d
els
u
s
e
th
e
s
am
e
co
n
f
ig
u
r
atio
n
,
as
well
as
m
u
ltimo
d
al
lear
n
in
g
—
co
m
b
in
in
g
clin
ical
d
ata
an
d
d
er
m
o
s
co
p
ic
im
a
g
es
—
o
n
th
e
I
SIC
2
0
1
9
test
d
ataset.
No
te:
C
lass
0
.
0
co
r
r
esp
o
n
d
s
to
n
o
n
-
m
ela
n
o
m
a
a
n
d
class
1
.
0
to
m
elan
o
m
a
.
Mu
ltimo
d
al
in
teg
r
atio
n
im
p
r
o
v
es
th
e
co
n
tr
i
b
u
tio
n
o
f
clin
ic
al
d
ata
an
d
p
er
f
o
r
m
an
ce
:
as
th
e
r
is
k
o
f
m
elan
o
m
a
in
cr
ea
s
es
with
ag
e,
th
e
ag
e/sex
m
etad
ata
lik
el
y
en
r
ich
ed
th
e
r
ep
r
esen
tatio
n
o
f
f
ea
tu
r
es.
T
h
is
ex
p
lain
s
th
e
h
ig
h
e
r
AUC
(
0
.
8
5
)
co
m
p
a
r
ed
to
Mo
b
ileNet
alo
n
e
(
AUC=0
.
8
2
in
T
a
b
le
1
)
.
Atten
tio
n
p
er
tr
an
s
f
o
r
m
er
was
f
o
cu
s
ed
o
n
ar
ea
s
o
f
th
e
im
a
g
e
with
d
iag
n
o
s
tic
s
ig
n
if
ican
ce
(
ir
r
eg
u
lar
b
o
r
d
er
s
,
c
o
lo
r
v
ar
iab
ilit
y
)
.
T
h
is
is
k
ey
t
o
in
cr
ea
s
in
g
m
elan
o
m
a
r
ec
all
f
r
o
m
3
8
% with
Mo
b
ileNet
to
5
1
%.
A
weig
h
ted
lo
s
s
f
u
n
ctio
n
wa
s
u
s
ed
to
m
itig
ate
class
im
b
alan
ce
:
m
elan
o
m
a
(
a
m
in
o
r
ity
class
)
was
p
r
io
r
itized
d
u
r
in
g
tr
ain
in
g
.
D
esp
ite
th
e
u
n
b
ala
n
ce
d
d
ata
(
m
elan
o
m
a
r
e
p
r
esen
ts
ap
p
r
o
x
i
m
ately
1
6
%
o
f
I
SIC
2
0
1
9
)
,
th
e
b
alan
ce
d
F1
s
co
r
e
r
ea
ch
ed
5
3
%.
T
h
e
tr
a
d
e
-
o
f
f
is
th
at
ac
cu
r
ac
y
(
5
5
%)
s
u
f
f
er
ed
f
r
o
m
a
h
ig
h
er
r
ec
all
(
5
1
%)
d
u
e
t
o
an
in
cr
ea
s
e
in
f
alse
p
o
s
itiv
es.
T
h
e
im
p
r
o
v
em
en
t
in
v
alid
atio
n
lo
s
s
r
ed
u
ctio
n
is
n
o
tab
le:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S I
n
t J Ro
b
&
Au
to
m
I
SS
N:
2722
-
2
5
8
6
Mu
lti
-
mo
d
a
l tra
n
s
fo
r
mer a
n
d
co
n
vo
lu
tio
n
a
l a
tten
tio
n
a
r
ch
it
ec
tu
r
es fo
r
mela
n
o
ma
…
(
Gu
i
d
o
u
m
A
min
a
)
143
v
alid
atio
n
lo
s
s
d
ec
r
ea
s
ed
f
r
o
m
0
.
3
4
3
(
in
itial)
t
o
0
.
3
1
7
(
f
in
al)
af
ter
f
in
e
-
tu
n
in
g
,
in
d
icatin
g
in
cr
ea
s
ed
g
en
er
alize
d
ca
p
ab
ilit
y
f
o
r
test
in
g
d
ata
with
8
5
% a
cc
u
r
ac
y
.
