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
5
,
No
.
2
,
A
p
r
il 2
0
2
6
,
p
p
.
1
6
3
2
~
1
6
4
5
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
2
.
p
p
1
6
3
2
-
1
6
4
5
1632
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Brea
st cancer
det
ection using
r
esid
ua
l DenseNe
ts in
deep
lea
rning
Na
g
a
na
nd
ini
G
urura
j
a
ra
o
,
Vis
hwa
na
t
h R.
H
uli
pa
lled
S
c
h
o
o
l
o
f
C
o
m
p
u
t
i
n
g
a
n
d
I
n
f
o
r
m
a
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
R
EV
A
U
n
i
v
e
r
s
i
t
y
,
B
e
n
g
a
l
u
r
u
,
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
Au
g
2
2
,
2
0
2
4
R
ev
is
ed
No
v
5
,
2
0
2
5
Acc
ep
ted
Feb
6
,
2
0
2
6
Bre
a
st
c
a
n
c
e
r,
th
e
lea
d
i
n
g
c
a
u
s
e
o
f
c
a
n
c
e
r
-
re
late
d
d
e
a
th
s
a
m
o
n
g
wo
m
e
n
g
lo
b
a
ll
y
,
re
q
u
ires
a
p
ro
m
p
t
a
n
d
p
re
c
ise
d
iag
n
o
sis
i
n
o
rd
e
r
t
o
i
n
c
re
a
se
su
rv
iv
a
l
ra
tes
v
ia
t
h
e
ra
p
y
.
T
h
e
re
is
a
p
o
ss
ib
i
li
ty
o
f
b
ias
a
n
d
i
n
c
o
n
siste
n
c
y
i
n
th
e
re
su
lt
s
o
f
trad
it
i
o
n
a
l
d
ia
g
n
o
st
ic
p
ro
c
e
d
u
re
s
li
k
e
m
a
m
m
o
g
ra
p
h
y
,
u
lt
ra
so
u
n
d
,
a
n
d
h
ist
o
l
o
g
ica
l
tes
ti
n
g
si
n
c
e
th
e
y
re
l
y
o
n
t
h
e
e
x
p
e
rti
se
o
f
ra
d
io
l
o
g
ists
a
n
d
p
a
th
o
lo
g
ists.
T
h
e
re
a
re
e
x
c
it
in
g
n
e
w
o
p
p
o
rt
u
n
i
ti
e
s
fo
r
b
re
a
st
c
a
n
c
e
r
d
iag
n
o
stics
t
o
b
e
e
n
h
a
n
c
e
d
b
y
a
rti
ficia
l
in
telli
g
e
n
c
e
(AI)
a
n
d
d
e
e
p
lea
rn
in
g
.
Th
e
p
u
r
p
o
se
o
f
th
is
re
se
a
rc
h
is
to
e
x
a
m
in
e
t
h
e
fe
a
sib
il
it
y
o
f
u
sin
g
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s
(CNN
s)
a
n
d
re
sid
u
a
l
d
e
n
se
n
e
two
rk
s
(Re
sD
e
n
se
Ne
ts)
u
se
d
fo
r
b
re
a
st
c
a
n
c
e
r
a
u
to
m
a
ted
d
e
tec
ti
o
n
in
m
e
d
ica
l
ima
g
e
s.
Be
c
a
u
se
o
f
t
h
e
ir
s
u
p
e
ri
o
r
c
a
p
a
c
it
y
to
lea
rn
h
iera
rc
h
ica
l
fe
a
t
u
re
s
fro
m
ra
w
ima
g
e
d
a
ta,
CNN
s
a
re
i
d
e
a
l
fo
r
m
e
d
ica
l
ima
g
e
i
n
terp
re
t
a
ti
o
n
.
By
in
c
lu
d
in
g
re
si
d
u
a
l
c
o
n
n
e
c
ti
o
n
s
,
w
h
ich
a
ll
o
w
fo
r
th
e
train
i
n
g
o
f
c
o
n
sid
e
ra
b
ly
d
e
e
p
e
r
m
o
d
e
ls,
Re
sD
e
n
se
Ne
ts
—
a
n
e
x
ten
sio
n
o
f
CNN
s
—
m
i
ti
g
a
te
th
e
p
ro
b
lem
o
f
v
a
n
ish
i
n
g
g
ra
d
ie
n
t
i
n
d
e
e
p
n
e
two
rk
s
.
Re
sD
e
n
se
Ne
t
a
n
d
CNN
s
c
o
n
sid
e
ra
b
l
y
e
n
h
a
n
c
e
th
e
a
c
c
u
ra
c
y
o
f
b
re
a
st
c
a
n
c
e
r
d
iag
n
o
sis
in
c
o
m
p
a
riso
n
to
c
o
n
v
e
n
ti
o
n
a
l
a
p
p
ro
a
c
h
e
s,
a
c
c
o
rd
in
g
t
o
t
h
e
fin
d
i
n
g
s.
No
tab
l
y
,
Re
sD
e
n
se
Ne
ts
o
u
tp
e
rfo
rm
o
th
e
r
ty
p
e
s
o
f
n
e
two
rk
s
b
e
c
a
u
se
th
e
y
a
re
a
b
le
to
lea
rn
in
tri
c
a
te an
d
n
u
a
n
c
e
d
p
ro
p
e
rti
e
s d
irec
tl
y
fr
o
m
th
e
d
a
ta.
K
ey
w
o
r
d
s
:
Ar
tific
ial
in
tellig
en
ce
B
r
ea
s
t c
an
ce
r
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Den
s
e
n
etwo
r
k
Ma
m
m
o
g
r
am
s
R
esid
u
al
Den
s
eNe
t
s
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
:
Nag
an
an
d
in
i
Gu
r
u
r
ajar
a
o
Sch
o
o
l o
f
C
o
m
p
u
tin
g
an
d
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
,
R
E
VA
Un
iv
er
s
ity
B
en
g
alu
r
u
,
Kar
n
atak
a,
I
n
d
ia
E
m
ail:
r
ao
n
ag
a
n
an
d
in
i@
g
m
ail
.
co
m
1.
I
NT
RO
D
UCT
I
O
N
B
r
ea
s
t
ca
n
ce
r
co
n
tin
u
es
to
b
e
o
n
e
o
f
th
e
m
o
s
t
co
m
m
o
n
an
d
p
o
ten
tially
f
atal
ca
n
ce
r
s
af
f
ec
tin
g
wo
m
en
wo
r
ld
wid
e.
Acc
o
r
d
in
g
to
t
h
e
W
o
r
ld
Hea
lth
Or
g
a
n
izatio
n
,
b
r
ea
s
t
ca
n
ce
r
ac
c
o
u
n
ts
f
o
r
a
s
ig
n
if
ican
t
p
r
o
p
o
r
tio
n
o
f
f
e
m
ale
ca
n
ce
r
-
r
elate
d
d
ea
th
s
,
em
p
h
asizin
g
th
e
im
p
o
r
tan
ce
o
f
ea
r
ly
d
etec
tio
n
an
d
ac
cu
r
ate
d
iag
n
o
s
is
.
T
im
ely
id
en
tific
ati
o
n
o
f
m
alig
n
an
t
tu
m
o
r
s
s
ig
n
if
ican
tly
im
p
r
o
v
es
tr
ea
tm
e
n
t
o
u
tco
m
es
an
d
s
u
r
v
iv
al
r
ates.
T
r
ad
itio
n
al
d
iag
n
o
s
tic
tech
n
iq
u
es
s
u
ch
as
m
am
m
o
g
r
a
p
h
y
,
u
ltra
s
o
u
n
d
im
ag
i
n
g
,
an
d
b
io
p
s
y
ar
e
wid
ely
u
s
ed
in
clin
ical
p
r
ac
tice;
h
o
wev
er
,
th
ese
ap
p
r
o
ac
h
es
o
f
ten
r
ely
h
ea
v
ily
o
n
th
e
ex
p
e
r
tis
e
o
f
r
ad
io
l
o
g
is
ts
an
d
p
ath
o
lo
g
is
ts
,
wh
ich
ca
n
lead
t
o
d
iag
n
o
s
tic
v
ar
iab
ilit
y
an
d
d
e
lay
s
in
tr
ea
tm
en
t d
ec
is
io
n
s
[
1
]
,
[
2
]
.