T
h
e
MM
T
N
m
o
d
el
s
u
r
p
ass
es
o
th
er
m
o
d
els
th
an
k
s
to
its
ex
ce
p
tio
n
al
o
v
er
all
p
er
f
o
r
m
a
n
ce
i
n
ac
cu
r
ac
y
an
d
AUC,
ac
h
iev
in
g
th
e
h
ig
h
est
AUC
(
0
.
8
5
)
,
d
em
o
n
s
tr
atin
g
s
tr
o
n
g
class
s
ep
ar
ab
ilit
y
,
an
d
an
ac
c
u
r
ac
y
o
f
8
7
.
3
7
%,
r
ep
r
esen
tin
g
th
e
m
o
s
t
p
r
ec
is
e
p
r
ed
ictio
n
.
I
ts
ad
v
an
tag
e
lies
in
its
s
u
p
er
io
r
ab
ilit
y
to
b
alan
c
e
s
en
s
itiv
ity
an
d
s
p
ec
if
icity
co
m
p
ar
ed
to
its
r
iv
als.
I
t
s
h
o
ws
a
s
ig
n
if
ican
t
im
p
r
o
v
e
m
en
t
i
n
r
ec
all
f
o
r
C
lass
1
(
m
elan
o
m
a)
at
5
1
%,
n
ea
r
l
y
1
3
%
h
ig
h
er
th
a
n
Mo
b
ileNet
’
s
3
8
%,
th
u
s
in
cr
ea
s
in
g
th
e
n
u
m
b
er
o
f
tr
u
e
p
o
s
itiv
es,
wh
ich
is
cr
u
cial
in
m
ed
ical
d
iag
n
o
s
is
wh
er
e
a
d
iag
n
o
s
tic
f
ailu
r
e
ca
n
h
av
e
s
er
io
u
s
co
n
s
eq
u
en
ce
s
.
I
t
also
ac
h
iev
es
th
e
h
i
g
h
est
b
alan
ce
d
F1
s
co
r
e
f
o
r
C
lass
1
.
0
(
5
3
%
v
s
.
4
7
%
f
o
r
Mo
b
ileNet)
,
in
d
ica
tin
g
a
b
etter
tr
ad
e
-
o
f
f
b
etwe
en
p
r
ec
is
io
n
an
d
r
e
ca
ll
an
d
av
o
id
in
g
an
o
v
e
r
-
r
el
ian
ce
o
n
p
r
ec
is
io
n
th
at
wo
u
l
d
p
r
io
r
itize
a
lo
w
n
u
m
b
er
o
f
f
alse p
o
s
itiv
es a
t
th
e
ex
p
en
s
e
o
f
tr
u
e
p
o
s
itiv
es.
C
o
m
p
ar
ed
to
th
e
r
ef
e
r
en
ce
m
o
d
els E
f
f
icien
tNet
an
d
ViT
,
wh
ich
ex
h
ib
it
ca
tast
r
o
p
h
ic
f
ailu
r
e
f
o
r
class
1
.
0
(
r
ec
all
≤
2
4
%),
th
e
MM
T
N
p
r
o
v
es
to
b
e
s
ig
n
if
ican
tly
m
o
r
e
r
o
b
u
s
t.
Fo
r
c
r
itical
ca
s
es
,
it p
r
io
r
itizes r
ec
all
o
v
e
r
ac
cu
r
ac
y
to
o
p
tim
ize
p
r
ac
tical
clin
i
ca
l v
alu
e.
I
n
co
n
clu
s
io
n
,
th
e
m
o
d
el
co
m
b
in
es
atten
tio
n
-
by
-
t
r
an
s
f
o
r
m
i
n
g
lesi
o
n
-
f
o
cu
s
in
g
with
th
e
ef
f
icien
cy
o
f
Mo
b
ileNet.
Ag
e
-
r
elate
d
r
is
k
i
s
an
ex
am
p
le
o
f
h
o
w
clin
ical
d
ata
b
r
id
g
es
th
e
g
ap
s
in
im
a
g
e
-
o
n
ly
m
o
d
els.
T
h
e
d
ataset
im
b
alan
ce
(
m
elan
o
m
a
r
ar
ity
)
was
c
o
m
p
en
s
ated
f
o
r
b
y
class
weig
h
tin
g
in
th
e
lo
s
s
f
u
n
ctio
n
d
esig
n
.