R
ec
en
t a
d
v
an
ce
m
en
ts
in
ar
tifi
cial
in
tellig
en
ce
(
AI
)
an
d
d
ee
p
lear
n
in
g
h
a
v
e
o
p
e
n
ed
n
ew
o
p
p
o
r
tu
n
ities
f
o
r
im
p
r
o
v
in
g
b
r
ea
s
t c
an
ce
r
d
etec
tio
n
an
d
d
iag
n
o
s
is
.
I
n
p
ar
t
icu
lar
,
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
two
r
k
s
(
C
NNs)
an
d
r
esid
u
al
n
etwo
r
k
s
(
R
esNets
)
h
av
e
d
em
o
n
s
tr
ated
r
em
ar
k
ab
l
e
s
u
cc
ess
in
m
ed
ical
im
ag
e
an
aly
s
is
d
u
e
to
th
eir
ab
ilit
y
to
a
u
to
m
atica
lly
e
x
tr
ac
t
co
m
p
lex
f
ea
tu
r
es
f
r
o
m
im
a
g
in
g
d
ata.
C
NN
-
b
ased
m
o
d
els
h
av
e
b
ee
n
wid
el
y
ad
o
p
ted
in
co
m
p
u
ter
v
is
io
n
a
p
p
licatio
n
s
an
d
h
av
e
s
h
o
wn
s
tr
o
n
g
p
er
f
o
r
m
an
ce
in
d
etec
tin
g
ab
n
o
r
m
alities
in
m
am
m
o
g
r
a
m
s
,
u
ltra
s
o
u
n
d
s
ca
n
s
,
an
d
h
is
to
p
ath
o
lo
g
y
im
ag
e
s
[
3
]
,
[
4
]
.
R
esNet
ar
ch
itectu
r
es
ex
ten
d
tr
ad
itio
n
al
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
B
r
ea
s
t c
a
n
ce
r
d
etec
tio
n
u
s
in
g
r
esid
u
a
l D
en
s
eNets
in
d
ee
p
lea
r
n
in
g
(
N
a
g
a
n
a
n
d
in
i G
u
r
u
r
a
j
a
r
a
o
)
1633
C
NN
m
o
d
els
b
y
in
tr
o
d
u
cin
g
r
esid
u
al
co
n
n
ec
tio
n
s
,
wh
ich
h
elp
ad
d
r
ess
th
e
v
a
n
is
h
in
g
g
r
a
d
ien
t
p
r
o
b
lem
an
d
en
ab
le
th
e
tr
ain
i
n
g
o
f
m
u
c
h
d
e
ep
er
n
eu
r
al
n
etwo
r
k
s
with
im
p
r
o
v
ed
ac
c
u
r
ac
y
an
d
s
tab
ilit
y
[
5
]
,
[
6
]
.
C
NNs
wer
e
o
r
ig
in
ally
d
ev
elo
p
ed
to
m
im
ic
th
e
h
ier
ar
ch
ical
p
r
o
ce
s
s
in
g
o
f
v
is
u
al
in
f
o
r
m
a
tio
n
in
th
e
h
u
m
an
b
r
ain
.
T
h
ese
n
etwo
r
k
s
co
n
s
is
t
o
f
m
u
ltip
le
lay
er
s
th
at
p
er
f
o
r
m
co
n
v
o
l
u
tio
n
o
p
e
r
a
tio
n
s
,
p
o
o
lin
g
,
a
n
d
n
o
n
lin
ea
r
ac
tiv
atio
n
f
u
n
ctio
n
s
to
lear
n
f
ea
tu
r
es
at
d
if
f
er
e
n
t
lev
els
o
f
ab
s
tr
ac
tio
n
.
T
h
r
o
u
g
h
th
is
m
u
lti
-
lay
er
ed
s
tr
u
ctu
r
e,
C
NNs
ca
n
ef
f
ec
tiv
e
ly
ca
p
tu
r
e
s
p
atial
p
atter
n
s
,
tex
tu
r
es,
an
d
s
tr
u
ctu
r
es
in
m
ed
ical
im
ag
es,
allo
win
g
th
em
to
d
is
tin
g
u
is
h
b
etwe
en
n
o
r
m
al
a
n
d
ca
n
ce
r
o
u
s
tis
s
u
es
with
h
ig
h
p
r
ec
is
io
n
[
7
]
,
[
8
]
.
T
h
e
in
tr
o
d
u
ctio
n
o
f
r
esid
u
al
lear
n
in
g
i
n
R
esNet
a
r
ch
itectu
r
es
f
u
r
th
er
en
h
an
ce
s
d
ee
p
lear
n
in
g
p
e
r
f
o
r
m
an
ce
b
y
allo
win
g
id
e
n
tity
m
ap
p
in
g
s
b
etwe
en
lay
er
s
,
wh
ich
f
ac
ilit
ate
ef
f
icien
t
g
r
ad
ien
t
p
r
o
p
a
g
atio
n
d
u
r
in
g
b
ac
k
p
r
o
p
ag
atio
n
an
d
e
n
ab
le
d
ee
p
er
n
etwo
r
k
s
to
b
e
tr
ain
ed
with
o
u
t d
eg
r
a
d
atio
n
in
p
er
f
o
r
m
an
ce
[
5
]
.
T
h
is
s
tu
d
y
in
v
esti
g
ates
th
e
p
o
ten
tial
o
f
C
NN
an
d
R
esNet
ar
ch
itectu
r
es
f
o
r
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
u
s
in
g
m
ed
ical
im
ag
in
g
d
atase
ts
.
T
h
e
f
o
cu
s
is
o
n
ap
p
ly
in
g
t
h
ese
d
ee
p
lear
n
in
g
tech
n
iq
u
es
to
v
ar
io
u
s
im
ag
in
g
m
o
d
alities
,
in
clu
d
in
g
m
am
m
o
g
r
am
s
,
u
ltra
s
o
u
n
d
im
ag
es,
an
d
h
is
to
p
ath
o
lo
g
y
s
lid
es.
B
y
lev
er
ag
i
n
g
t
h
e
ca
p
ab
ilit
ies
o
f
th
ese
ad
v
an
c
ed
n
eu
r
al
n
etwo
r
k
a
r
ch
itectu
r
es,
th
e
p
r
o
p
o
s
ed
m
o
d
els
aim
to
ac
cu
r
ately
d
if
f
er
en
tiate
b
etwe
en
m
alig
n
a
n
t
an
d
b
e
n
ig
n
b
r
ea
s
t
tis
s
u
es.
Usi
n
g
p
u
b
licly
av
ailab
le
d
atasets
an
d
ad
v
an
ce
d
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es,
t
h
e
s
tu
d
y
d
em
o
n
s
tr
ates
h
o
w
d
e
ep
lear
n
in
g
ap
p
r
o
ac
h
es
ca
n
i
m
p
r
o
v
e
d
iag
n
o
s
tic
ac
cu
r
ac
y
an
d
s
u
p
p
o
r
t c
lin
ical
d
ec
is
io
n
-
m
ak
in
g
p
r
o
ce
s
s
es
[
3
]
,
[
7
]
.
AI
-
p
o
wer
ed
d
iag
n
o
s
tic
s
y
s
tem
s
o
f
f
er
s
ev
er
al
ad
v
an
tag
es
i
n
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
.
T
h
ese
s
y
s
tem
s
ca
n
r
ed
u
ce
th
e
wo
r
k
lo
ad
o
f
h
ea
lth
ca
r
e
p
r
o
f
ess
io
n
als,
m
in
im
ize
h
u
m
an
er
r
o
r
,
an
d
p
r
o
v
id
e
co
n
s
is
ten
t
an
d
r
ep
r
o
d
u
cib
le
d
ia
g
n
o
s
tic
r
esu
lt
s
.