Ou
r
MM
T
N
m
o
d
el
ac
h
iev
es
clin
i
ca
lly
s
ig
n
if
ican
t
p
er
f
o
r
m
an
ce
(
AUC=0
.
8
5
)
b
y
e
f
f
ec
tiv
ely
le
v
er
ag
in
g
atte
n
tio
n
-
by
-
tr
an
s
f
o
r
m
in
g
an
d
m
u
ltimo
d
al
lear
n
in
g
(
im
a
g
es
+
clin
ic
al
d
ata)
.
T
h
e
em
p
h
asis
o
n
r
e
ca
ll
alig
n
s
with
th
e
v
ital
g
o
al
o
f
ea
r
ly
m
elan
o
m
a
d
etec
tio
n
,
alth
o
u
g
h
ac
c
u
r
ac
y
f
o
r
m
ela
n
o
m
a
n
ee
d
s
im
p
r
o
v
e
m
en
t.
T
h
is
co
n
f
ir
m
s
o
u
r
ass
er
tio
n
th
at
m
u
ltimo
d
al
ar
ch
itectu
r
es o
u
tp
er
f
o
r
m
s
in
g
l
e
-
m
o
d
ality
tech
n
i
q
u
es in
ar
tific
ial
in
tellig
en
ce
f
o
r
d
er
m
ato
lo
g
y
.
T
o
p
r
o
v
id
e
a
co
m
p
ar
ativ
e
m
e
asu
r
e
o
f
r
o
b
u
s
tn
ess
,
th
e
MM
T
N
m
o
d
el
was
tr
ain
ed
an
d
ev
al
u
ated
th
r
e
e
tim
es
with
d
if
f
er
en
t
r
an
d
o
m
s
ee
d
s
.
T
h
e
p
er
f
o
r
m
an
ce
m
et
r
ics
r
ep
o
r
ted
r
ep
r
esen
t
th
e
m
ea
n
v
alu
es
ac
r
o
s
s
th
ese
r
u
n
s
.
T
h
e
m
o
d
el
ac
h
iev
ed
an
av
er
ag
e
AUC
o
f
0
.
8
5
±
0
.
0
1
5
an
d
an
av
er
a
g
e
test
ac
cu
r
ac
y
o
f
8
7
.
3
7
%
±
0
.
3
%,
in
d
icatin
g
s
tab
le
p
er
f
o
r
m
a
n
ce
.
Fig
u
r
e
4
.
C
o
n
f
u
s
io
n
m
atr
i
x
f
o
r
MM
T
N
Fig
u
r
e
5
.
AUC cu
r
v
e
f
o
r
MM
T
N
T
ab
le
1
.
An
aly
s
is
o
f
p
er
f
o
r
m
a
n
ce
m
etr
ics f
o
r
d
if
f
er
e
n
t m
o
d
e
ls
M
o
d
e
l
AUC
(
t
e
s
t
)
A
c
c
u
r
a
c
y
(
t
e
st
)
P
r
e
c
i
s
i
o
n
F
1
sc
o
r
e
R
e
c
a
l
l
0
.
0
1
.
0
0
.
0
1
.
0
0
.
0
1
.
0
V
i
t
63
81
84
24
91
10
90
10
Ef
f
i
c
i
e
n
t
N
e
t
62
82
84
10
91
10
90
10
M
o
b
i
l
e
N
e
t
82
85
89
62
92
47
92
38
M
M
TN
8
5
±
0
.
0
1
5
8
7
.
3
7
±
0
.
3
%
91
55
93
53
92
51
C
A
M
M
0
.
8
7
±
0
.
0
2
80
82
76
86
67
91
59
3
.
2
.
E
v
a
lua
t
i
o
n o
f
re
s
ults f
o
r
CAM
M
T
h
e
m
o
d
el
d
em
o
n
s
tr
ates
ex
ce
p
tio
n
al
p
e
r
f
o
r
m
an
ce
in
id
en
tif
y
in
g
n
o
n
-
m
elan
o
m
a
lesi
o
n
s
,
a
ch
iev
in
g
a
h
ig
h
r
ec
all
(
0
.
9
1
)
a
n
d
a
b
ala
n
c
ed
F1
s
co
r
e
(
0
.
8
6
)
.
T
h
is
tr
a
n
s
lates
to
a
l
o
w
f
alse
-
n
e
g
ativ
e
r
at
e
f
o
r
b
e
n
ig
n
ca
s
es,
en
s
u
r
in
g
th
at
m
o
s
t
h
a
r
m
less
lesi
o
n
s
ar
e
co
r
r
ec
tly
r
u
le
d
o
u
t,
wh
ich
is
ef
f
ec
tiv
e
f
o
r
s
cr
ee
n
i
n
g
.
Fo
r
th
e
cr
itical
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2722
-
2
5
8
6
I
AE
S I
n
t J Ro
b
&
Au
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
1
36
-
1
48
144
m
elan
o
m
a
class
,
th
e
m
o
d
el
ex
h
ib
its
h
ig
h
ac
c
u
r
ac
y
(
0
.
7
6
)
b
u
t
m
o
d
er
ate
r
ec
all
(
0
.
5
9
)
.
T
h
is
in
d
icate
s
th
at
wh
en
C
AM
M
p
r
ed
icts
m
elan
o
m
a,
it
is
co
r
r
ec
t
7
6
%
o
f
th
e
tim
e,
th
u
s
m
in
im
izin
g
u
n
n
ec
ess
ar
y
b
io
p
s
ies
(
f
alse
p
o
s
itiv
es).
Ho
wev
er
,
its
s
en
s
itiv
ity
o
f
5
9
%
m
ea
n
s
t
h
at
4
1
%
o
f
tr
u
e
m
elan
o
m
as
ar
e
m
is
s
ed
(
f
alse
n
eg
ativ
es),
r
ep
r
esen
tin
g
th
e
m
ain
lim
itatio
n
f
o
r
f
u
lly
s
tan
d
alo
n
e
d
iag
n
o
s
is
.
T
h
e
d
is
tr
ib
u
tio
n
o
f
co
r
r
ec
t
an
d
in
c
o
r
r
ec
t
p
r
ed
ictio
n
s
ac
r
o
s
s
b
o
th
class
es
ca
n
b
e
o
b
s
er
v
ed
i
n
th
e
c
o
n
f
u
s
io
n
m
atr
ix
s
h
o
wn
in
Fig
u
r
e
6
,
wh
ich
h
ig
h
lig
h
ts
th
e
r
elativ
ely
h
ig
h
er
n
u
m
b
e
r
o
f
m
is
s
ed
m
elan
o
m
a
ca
s
es
co
m
p
ar
ed
t
o
m
is
class
if
ied
n
o
n
-
m
elan
o
m
a
s
am
p
les
T
h
is
p
er
f
o
r
m
a
n
ce
p
r
o
f
ile
is
in
ten
tio
n
al.
W
ith
an
o
v
er
all
AUC
>
0
.
8
7
,
th
e
m
o
d
el
’
s
d
is
cr
im
in
ato
r
y
p
o
wer
f
alls
with
in
th
e
r
an
g
e
o
f
r
ec
en
t
s
tate
-
of
-
th
e
-
ar
t
m
o
d
els
(
AUC
~0
.
8
5
–
0
.
9
1
)
o
n
th
e
I
SIC
d
ataset.
Fu
r
th
er
m
o
r
e,
ac
h
iev
in
g
9
5
%
o
f
t
h
e
p
er
f
o
r
m
an
ce
o
f
lar
g
er
m
o
d
els
lik
e
E
f
f
icien
tNet
with
th
r
ee
tim
e
s
f
ewe
r
p
ar
am
eter
s
(
u
s
in
g
Mo
b
ileNetV2
)
u
n
d
er
s
co
r
es
its
ef
f
ec
tiv
en
ess
.
T
h
e
h
ig
h
r
ec
all
f
o
r
th
e
n
o
n
-
m
ela
n
o
m
a
class
(
0
.
9
1
)
en
ab
les
ef
f
icien
t
tr
iag
e
b
y
r
eli
ab
ly
f
ilter
in
g
o
u
t
b
en
i
g
n
ca
s
es.