Fu
r
th
er
m
o
r
e
,
d
ee
p
lear
n
in
g
m
o
d
els
ar
e
ca
p
a
b
le
o
f
id
en
tif
y
in
g
s
u
b
tle
p
atter
n
s
with
in
m
ed
ical
im
ag
es
th
at
m
ay
b
e
d
if
f
icu
lt
f
o
r
h
u
m
an
o
b
s
e
r
v
er
s
to
d
etec
t,
th
er
e
b
y
im
p
r
o
v
in
g
t
h
e
c
h
an
ce
s
o
f
ea
r
ly
d
iag
n
o
s
is
an
d
tim
ely
tr
ea
tm
en
t
[
1
]
,
[
4
]
.
T
h
e
f
o
llo
win
g
s
ec
tio
n
s
p
r
o
v
i
d
e
a
s
u
r
v
ey
o
n
v
ar
io
u
s
AI
b
ased
b
r
ea
s
t
ca
n
ce
r
tech
n
iq
u
es
an
d
an
ex
h
au
s
tiv
e
ac
co
u
n
t
o
f
t
h
e
ap
p
r
o
ac
h
,
e
v
er
y
th
in
g
f
r
o
m
g
at
h
er
in
g
an
d
clea
n
in
g
d
ata
to
cr
ea
ti
n
g
m
o
d
els,
tr
ain
in
g
p
r
o
ce
d
u
r
es,
an
d
ev
al
u
atio
n
to
o
ls
.
T
h
e
f
o
llo
win
g
tex
t
s
h
o
w
ca
s
es
th
e
o
u
tco
m
es
o
f
o
u
r
te
s
ts
,
wh
er
e
we
h
av
e
co
m
p
ar
ed
th
e
ef
f
icien
cy
o
f
C
NN
an
d
R
esNet
m
o
d
els.
Ad
d
itio
n
ally
,
we
d
elv
e
i
n
to
th
e
s
ig
n
if
ican
ce
o
f
th
ese
f
in
d
in
g
s
f
o
r
clin
ical
p
r
ac
tice.
I
n
th
is
im
p
o
r
tan
t
ar
ea
o
f
h
ea
lt
h
ca
r
e,
o
u
r
s
tu
d
y
'
s
f
in
d
in
g
s
h
ig
h
lig
h
t
th
e
p
o
ten
tial
ef
f
ec
t
o
f
A
I
-
p
o
we
r
ed
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
an
d
p
o
in
t
to
p
o
te
n
tial
av
en
u
es
f
o
r
f
u
r
th
er
r
esear
ch
an
d
d
ev
elo
p
m
e
n
t.
A
m
ajo
r
s
tep
f
o
r
war
d
in
th
e
f
i
g
h
t
ag
ain
s
t
b
r
e
ast
ca
n
ce
r
h
as
b
ee
n
th
e
in
c
o
r
p
o
r
atio
n
o
f
AI
in
to
th
e
d
iag
n
o
s
tic
p
r
o
ce
s
s
,
wh
ich
h
as
en
h
an
ce
d
d
iag
n
o
s
tic
ca
p
ab
ilit
ies
an
d
,
in
th
e
lo
n
g
r
u
n
,
p
r
o
d
u
ce
d
b
etter
p
atien
t o
u
tco
m
es.
2.
L
I
T
E
R
AT
U
RE
SU
RVE
Y
Ksh
ir
s
ag
ar
et
a
l.
[
9
]
u
s
es
d
e
ep
lear
n
in
g
tech
n
iq
u
es,
s
p
ec
i
f
ically
Mo
b
ileNetV2
ar
ch
itectu
r
e
with
tr
an
s
f
er
lear
n
in
g
,
to
an
al
y
ze
r
is
k
f
ac
to
r
s
in
b
r
ea
s
t
ca
n
ce
r
i
m
ag
in
g
.
T
h
e
ac
cu
r
ac
y
with
wh
ich
it
d
if
f
er
e
n
tiates
b
etwe
en
b
en
ig
n
an
d
m
alig
n
an
t
in
s
tan
ce
s
s
u
g
g
ests
p
r
o
m
is
in
g
d
ir
ec
tio
n
s
f
o
r
t
h
e
d
e
v
elo
p
m
en
t
o
f
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
s
y
s
tem
s
in
th
e
f
u
tu
r
e.
Ho
wev
er
,
th
e
s
tu
d
y
'
s
0
.
6
1
6
to
tal
ac
cu
r
ac
y
s
co
r
e
s
u
g
g
es
ts
co
m
p
lex
ity
an
d
r
o
o
m
f
o
r
im
p
r
o
v
em
e
n
t.
T
h
e
m
o
d
el'
s
p
r
ec
is
io
n
in
b
en
ig
n
s
i
tu
atio
n
s
is
also
in
f
lu
e
n
ce
d
b
y
s
o
cieta
l
an
d
cu
lt
u
r
al
f
ac
to
r
s
.
T
h
e
s
tu
d
y
'
s
g
en
er
aliz
ab
ilit
y
d
ep
en
d
s
o
n
th
e
a
v
ailab
ilit
y
an
d
q
u
ality
o
f
d
atasets
u
s
ed
,
an
d
th
e
s
tu
d
y
d
o
es
n
o
t
ad
d
r
ess
s
o
cieta
l
an
d
cu
ltu
r
al
f
ac
to
r
s
th
at
in
f
lu
en
ce
ac
ce
s
s
to
s
cr
ee
n
in
g
a
n
d
ca
r
e,
wh
ich
ar
e
ess
en
tial
asp
ec
ts
o
f
ea
r
ly
d
iag
n
o
s
is
.
C
io
b
o
tar
u
et
a
l.
[
1
0
]
p
r
o
p
o
s
e
a
d
ee
p
lear
n
in
g
-
b
ased
f
r
am
ew
o
r
k
f
o
r
b
r
ea
s
t
tu
m
o
r
class
if
icatio
n
u
s
in
g
C
NN
s
an
d
tr
an
s
f
er
lear
n
in
g
tech
n
iq
u
es
a
p
p
lied
t
o
u
ltra
s
o
u
n
d
im
ag
es.
T
h
e
s
tu
d
y
f
o
cu
s
e
s
o
n
m
u
lti
-
in
s
tan
ce
lear
n
in
g
t
o
im
p
r
o
v
e
th
e
clas
s
if
icatio
n
p
er
f
o
r
m
an
ce
o
f
b
r
ea
s
t
tu
m
o
r
im
a
g
es
b
y
ex
tr
ac
tin
g
d
is
cr
im
in
ativ
e
f
ea
tu
r
es
f
r
o
m
m
ed
ical
im
a
g
in
g
d
atasets
.
E
x
p
e
r
im
en
tal
r
e
s
u
lts
d
em
o
n
s
tr
ate
th
at
th
e
p
r
o
p
o
s
ed
C
NN
-
b
ased
ap
p
r
o
ac
h
im
p
r
o
v
es
class
if
icat
io
n
ac
cu
r
ac
y
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
b
en
ig
n
an
d
m
alig
n
an
t
b
r
ea
s
t
lesi
o
n
s
,
h
ig
h
lig
h
tin
g
th
e
ef
f
ec
tiv
en
ess
o
f
tr
an
s
f
er
lear
n
in
g
in
m
ed
i
ca
l
im
ag
e
an
aly
s
is
.
Ho
wev
er
,
th
e
s
tu
d
y
m
ain
ly
f
o
cu
s
es
o
n
u
ltra
s
o
u
n
d
im
a
g
in
g
d
atasets
an
d
d
o
es
n
o
t
ex
ten
s
iv
ely
ev
alu
ate
th
e
m
o
d
el
ac
r
o
s
s
m
u
ltip
le
im
ag
in
g
m
o
d
alities
s
u
ch
as
m
am
m
o
g
r
a
p
h
y
o
r
h
is
to
p
ath
o
lo
g
ical
im
ag
es.
Ad
d
itio
n
ally
,
f
u
r
th
er
in
v
esti
g
atio
n
is
r
eq
u
ir
ed
to
ass
es
s
th
e
m
o
d
el’
s
g
en
er
aliza
b
ilit
y
an
d
p
er
f
o
r
m
an
ce
in
d
i
v
er
s
e
clin
ical
en
v
ir
o
n
m
en
ts
with
lar
g
er
an
d
m
o
r
e
h
eter
o
g
en
e
o
u
s
d
atasets
.