W
h
ile
th
e
m
o
d
er
ate
r
ec
all
f
o
r
m
elan
o
m
a
lim
its
its
u
s
e
as
a
s
tan
d
alo
n
e
d
iag
n
o
s
tic
to
o
l,
it
r
em
ai
n
s
v
alu
a
b
le
as
a
clin
ical
d
ec
is
io
n
s
u
p
p
o
r
t
s
y
s
tem
.
T
h
e
g
en
er
ated
atte
n
tio
n
m
a
p
s
as
s
h
o
wn
in
Fig
u
r
e
7
p
r
o
v
i
d
e
in
t
er
p
r
etab
ilit
y
,
allo
win
g
clin
ici
an
s
to
v
is
u
alize
th
e
m
o
d
el
’
s
f
o
c
u
s
ar
ea
s
an
d
ef
f
ec
t
iv
ely
in
teg
r
ate
its
r
esu
lts
in
to
t
h
eir
ex
p
e
r
tis
e.
Fig
u
r
e
6
.
Ma
tr
ix
co
n
f
u
s
io
n
f
o
r
C
AM
M
Fig
u
r
e
7
.
Atten
tio
n
m
ap
v
is
u
aliza
tio
n
f
o
r
m
elan
o
m
a
d
etec
tio
n
T
o
en
s
u
r
e
th
e
s
tatis
tical
r
o
b
u
s
tn
ess
o
f
o
u
r
r
esu
lts
an
d
to
ac
co
u
n
t
f
o
r
tr
ain
i
n
g
v
ar
iab
ilit
y
,
t
h
e
C
AM
M
m
o
d
el
was
tr
ain
ed
an
d
ev
alu
at
ed
f
iv
e
tim
es
with
d
if
f
er
en
t
r
an
d
o
m
s
ee
d
s
.
Per
f
o
r
m
a
n
ce
is
p
r
esen
ted
as
m
ea
n
±
s
tan
d
ar
d
d
ev
iatio
n
as
p
r
esen
ted
in
T
ab
le
2
.
T
h
e
m
o
d
el
ac
h
i
ev
ed
a
m
ea
n
AUC
o
f
0
.
8
7
±
0
.
0
2
,
d
em
o
n
s
tr
atin
g
co
n
s
is
ten
t
d
is
cr
im
in
atio
n
ca
p
ab
ilit
y
(
ex
h
ib
itin
g
g
r
ea
ter
co
n
s
is
ten
cy
co
m
p
ar
ed
to
th
e
M
MT
N
m
o
d
el
)
.
T
h
ese
co
n
f
id
en
ce
in
ter
v
als
in
d
icate
s
tab
le
p
er
f
o
r
m
an
ce
d
esp
ite
th
e
in
h
er
en
t
r
an
d
o
m
n
ess
o
f
weig
h
t
in
itializatio
n
an
d
d
ata
s
h
u
f
f
lin
g
d
u
r
in
g
tr
ain
i
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S I
n
t J Ro
b
&
Au
to
m
I
SS
N:
2722
-
2
5
8
6
Mu
lti
-
mo
d
a
l tra
n
s
fo
r
mer a
n
d
co
n
vo
lu
tio
n
a
l a
tten
tio
n
a
r
ch
it
ec
tu
r
es fo
r
mela
n
o
ma
…
(
Gu
i
d
o
u
m
A
min
a
)
145
T
ab
le
2
.
An
aly
s
is
o
f
p
er
f
o
r
m
a
n
ce
m
etr
ics f
o
r
C
AM
M
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
n
-
M
e
l
a
n
o
m
a
0
.
8
2
0
.
9
1
0
.
8
6
1
3
5
7
M
e
l
a
n
o
m
a
0
.
7
6
0
.
5
9
0
.
6
7
6
7
8
T
h
e
atten
tio
n
m
ap
v
is
u
aliza
tio
n
[
2
8
]
p
r
esen
ts
two
elem
en
ts
.
T
h
e
lef
t
p
an
el
s
h
o
ws
th
e
o
r
i
g
in
al
in
p
u
t
im
ag
e
o
f
th
e
s
k
in
lesi
o
n
.
T
h
e
r
ig
h
t
p
a
n
el
o
v
e
r
lay
s
an
atten
t
io
n
m
ap
i
n
th
e
f
o
r
m
o
f
a
c
o
lo
r
-
co
d
e
d
h
ea
t
m
a
p
,
wh
ich
in
d
icate
s
th
e
m
o
d
el
’
s
f
o
cu
s
ar
ea
s
.
R
ed
,
o
r
an
g
e,
a
n
d
y
ello
w
s
ig
n
al
h
ig
h
atten
tio
n
.
B
lu
e
an
d
p
u
r
p
le
in
d
icate
r
eg
io
n
s
to
wh
ich
th
e
m
o
d
el
p
aid
litt
le
atten
tio
n
.
Gr
ee
n
co
r
r
esp
o
n
d
s
to
a
m
o
d
e
r
ate
lev
el
o
f
f
o
cu
s
.
3
.
3
.
Dis
cus
s
io
n o
n t
he
inte
g
ra
t
io
n o
f
a
uto
ma
t
ed
dia
g
no
s
t
i
cs
T
h
e
p
er
f
o
r
m
an
ce
d
em
o
n
s
tr
ated
b
y
o
u
r
m
u
ltimo
d
al
ar
c
h
itectu
r
es,
co
m
b
in
e
d
with
th
eir
co
m
p
u
tatio
n
a
l
ef
f
icien
cy
,
s
u
g
g
ests
th
eir
p
o
t
en
tial
u
tili
ty
in
au
to
m
ated
d
iag
n
o
s
tic
s
y
s
tem
s
.
I
n
s
u
ch
a
co
n
tex
t,
o
u
r
m
o
d
els
co
u
ld
b
e
in
teg
r
ated
in
to
clin
ic
al
wo
r
k
f
lo
ws
wh
e
r
e
d
er
m
o
s
co
p
ic
im
ag
e
ac
q
u
is
itio
n
an
d
p
ati
en
t
d
ata
co
llectio
n
ar
e
au
to
m
ated
.
Fo
r
e
x
am
p
le,
i
n
teled
er
m
ato
lo
g
y
o
r
m
o
b
ile
s
cr
ee
n
in
g
s
ce
n
ar
io
s
,
an
o
p
er
at
o
r
co
u
ld
ca
p
t
u
r
e
an
im
ag
e
an
d
in
p
u
t
b
asic
clin
ical
m
etad
ata
to
o
b
tain
an
au
t
o
m
ated
p
r
elim
in
ar
y
ass
ess
m
e
n
t.
T
h
e
m
u
ltimo
d
al
n
atu
r
e
o
f
o
u
r
m
o
d
els,
p
r
o
ce
s
s
in
g
b
o
th
v
is
u
al
a
n
d
co
n
te
x
tu
al
in
f
o
r
m
atio
n
,
is
well
-
s
u
ited
to
th
ese
ap
p
licatio
n
s
.
Fu
r
th
er
m
o
r
e
,
th
e
g
e
n
er
ated
att
en
tio
n
m
ap
s
o
f
f
er
a
d
eg
r
ee
o
f
in
ter
p
r
etab
ilit
y
,
wh
ich
c
o
u
ld
e
n
h
an
ce
co
n
f
id
e
n
ce
i
n
a
u
t
o
m
a
t
e
d
s
y
s
t
e
m
s
.
T
h
es
e
p
r
o
s
p
e
c
ts
w
o
u
l
d
r
e
q
u
i
r
e
f
u
r
t
h
e
r
v
a
l
i
d
a
t
i
o
n
w
o
r
k
a
n
d
s
p
ec
i
f
i
c
t
e
ch
n
i
c
a
l
i
n
t
e
g
r
a
ti
o
n
.
3
.
4
.
Co
m
pa
riso
n wit
h e
x
is
t
ing
m
ultim
o
da
l
a
pp
ro
a
ches o
n t
he
I
SI
C
da
t
a
s
et
T
ab
le
3
c
o
m
p
a
r
es
o
u
r
wo
r
k
with
r
ec
en
t
s
tate
-
of
-
th
e
-
a
r
t
m
u
ltimo
d
al
m
et
h
o
d
s
o
n
th
e
I
SIC
2
0
1
9
d
ataset.
T
h
e
an
aly
s
is
r
ev
ea
ls
th
at
o
u
r
two
p
r
o
p
o
s
ed
ar
ch
itec
tu
r
es,
MM
T
N
an
d
C
AM
M,
ac
h
iev
e
co
m
p
etitiv
e
p
er
f
o
r
m
an
ce
wh
ile
in
tr
o
d
u
ci
n
g
in
n
o
v
ativ
e
co
n
tr
ib
u
tio
n
s
.