Oy
elad
e
et
a
l.
[
1
1
]
in
tr
o
d
u
ce
s
T
win
C
NN
,
an
ap
p
r
o
ac
h
f
o
r
b
r
ea
s
t
ca
n
ce
r
im
ag
e
class
if
icatio
n
f
r
o
m
m
u
ltimo
d
al
d
ata
s
tr
ea
m
s
.
I
t
u
s
es
m
o
d
ality
-
b
ased
f
ea
t
u
r
e
lear
n
i
n
g
,
b
in
ar
y
o
p
tim
i
za
tio
n
f
o
r
f
ea
tu
r
e
d
im
en
s
io
n
ality
r
ed
u
ctio
n
,
an
d
a
n
ew
m
eth
o
d
f
o
r
f
ea
t
u
r
e
f
u
s
io
n
.
Acc
o
r
d
in
g
to
th
e
ex
p
er
im
en
tal
d
ata,
m
u
ltimo
d
al
class
if
icatio
n
o
u
tp
er
f
o
r
m
s
s
in
g
le
-
m
o
d
al
class
if
ic
atio
n
in
ter
m
s
o
f
b
o
th
class
if
icatio
n
ac
cu
r
ac
y
an
d
ar
ea
u
n
d
er
th
e
cu
r
v
e
(
AUC)
.
Ho
wev
er
,
th
e
s
tu
d
y
'
s
ev
alu
atio
n
is
b
ased
o
n
b
en
ch
m
a
r
k
d
a
tasets
,
wh
ich
m
ay
lim
it
its
g
en
er
aliza
b
ilit
y
to
r
e
al
-
wo
r
ld
d
atasets
.
T
h
e
s
p
ec
if
i
c
p
er
f
o
r
m
an
ce
c
h
ar
ac
ter
is
tics
an
d
co
m
p
u
tatio
n
al
r
eq
u
ir
em
e
n
ts
o
f
T
win
C
NN
ar
e
n
o
t e
x
ten
s
iv
ely
d
is
cu
s
s
ed
,
af
f
ec
tin
g
p
r
ac
tical
im
p
lem
e
n
tatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
3
2
-
1
6
4
5
1634
T
h
e
p
u
r
p
o
s
e
o
f
t
h
e
r
esear
ch
[
3
]
was
to
u
s
e
a
d
ee
p
lea
r
n
i
n
g
m
o
d
el
(
b
r
ea
s
t
ca
n
ce
r
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
B
C
C
NN)
)
t
o
ca
teg
o
r
ize
b
r
ea
s
t
tu
m
o
r
s
in
to
eig
h
t
g
r
o
u
p
s
in
o
r
d
er
t
o
aid
in
th
e
ea
r
ly
id
en
tific
atio
n
a
n
d
d
iag
n
o
s
is
o
f
th
e
d
is
ea
s
e.
W
e
u
s
ed
th
e
B
C
C
NN
m
o
d
el
to
g
eth
er
with
f
i
v
e
f
in
e
-
t
u
n
ed
m
o
d
els
to
ca
teg
o
r
ize
m
a
g
n
etic
r
eso
n
a
n
ce
im
ag
in
g
(
MRI)
im
ag
es
o
f
b
r
ea
s
t
ca
n
ce
r
.
W
ith
an
F1
-
s
co
r
e
o
f
9
8
.
2
8
%,
th
e
m
o
d
el
was
th
e
m
o
s
t
ac
cu
r
ate
.
T
h
e
d
etec
tio
n
an
d
class
if
ica
tio
n
o
f
b
r
ea
s
t
ca
n
ce
r
wer
e
g
r
ea
tly
im
p
r
o
v
ed
b
y
d
ataset
b
o
o
s
tin
g
,
p
r
ep
r
o
ce
s
s
in
g
,
an
d
b
alan
ci
n
g
.
T
h
e
g
r
ea
t
q
u
ality
o
f
MRI
s
ca
n
s
at
4
0
0
X
m
ag
n
i
f
icatio
n
allo
wed
f
o
r
th
e
h
ig
h
est
ac
c
u
r
ac
y
.
I
t
is
u
n
clea
r
h
o
w
well
th
e
m
o
d
el
p
er
f
o
r
m
s
in
v
ar
ie
d
p
atien
t
d
em
o
g
r
ap
h
ics
an
d
r
ea
l
-
wo
r
ld
clin
ical
s
itu
atio
n
s
,
s
in
ce
th
e
wo
r
k
r
elies
o
n
a
s
in
g
le
Kag
g
le
d
ataset,
wh
i
ch
m
ay
r
estrict
its
g
en
er
aliza
b
ilit
y
to
o
th
e
r
d
ata
s
ets.
Valid
atio
n
o
f
th
e
m
o
d
e
l's
ef
f
ec
tiv
en
ess
in
clin
ical
p
r
ac
tice
an
d
ac
r
o
s
s
d
if
f
er
en
t
im
ag
in
g
m
o
d
alities
r
eq
u
ir
es
m
o
r
e
i
n
v
esti
g
atio
n
.
No
o
n
e
k
n
o
ws
h
o
w
well
o
r
f
o
r
h
o
w
lo
n
g
t
h
e
s
u
g
g
ested
m
eth
o
d
o
lo
g
y
will wo
r
k
to
i
n
cr
ea
s
e
b
r
ea
s
t c
an
ce
r
s
u
r
v
iv
al
r
ates.
T
h
e
au
t
h
o
r
s
p
r
o
v
i
d
e
a
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
en
s
em
b
le
class
if
ier
[
1
2
]
t
h
at
u
s
es
d
ee
p
lear
n
in
g
,
tr
an
s
f
er
lear
n
in
g
m
o
d
els,
an
d
o
th
er
m
eth
o
d
s
.
T
h
e
ap
p
r
o
ac
h
wo
r
k
s
ad
m
ir
ab
l
y
o
n
u
ltra
s
o
u
n
d
d
atasets
an
d
attain
s
a
g
o
o
d
lev
el
o
f
ac
cu
r
ac
y
o
n
th
e
m
in
i
d
ig
ital
d
atab
ase
f
o
r
s
cr
ee
n
in
g
(
m
in
i
-
D
DSM
)
d
ataset.
I
ts
ad
ap
tab
ilit
y
an
d
d
e
p
en
d
a
b
ilit
y
m
ak
e
it
a
n
id
ea
l
ca
n
d
id
ate
f
o
r
m
u
ltimo
d
al
b
r
ea
s
t
ca
n
ce
r
d
ia
g
n
o
s
is
.
T
h
e
s
tu
d
y
'
s
r
esu
lts
,
h
o
wev
er
,
ar
e
o
n
ly
a
p
p
licab
le
to
th
e
d
atasets
u
s
ed
to
an
aly
ze
th
em
,
th
er
ef
o
r
e
th
eir
ap
p
licab
ilit
y
is
r
estricte
d
.
C
o
n
ce
r
n
s
ab
o
u
t
its
th
er
ap
eu
tic
ef
f
icac
y
ar
is
e
f
r
o
m
th
e
lack
o
f
d
is
cu
s
s
io
n
o
n
th
e
in
ter
p
r
eta
b
ilit
y
o
f
th
e
d
ee
p
lear
n
in
g
m
o
d
el
a
n
d
t
h
e
in
f
lu
en
ce
o
f
f
alse p
o
s
itiv
es a
n
d
f
alse n
eg
ativ
es o
n
p
atien
t
o
u
tco
m
es.
Usi
n
g
m
am
m
o
g
r
am
s
an
d
u
lt
r
aso
u
n
d
p
ictu
r
es,
M
u
d
u
li
et
a
l.
[
1
3
]
in
tr
o
d
u
ce
a
C
NN
m
o
d
el
f
o
r
au
to
m
ated
b
r
ea
s
t
ca
n
ce
r
ca
te
g
o
r
izatio
n
.