MM
T
N
ac
h
iev
es
th
e
h
ig
h
est
ac
cu
r
ac
y
(
8
7
.
3
7
%),
s
u
r
p
ass
in
g
u
n
im
o
d
al
m
et
h
o
d
s
s
u
ch
as
Du
alAu
to
E
L
M
an
d
C
S
-
AF.
T
h
is
v
alid
ates th
e
s
ig
n
if
ican
t c
o
n
tr
ib
u
tio
n
o
f
m
e
r
g
in
g
clin
ical
d
ata
(
ag
e,
s
ex
)
with
v
is
u
al
ch
ar
ac
ter
is
tics
v
ia
a
T
r
an
s
f
o
r
m
er
-
lik
e
ar
c
h
itectu
r
e.
C
AM
M,
alth
o
u
g
h
with
s
lig
h
t
ly
lo
wer
ac
cu
r
ac
y
(
8
0
.
3
%),
s
tan
d
s
o
u
t
f
o
r
its
ef
f
i
cien
cy
an
d
in
te
r
p
r
etab
ilit
y
.
I
ts
lig
h
tweig
h
t
a
r
ch
itectu
r
e
b
ase
d
o
n
Mo
b
ileNetV2
an
d
its
atten
tio
n
m
ec
h
an
is
m
(
C
B
AM
)
m
ak
e
it
p
ar
ticu
lar
ly
well
-
s
u
ited
f
o
r
em
b
ed
d
e
d
d
e
p
lo
y
m
en
t,
o
f
f
er
i
n
g
an
o
p
tim
al
b
alan
ce
b
etwe
en
p
er
f
o
r
m
an
ce
,
t
r
an
s
p
ar
en
c
y
,
an
d
co
m
p
u
tatio
n
al
ef
f
icien
cy
.
Un
lik
e
p
r
e
v
io
u
s
wo
r
k
th
at
f
o
cu
s
ed
m
ain
ly
o
n
u
n
im
o
d
a
l
im
p
r
o
v
em
en
t
o
r
co
m
p
lex
f
u
s
io
n
o
f
class
if
ier
s
,
o
u
r
ap
p
r
o
ac
h
es
d
e
m
o
n
s
tr
ate
th
at
s
tr
u
ctu
r
ed
an
d
tar
g
eted
m
u
ltimo
d
al
f
u
s
io
n
–
wh
eth
er
b
ased
o
n
tr
an
s
f
o
r
m
atio
n
al
atten
tio
n
(
MM
T
N)
o
r
co
n
v
o
l
u
tio
n
al
at
ten
tio
n
(
C
AM
M)
–
is
a
p
r
o
m
is
in
g
av
en
u
e
f
o
r
im
p
r
o
v
in
g
b
o
th
th
e
ac
c
u
r
ac
y
a
n
d
clin
ical
u
tili
ty
o
f
au
to
m
ate
d
m
elan
o
m
a
d
iag
n
o
s
tic
s
y
s
te
m
s
.
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
with
ex
i
s
tin
g
m
u
ltimo
d
al
ap
p
r
o
ac
h
es o
n
th
e
I
SIC d
ataset
M
e
t
h
o
d
M
a
i
n
a
r
c
h
i
t
e
c
t
u
r
e
M
o
d
a
l
i
t
i
e
s Us
e
d
A
c
c
u
r
a
c
y
(
Te
st
)
M
a
i
n
c
o
n
t
r
i
b
u
t
i
o
n
D
u
a
l
A
u
t
o
E
LM
[
1
0
]
D
u
a
l
a
u
t
o
-
e
n
c
o
d
e
r
s (F
F
T
+
sp
a
t
i
a
l
)
D
e
r
mo
sc
o
p
i
c
i
ma
g
e
o
n
l
y
8
6
.
6
8
%
U
se
o
f
F
o
u
r
i
e
r
t
r
a
n
sf
o
r
m
f
o
r
t
e
x
t
u
r
e
a
n
a
l
y
si
s
.
CS
-
A
F
(
A
c
t
i
v
e
F
u
si
o
n
)
[
1
7
]
En
se
mb
l
e
o
f
1
2
C
N
N
mo
d
e
l
s
(
a
c
t
i
v
e
f
u
s
i
o
n
)
D
e
r
mo
sc
o
p
i
c
i
ma
g
e
o
n
l
y
7
7
.
7
4
%
A
d
a
p
t
i
v
e
c
o
s
t
m
u
l
t
i
-
c
l
a
ss
i
f
i
e
r
f
u
s
i
o
n
f
r
a
mew
o
r
k
(
I
S
I
C
2
0
1
9
)
.
M
M
TN
(
O
u
r
w
o
r
k
)
M
u
l
t
i
m
o
d
a
l
t
r
a
n
sf
o
r
m
e
r
(
En
c
o
d
e
r
+
C
l
i
n
i
c
a
l
D
a
t
a
)
I
mag
e
+
A
g
e
+
S
e
x
8
7
.
3
7
%
F
i
r
st
u
se
o
f
a
t
r
a
n
sf
o
r
mer
-
l
i
k
e
e
n
c
o
d
e
r
f
o
r
i
ma
g
e
/
c
l
i
n
i
c
a
l
d
a
t
a
f
u
si
o
n
o
n
I
S
I
C
2
0
1
9
.
C
A
M
M
(
O
u
r
w
o
r
k
)
M
o
b
i
l
e
N
e
t
V
2
C
N
N
+
A
t
t
e
n
t
i
o
n
(
C
B
A
M
)
+
C
l
i
n
i
c
a
l
D
a
t
a
I
mag
e
+
A
g
e
+
S
e
x
8
0
.
3
%
Li
g
h
t
w
e
i
g
h
t
a
r
c
h
i
t
e
c
t
u
r
e
w
i
t
h
i
n
t
e
r
p
r
e
t
a
b
l
e
a
t
t
e
n
t
i
o
n
ma
p
s,
s
u
i
t
a
b
l
e
f
o
r
e
mb
e
d
d
e
d
d
e
p
l
o
y
me
n
t
.
3
.
5
.
Abla
t
io
n study
o
n t
he
c
o
ntr
ibu
t
io
n o
f
clinica
l
d
a
t
a
T
h
is
s
tu
d
y
p
r
esen
ts
th
e
r
esu
lts
o
f
an
ex
p
er
im
e
n
t
in
wh
ich
th
e
MM
T
N
an
d
C
AM
M
m
o
d
els
wer
e
r
etr
ain
ed
with
o
u
t
clin
ical
d
ata
(
im
ag
es
o
n
ly
)
.
T
h
e
r
esu
lts
as
p
r
esen
ted
in
T
ab
le
4
s
h
o
w
a
s
ig
n
if
ican
t
d
ec
r
ea
s
e
in
AUC (
e.
g
.
,
-
0
.
0
4
f
o
r
C
AM
M)
,
q
u
an
titativ
ely
c
o
n
f
ir
m
in
g
th
e
u
s
ef
u
ln
ess
o
f
m
u
ltimo
d
al
f
u
s
io
n
.
T
ab
le
4
.
R
esu
lts
o
f
th
e
ab
latio
n
s
tu
d
y
M
o
d
e
l
C
o
n
f
i
g
u
r
a
t
i
o
n
(
M
o
d
a
l
i
t
i
e
s)
A
U
C
(
Te
s
t
)
Δ A
U
C
R
e
c
a
l
l
(
M
e
l
a
n
o
m
a
)
P
r
e
c
i
s
i
o
n
(
M
e
l
a
n
o
m
a
)
M
M
TN
I
mag
e
+
C
l
i
n
i
c
a
l
D
a
t
a
0
.
8
5
+
0
.
0
0
.
5
1
0
.
5
5
M
M
TN
I
mag
e
o
n
l
y
0
.
8
1
-
0
.
0
4
0
.
4
2
0
.
5
8
C
A
M
M
I
mag
e
+
C
l
i
n
i
c
a
l
D
a
t
a
0
.
8
7
+
0
.
0
0
.
5
9
0
.
7
6
C
A
M
M
I
mag
e
o
n
l
y
0
.
8
3
-
0
.
0
4
0
.
5
4
0
.
7
1
Evaluation Warning : The document was created with Spire.PDF for Python.