Au
to
m
atic
f
ea
tu
r
e
ex
tr
ac
tio
n
with
m
in
im
u
m
p
ar
am
eter
s
is
m
ad
e
p
o
s
s
ib
le
b
y
th
e
m
o
d
el'
s
f
o
u
r
co
n
v
o
lu
tio
n
al
lay
er
s
an
d
f
u
lly
lin
k
e
d
lay
e
r
.
I
t
o
u
tp
er
f
o
r
m
s
cu
r
r
e
n
t
m
eth
o
d
s
,
ac
h
iev
in
g
h
ig
h
ac
cu
r
ac
ies
o
n
v
ar
io
u
s
d
atasets
th
r
o
u
g
h
d
ata
au
g
m
en
tatio
n
to
m
itig
ate
o
v
e
r
f
itti
n
g
an
d
en
h
an
ce
g
en
er
aliza
tio
n
.
Ho
wev
er
,
th
e
p
ap
er
d
o
es
n
o
t
d
is
cu
s
s
p
o
ten
ti
al
d
r
awb
ac
k
s
o
r
lim
itatio
n
s
,
s
u
ch
as
p
er
f
o
r
m
an
ce
o
n
d
i
v
er
s
e
d
em
o
g
r
a
p
h
ics
o
r
im
ag
e
q
u
ality
v
a
r
iatio
n
s
,
a
n
d
d
o
es
n
o
t
m
en
tio
n
p
o
te
n
tial
co
m
p
u
tatio
n
a
l
r
eq
u
ir
em
e
n
ts
o
r
ef
f
icien
c
y
lim
itatio
n
s
,
wh
ich
co
u
ld
b
e
cr
u
cial
f
o
r
r
ea
l
-
w
o
r
ld
h
ea
lth
ca
r
e
d
e
p
lo
y
m
en
t.
T
o
au
to
m
ate
b
r
ea
s
t
ca
n
ce
r
p
r
ed
ictio
n
,
Ab
d
ar
et
a
l.
[
1
4
]
e
m
p
lo
y
s
a
lay
er
ed
en
s
em
b
le
m
eth
o
d
t
h
at
co
m
b
in
es
m
ac
h
in
e
lea
r
n
in
g
(
ML
)
an
d
d
ata
m
in
in
g
a
p
p
r
o
a
ch
es.
W
h
en
it
ca
m
e
to
d
is
tin
g
u
is
h
in
g
b
etwe
en
b
en
ig
n
an
d
m
alig
n
an
t
b
r
ea
s
t
ca
n
ce
r
s
,
th
e
two
-
lay
e
r
n
ested
en
s
em
b
le
class
if
ier
s
,
SV
-
B
ay
esNet
-
3
-
Me
taC
las
s
if
ier
an
d
SV
-
n
aïv
e
B
ay
es
-
3
-
Me
taC
las
s
if
ier
,
p
er
f
o
r
m
ed
b
etter
.
W
ith
an
ac
c
u
r
ac
y
o
f
9
8
.
0
7
%,
th
ese
m
o
d
els
b
lew
awa
y
s
in
g
le
class
if
ier
s
an
d
th
e
m
ajo
r
ity
o
f
p
r
io
r
wo
r
k
in
ex
p
er
im
en
ts
co
n
d
u
cte
d
o
n
th
e
W
is
co
n
s
in
d
iag
n
o
s
tic
b
r
ea
s
t
ca
n
ce
r
(
W
DB
C
)
d
ataset
u
s
in
g
th
e
k
-
f
o
l
d
cr
o
s
s
v
alid
atio
n
a
p
p
r
o
ac
h
.
H
o
wev
er
,
th
e
s
tu
d
y
'
s
f
in
d
in
g
s
ar
e
b
ase
d
o
n
th
e
W
DB
C
d
ataset
an
d
th
eir
g
en
er
aliza
b
ilit
y
to
o
th
er
d
atasets
is
s
ti
ll
u
n
k
n
o
wn
.
T
h
e
co
m
p
lex
n
atu
r
e
o
f
en
s
em
b
le
m
o
d
els
m
ay
lim
it
tr
an
s
p
ar
en
cy
an
d
in
ter
p
r
etab
ilit
y
,
p
o
s
in
g
ch
allen
g
es
f
o
r
clin
ical
ad
o
p
ti
o
n
an
d
tr
u
s
t.
T
h
e
p
o
ten
tial
p
r
esen
ce
o
f
b
iases
in
th
e
s
elec
ted
class
if
icatio
n
alg
o
r
ith
m
s
r
eq
u
ir
es
ca
r
ef
u
l
c
o
n
s
id
er
atio
n
,
esp
ec
ially
f
o
r
m
ar
g
in
alize
d
p
o
p
u
latio
n
s
.
C
o
m
p
a
r
is
o
n
with
p
r
ev
i
o
u
s
wo
r
k
s
m
ay
b
e
lim
ited
b
y
v
ar
ia
tio
n
s
in
ex
p
er
im
e
n
tal
s
etu
p
s
,
d
ata
p
r
ep
r
o
ce
s
s
in
g
,
an
d
e
v
alu
at
io
n
m
etr
ics.
A
n
o
v
el
d
ata
m
in
i
n
g
ap
p
r
o
ac
h
f
o
r
p
r
ec
is
e
b
r
ea
s
t
ca
n
ce
r
p
r
e
d
ictio
n
u
tili
zin
g
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es
(
SVMs)
an
d
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
(
ANNs)
o
n
W
is
co
n
s
in
b
r
ea
s
t
ca
n
ce
r
d
ataset
(
W
B
C
D
)
i
s
p
r
esen
ted
[
1
5
]
.
T
h
e
c
o
n
f
id
e
n
ce
-
weig
h
ted
v
o
tin
g
-
b
o
o
s
tin
g
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
CWV
-
B
ANNS
VM
)
m
o
d
el,
a
co
m
b
in
atio
n
o
f
ANNs
an
d
SVMs,
im
p
r
o
v
ed
p
er
f
o
r
m
a
n
ce
to
1
0
0
%
u
s
in
g
en
s
em
b
le
tech
n
i
q
u
es.
T
h
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
ev
alu
atio
n
r
elies
o
n
s
p
ec
if
ic
m
etr
ics,
p
o
te
n
tially
o
v
e
r
lo
o
k
i
n
g
o
th
er
asp
ec
ts
.
T
h
e
g
en
e
r
aliza
b
ilit
y
o
f
th
e
m
o
d
el
b
e
y
o
n
d
th
e
W
B
C
D
d
ataset
m
ay
b
e
lim
ite
d
,
an
d
th
e
f
o
cu
s
o
n
h
ig
h
ac
cu
r
ac
y
s
co
r
es
m
ay
n
o
t
tr
an
s
late
to
clin
ical
ap
p
licab
ilit
y
o
r
r
ea
l
-
wo
r
ld
e
f
f
ec
tiv
en
e
s
s
.
T
h
e
s
tu
d
y
also
lack
s
ex
ten
s
iv
e
d
is
cu
s
s
io
n
o
n
f
ea
tu
r
e
s
elec
tio
n
a
n
d
d
ata
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es,
in
ter
p
r
etab
ilit
y
,
a
n
d
s
ca
lab
ilit
y
an
d
co
m
p
u
tatio
n
al
ef
f
icien
cy
f
o
r
lar
g
e
-
s
ca
le
b
r
ea
s
t c
an
ce
r
p
r
ed
ictio
n
ap
p
licatio
n
s
.
Usi
n
g
th
e
B
r
ea
k
His
d
ataset
a
n
d
d
ee
p
n
eu
r
al
n
etwo
r
k
s
,
Alj
u
aid
et
a
l.
[
7
]
d
em
o
n
s
tr
ates
a
co
m
p
u
ter
-
aid
ed
d
iag
n
o
s
tic
ap
p
r
o
ac
h
f
o
r
b
r
ea
s
t
ca
n
ce
r
ca
teg
o
r
izatio
n
.
T
h
e
ap
p
r
o
ac
h
u
s
es
R
esNet,
I
n
ce
p
tio
n
-
V3
Net,
an
d
Sh
u
f
f
leNe
t
to
p
r
o
d
u
ce
h
ig
h
a
v
er
ag
e
ac
c
u
r
ac
ies
f
o
r
b
in
ar
y
a
n
d
m
u
lti
-
class
class
if
icatio
n
.
Ad
d
itio
n
al
r
esear
ch
is
n
ee
d
ed
t
o
s
ee
wh
eth
e
r
th
e
s
u
g
g
ested
s
tr
ateg
y
ca
n
b
e
a
p
p
lied
to
o
th
e
r
d
atasets
an
d
r
ea
l
-
wo
r
ld
clin
ical
co
n
tex
ts
,
s
in
ce
th
e
f
in
d
in
g
s
m
ig
h
t
b
e
im
p
ac
ted
b
y
th
e
co
n
te
n
t
an
d
q
u
ality
o
f
th
e
B
r
ea
k
His
d
ataset.
C
r
u
cial
in
m
ed
ical
d
iag
n
o
s
is
,
th
e
r
esear
ch
f
ails
to
p
r
o
v
id
e
lig
h
t
o
n
t
h
e
in
ter
p
r
etab
ilit
y
o
f
co
n
clu
s
io
n
s
m
ad
e
b
y
d
ee
p
lear
n
in
g
alg
o
r
ith
m
s
.
T
r
ain
i
n
g
d
ata
t
h
at
co
n
tain
s
b
iases
m
i
g
h
t
af
f
ec
t
h
o
w
well
t
h
e
m
o
d
el
p
er
f
o
r
m
s
in
th
e
ac
tu
al
wo
r
ld
.
T
h
e
s
u
g
g
ested
m
eth
o
d
'
s
r
esil
ien
ce
an
d
r
eliab
ilit
y
ac
r
o
s
s
m
u
ltip
le
d
em
o
g
r
a
p
h
ics
an
d
im
a
g
in
g
m
o
d
alities
m
u
s
t b
e
ev
alu
ated
v
ia
ex
ter
n
al
v
ali
d
atio
n
o
n
v
ar
i
ed
an
d
b
ig
g
er
d
atasets
.
I
n
o
r
d
er
to
a
u
to
m
ate
th
e
m
u
lti
-
class
if
icatio
n
o
f
b
r
ea
s
t
ca
n
c
er
h
is
to
lo
g
ical
p
ictu
r
es
o
n
th
e
B
r
ea
k
His
d
ataset,
Sh
ar
m
a
an
d
Me
h
r
a
[
1
6
]
co
n
tr
asts
h
an
d
-
cr
af
te
d
f
ea
tu
r
es
with
tr
an
s
f
er
lea
r
n
in
g
u
s
in
g
p
r
e
-
tr
ai
n
ed
n
etwo
r
k
s
.
R
eg
ar
d
less
o
f
th
e
m
ag
n
if
icatio
n
lev
el,
tr
a
n
s
f
er
l
ea
r
n
in
g
co
n
s
is
ten
tly
b
ea
ts
b
as
elin
e
tech
n
iq
u
es
an
d
cr
ea
ted
f
ea
tu
r
es.
W
h
en
it
ca
m
e
to
m
ag
n
if
icatio
n
-
d
ep
e
n
d
en
t
p
ictu
r
e
ca
teg
o
r
izatio
n
,
VGG1
6
with
lin
ea
r
SVM
was
th
e
m
o
s
t
ac
cu
r
ate.
T
h
e
two
m
o
s
t
d
if
f
icu
lt
ty
p
es
o
f
ca
n
ce
r
to
d
etec
t
at
v
ar
io
u
s
m
ag
n
if
icatio
n
s
ar
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
B
r
ea
s
t c
a
n
ce
r
d
etec
tio
n
u
s
in
g
r
esid
u
a
l D
en
s
eNets
in
d
ee
p
lea
r
n
in
g
(
N
a
g
a
n
a
n
d
in
i G
u
r
u
r
a
j
a
r
a
o
)
1635
f
ib
r
o
-
a
d
en
o
m
a
an
d
m
u
c
o
u
s
ca
r
cin
o
m
a.
Ho
wev
er
,
th
e
s
tu
d
y
'
s
f
o
cu
s
o
n
a
s
p
ec
if
ic
d
at
aset
m
ay
lim
it
its
g
en
er
aliza
b
ilit
y
.
I
t
also
d
id
n
o
t
co
n
s
id
er
o
th
er
ty
p
es
o
f
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
es,
t
h
e
im
p
ac
t
o
f
d
ata
im
b
alan
ce
with
in
d
if
f
er
e
n
t
class
es
o
n
cla
s
s
if
icatio
n
p
er
f
o
r
m
an
ce
,
th
e
in
ter
p
r
etab
ilit
y
o
f
th
e
m
o
d
els,
p
o
ten
tial
b
iases
in
th
e
d
ataset
o
r
m
o
d
el
p
r
ed
ictio
n
s
,
an
d
th
e
s
ca
la
b
ilit
y
an
d
co
m
p
u
tatio
n
al
r
es
o
u
r
ce
s
r
e
q
u
ir
ed
f
o
r
im
p
lem
en
tin
g
th
e
p
r
o
p
o
s
ed
m
o
d
els.
Usi
n
g
a
co
m
b
i
n
atio
n
o
f
tr
a
n
s
f
er
lear
n
in
g
an
d
p
r
e
-
tr
ain
e
d
C
NN
ar
ch
itectu
r
es
s
u
ch
a
s
R
esNet,
VGGN
et,
an
d
Go
o
g
L
eNe
t,
a
n
o
v
el
d
ee
p
lear
n
in
g
f
r
a
m
ewo
r
k
f
o
r
cy
to
lo
g
y
im
ag
e
d
etec
tio
n
an
d
b
r
ea
s
t
ca
n
ce
r
class
if
icatio
n
is
p
r
esen
ted
[
4
]
.
T
h
e
f
r
am
ewo
r
k
class
if
ies
ce
lls
as
eith
er
b
en
ig
n
o
r
m
ali
g
n
an
t
b
y
e
x
tr
ac
tin
g
in
f
o
r
m
atio
n
f
r
o
m
p
ictu
r
es
an
d
ap
p
ly
in
g
th
em
ac
r
o
s
s
a
f
u
lly
lin
k
ed
lay
er
.
W
h
en
co
m
p
ar
ed
to
ex
is
tin
g
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es,
th
e
f
r
a
m
ewo
r
k
a
ch
iev
es
b
etter
r
esu
lt
s
in
id
en
tif
y
in
g
an
d
ca
teg
o
r
izi
n
g
b
r
ea
s
t
ca
n
ce
r
s
in
ex
p
er
im
en
ts
co
n
d
u
cte
d
o
n
co
m
m
o
n
b
en
ch
m
ar
k
d
atasets
.
T
h
er
e
is
a
lac
k
o
f
d
is
cu
s
s
io
n
o
f
th
e
c
o
m
p
u
tin
g
r
eso
u
r
ce
s
n
ee
d
ed
f
o
r
im
p
lem
en
tatio
n
,
th
e
in
ter
p
r
etab
ilit
y
o
f
th
e
m
o
d
el'
s
d
ec
is
io
n
-
m
ak
in
g
p
r
o
ce
s
s
,
p
r
ec
is
e
f
alse
p
o
s
itiv
e
an
d
f
alse
n
eg
at
iv
e
r
ates,
an
d
m
o
r
e
in
th
e
p
ap
er
.
T
h
e
a
r
ticle
co
u
ld
b
e
n
ef
it
f
r
o
m
d
is
cu
s
s
in
g
p
o
ten
tial
b
iases
in
b
en
c
h
m
ar
k
d
atasets
an
d
th
e
s
tep
s
tak
en
to
m
itig
ate
th
em
.
A
d
d
itio
n
all
y
,
th
e
f
r
am
ewo
r
k
'
s
s
ca
lab
ilit
y
an
d
a
d
ap
tab
ilit
y
t
o
v
ar
y
i
n
g
im
ag
in
g
co
n
d
itio
n
s
an
d
e
q
u
ip
m
e
n
t
s
h
o
u
ld
b
e
e
x
p
lo
r
ed
to
en
s
u
r
e
its
r
o
b
u
s
tn
ess
ac
r
o
s
s
d
if
f
er
en
t
h
e
alth
ca
r
e
s
ettin
g
s
.
I
n
o
r
d
er
to
im
p
r
o
v
e
ea
r
ly
d
iag
n
o
s
is
o
f
b
r
ea
s
t
ca
n
ce
r
,
Alr
u
wa
ili
an
d
Go
u
d
a
[
5
]
in
v
esti
g
ates
th
e
u
s
e
o
f
d
ee
p
lear
n
in
g
m
o
d
els
in
m
am
m
o
g
r
a
p
h
y
.
W
e
u
tili
ze
tr
an
s
f
er
lear
n
in
g
t
o
d
is
tin
g
u
is
h
b
etwe
en
b
en
ig
n
a
n
d
m
alig
n
an
t
in
s
tan
ce
s
,
an
d
we
ap
p
ly
s
ev
er
al
au
g
m
e
n
tatio
n
p
r
o
ce
d
u
r
es
to
en
h
an
ce
th
e
am
o
u
n
t
o
f
p
ictu
r
es
a
n
d
m
in
im
ize
o
v
er
f
itti
n
g
.
Usi
n
g
R
esNe
t5
0
,
th
e
s
u
g
g
ested
s
y
s
tem
o
b
tain
s
an
ac
cu
r
ac
y
o
f
8
9
.
5
%
o
n
th
e
m
am
m
o
g
r
a
p
h
ic
im
a
g
e
a
n
aly
s
is
s
o
ciety
(
MI
AS
)
d
ataset,
w
h
er
ea
s
n
eu
r
al
ar
ch
itectu
r
e
s
ea
r
ch
n
etwo
r
k
-
m
o
b
ile
(
Nas
N
et
-
Mo
b
ile
)
ac
h
iev
es
an
ac
cu
r
ac
y
o
f
7
0
%.
Ho
wev
er
,
th
e
f
in
d
in
g
s
ar
e
lim
ited
to
s
p
ec
if
ic
d
atasets
an
d
m
o
d
els,
an
d
m
ay
n
o
t
g
en
er
ali
ze
to
o
th
e
r
d
atasets
o
r
ar
c
h
ite
ctu
r
es.
T
h
e
p
er
f
o
r
m
an
ce
e
v
al
u
atio
n
is
b
ased
o
n
ac
cu
r
ac
y
alo
n
e,
an
d
f
u
r
th
er
v
alid
atio
n
o
n
lar
g
er
d
atasets
is
n
ee
d
ed
.
T
h
e
s
tu
d
y
also
lack
s
in
s
ig
h
ts
in
to
th
e
in
ter
p
r
etab
ilit
y
o
f
d
ee
p
lear
n
i
n
g
m
o
d
els,
wh
ich
is
cr
u
cial
f
o
r
m
ed
ical
p
r
o
f
ess
io
n
als'
tr
u
s
t
.
T
h
e
co
m
p
u
tatio
n
al
r
eq
u
ir
em
e
n
ts
an
d
in
f
r
astru
ctu
r
e
f
o
r
d
e
p
lo
y
in
g
t
h
ese
m
o
d
els
i
n
clin
ical
s
ettin
g
s
a
r
e
n
o
t
a
d
d
r
ess
ed
,
lim
itin
g
th
e
p
r
ac
tical
ap
p
licab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
.
A
n
o
v
el
b
r
ea
s
t
ca
n
ce
r
m
ass
class
if
icat
io
n
m
o
d
el
b
ased
o
n
d
ee
p
co
n
v
o
lu
tio
n
al
n
eu
r
a
l
n
etwo
r
k
s
(
DC
NNs)
an
d
tr
an
s
f
er
lear
n
i
n
g
f
r
o
m
p
r
e
-
tr
ain
ed
m
o
d
els
s
u
ch
as
Go
o
g
leNe
t
a
n
d
Alex
Net
is
in
tr
o
d
u
ce
d
[
1
7
]
.
W
ith
b
etter
ac
cu
r
ac
y
an
d
AUC
r
atin
g
s
th
an
Go
o
g
leNe
t,
th
e
m
o
d
el
p
er
f
o
r
m
e
d
well
o
n
m
am
m
o
g
r
a
p
h
y
d
atasets
.
W
h
ile
th
e
s
tu
d
y
'
s
f
in
d
in
g
s
ar
e
p
r
o
m
is
in
g
,
th
e
r
e
ar
e
a
f
ew
ca
v
ea
ts
to
k
ee
p
in
m
in
d
b
e
f
o
r
e
im
p
lem
en
tin
g
th
em
i
n
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ettin
g
s
.
T
h
ese
in
clu
d
e
th
e
s
tu
d
y
'
s
n
ar
r
o
w
em
p
h
asis
o
n
m
am
m
o
g
r
a
p
h
ic
d
atasets
,
th
e
n
ee
d
f
o
r
ad
d
itio
n
al
v
ali
d
atio
n
o
n
lar
g
er
d
atasets
,
an
d
th
e
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
n
ee
d
e
d
to
tr
ain
a
n
d
t
est th
ese
d
ee
p
m
o
d
els.
T
h
e
s
u
r
v
ey
f
r
o
m
[
1
]
e
x
am
in
es
th
e
u
s
e
o
f
C
NNs
in
m
am
m
o
g
r
ap
h
y
,
h
ig
h
lig
h
tin
g
th
eir
s
tr
en
g
th
s
,
lim
itatio
n
s
,
an
d
p
er
f
o
r
m
a
n
ce
in
an
aly
zin
g
m
am
m
o
g
r
am
i
m
ag
es.
I
n
s
ig
h
ts
in
to
C
NN
ar
ch
itectu
r
e,
p
o
p
u
lar
p
u
b
lically
ac
ce
s
s
ib
le
m
a
m
m
o
g
r
am
s
r
e
p
o
s
ito
r
ies,
an
d
m
et
h
o
d
s
f
o
r
e
n
h
an
ci
n
g
t
h
e
ac
c
u
r
a
cy
o
f
d
iag
n
o
s
es
ar
e
p
r
esen
ted
.
I
n
ad
d
itio
n
to
h
elp
in
g
with
d
atab
ase
s
elec
tio
n
,
t
h
e
s
tu
d
y
g
o
es
o
v
e
r
o
th
e
r
m
et
h
o
d
s
f
o
r
im
p
r
o
v
in
g
C
NN
p
er
f
o
r
m
a
n
ce
,
s
u
ch
as
tr
a
n
s
f
er
lear
n
i
n
g
,
d
ata
a
u
g
m
e
n
tatio
n
,
b
atch
n
o
r
m
aliza
tio
n
,
an
d
d
r
o
p
o
u
t.
H
o
wev
er
,
it
m
ay
n
o
t
co
v
e
r
th
e
latest
d
ev
elo
p
m
en
ts
in
d
ee
p
lear
n
in
g
f
o
r
m
am
m
o
g
r
ap
h
y
,
o
m
itti
n
g
d
etailed
co
m
p
ar
ativ
e
an
aly
s
is
,
an
d
o
v
e
r
lo
o
k
i
n
g
p
r
o
p
r
ietar
y
o
r
s
p
ec
ialized
d
ataset
s
.
T
h
e
s
u
r
v
ey
'
s
lim
itatio
n
s
s
ec
tio
n
m
a
y
n
o
t
f
u
lly
ex
p
lo
r
e
p
r
ac
tical
co
n
s
tr
ain
ts
an
d
r
eso
u
r
ce
d
em
a
n
d
s
o
f
im
p
lem
en
tin
g
C
NN
-
b
ased
s
o
lu
t
io
n
s
in
r
ea
l
-
wo
r
l
d
clin
ical
s
ettin
g
s
.
T
o
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
o
f
d
i
g
ital
m
am
m
o
g
r
ap
h
y
in
th
e
d
ia
g
n
o
s
is
o
f
b
r
ea
s
t
ca
n
ce
r
,
a
m
u
l
tis
ca
le
all
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
MA
-
C
NN)
wa
s
cr
ea
ted
[
8
]
.
T
h
e
m
o
d
el
ac
cu
r
ately
class
if
ies
m
am
m
o
g
r
am
im
ag
es
as
n
o
r
m
al,
m
alig
n
an
t
,
o
r
b
e
n
ig
n
,
im
p
r
o
v
in
g
class
if
icatio
n
ac
cu
r
ac
y
th
r
o
u
g
h
m
u
ltis
ca
le
f
ilter
s
.
I
t
ac
h
iev
es
an
o
v
er
all
s
en
s
itiv
ity
o
f
9
6
%
an
d
a
0
.
9
9
AUC,
d
em
o
n
s
tr
atin
g
its
ef
f
ec
tiv
en
ess
in
ea
r
l
y
d
ia
g
n
o
s
is
.
T
h
e
f
ea
tu
r
e
lear
n
in
g
ca
p
ab
i
liti
es
o
f
C
NNs
en
h
an
ce
cla
s
s
if
icatio
n
ac
cu
r
ac
y
,
aid
in
g
in
p
r
o
p
er
clin
ical
tr
ea
tm
en
ts
an
d
im
p
r
o
v
in
g
s
u
r
v
iv
al
r
ates.
Nev
er
th
eless
,
MA
-
C
NN
'
s
ef
f
icac
y
co
u
ld
b
e
af
f
e
cted
b
y
a
n
u
m
b
er
o
f
f
ac
to
r
s
,
in
clu
d
in
g
th
e
v
ar
iety
a
n
d
q
u
ality
o
f
t
h
e
tr
ain
in
g
d
ata
s
et,
th
e
in
tr
icac
y
o
f
th
e
d
ee
p
l
ea
r
n
in
g
alg
o
r
ith
m
s
,
th
e
av
ailab
ilit
y
o
f
ac
ce
s
s
ib
le
ad
v
an
ce
d
im
ag
in
g
an
d
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
a
n
d
th
e
n
ec
ess
ity
o
f
o
n
g
o
i
n
g
u
p
d
ates
an
d
m
ain
ten
a
n
ce
to
ac
co
m
m
o
d
ate
ch
an
g
in
g
clin
ical
p
r
ac
tices
an
d
d
ev
elo
p
m
e
n
ts
in
b
r
ea
s
t
ca
n
ce
r
d
iag
n
o
s
is
.
Fo
r
th
e
p
u
r
p
o
s
e
o
f
id
en
tify
in
g
b
r
ea
s
t
lesi
o
n
s
u
s
in
g
d
ig
ital
X
-
r
a
y
m
am
m
o
g
r
a
m
s
,
a
co
m
p
u
ter
-
ai
d
ed
d
etec
tio
n
(
C
AD)
s
y
s
tem
b
ased
o
n
d
ee
p
lear
n
i
n
g
h
as
b
ee
n
cr
ea
ted
[
1
8
]
.
Fo
r
d
etec
tio
n
,
t
h
e
s
y
s
tem
em
p
lo
y
s
y
o
u
o
n
ly
lo
o
k
o
n
ce
(
YOL
O
)
; f
o
r
s
eg
m
en
tatio
n
,
it
em
p
lo
y
s
f
u
ll
-
r
eso
lu
tio
n
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
(
FrCN
)
;
an
d
f
o
r
class
if
icatio
n
,
it
em
p
lo
y
s
C
NN,
R
e
s
Net
-
5
0
,
an
d
I
n
ce
p
ti
o
n
R
esNet
-
V2
.
At
th
e
d
etec
tio
n
,
s
eg
m
en
tatio
n
,
an
d
class
if
icatio
n
p
h
ases
,
th
e
I
Nb
r
ea
s
t
d
atab
ase
o
u
tp
er
f
o
r
m
e
d
t
r
ad
itio
n
al
d
ee
p
lear
n
in
g
tech
n
iq
u
es.
T
h
e
s
y
s
tem
h
as
p
o
ten
tial
to
ass
is
t
r
ad
io
lo
g
is
ts
in
ac
cu
r
ately
d
iag
n
o
s
in
g
b
r
ea
s
t
lesi
o
n
s
.
Ho
wev
er
,
its
ef
f
ec
tiv
en
ess
in
r
ea
l
-
wo
r
ld
clin
ical
s
ettin
g
s
r
eq
u
ir
es
f
u
r
th
er
v
alid
atio
n
,
an
d
its
p
er
f
o
r
m
an
ce
m
etr
ics
m
ay
n
o
t
ca
p
tu
r
e
all
asp
ec
ts
o
f
clin
ical
r
elev
an
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
3
2
-
1
6
4
5
1636
I
n
o
r
d
er
to
id
e
n
tify
an
d
class
if
y
b
r
ea
s
t
lesi
o
n
s
,
Al
-
an
tar
i
e
t
a
l.
[
2
]
p
r
esen
ts
an
d
ass
ess
e
s
a
C
A
D
s
y
s
tem
th
at
is
in
teg
r
ated
an
d
u
s
es
d
ee
p
lear
n
in
g
.
Fo
r
m
am
m
o
g
r
am
s
f
r
o
m
th
e
DDSM
an
d
I
Nb
r
ea
s
t
d
atasets
,
th
e
YOL
O
d
etec
to
r
ac
h
iev
ed
e
x
ce
llen
t d
etec
tio
n
ac
cu
r
ac
ies a
n
d
F1
-
s
co
r
es.
W
h
en
it c
am
e
to
id
en
tify
in
g
b
r
ea
s
t
lesi
o
n
s
,
class
if
icatio
n
m
o
d
el
s
s
u
ch
as
C
NN,
R
esNet
-
5
0
,
an
d
I
n
ce
p
tio
n
R
esNet
-
V2
s
h
o
wn
e
n
co
u
r
ag
in
g
ac
cu
r
ac
y
.
T
h
e
YOL
O
d
etec
to
r
im
p
r
o
v
ed
class
if
icatio
n
m
o
d
e
ls
'
d
iag
n
o
s
tic
ca
p
ab
ilit
ies,
wh
ich
m
ay
h
av
e
aid
e
d
in
th
e
cr
ea
tio
n
o
f
a
r
eliab
le
C
AD
s
y
s
tem
f
o
r
th
e
d
etec
tio
n
o
f
b
r
ea
s
t
ca
n
ce
r
.
T
h
e
r
esea
r
ch
o
n
ly
em
p
lo
y
ed
DDSM
an
d
I
Nb
r
ea
s
t a
s
d
atase
ts
,
th
u
s
h
o
w
well
it d
o
es o
n
b
i
g
g
er
an
d
m
o
r
e
d
iv
er
s
e
d
atasets
is
u
n
clea
r
.
No
o
n
e
ad
d
r
ess
ed
th
e
co
m
p
u
tin
g
ef
f
ic
ien
cy
o
r
r
eso
u
r
ce
n
ee
d
s
o
f
im
p
lem
en
tin
g
s
u
ch
a
d
ee
p
lear
n
i
n
g
s
y
s
tem
in
ac
tu
al
h
ea
lth
ca
r
e
en
v
ir
o
n
m
en
ts
.
T
h
e
im
p
ac
t
o
f
f
alse
p
o
s
itiv
es
an
d
f
alse
n
eg
ativ
es
o
n
clin
ical
d
e
cisi
o
n
-
m
ak
in
g
was
n
o
t th
o
r
o
u
g
h
ly
an
al
y
ze
d
,
r
aisi
n
g
co
n
ce
r
n
s
ab
o
u
t th
e
s
y
s
tem
'
s
r
eliab
ilit
y
in
a
r
ea
l
-
wo
r
ld
s
ettin
g
.
E
a
r
l
y
d
ia
g
n
o
s
is
is
v
it
al
f
o
r
b
r
e
ast
c
an
ce
r
[
1
9
]
,
a
s
e
r
i
o
u
s
w
o
r
l
d
wi
d
e
h
ea
l
th
c
o
n
ce
r
n
.
T
h
e
c
at
eg
o
r
iz
ati
o
n
o
f
b
r
ea
s
t
ca
n
c
e
r
u
s
i
n
g
a
d
ee
p
e
n
s
e
m
b
le
t
r
an
s
f
e
r
le
a
r
n
in
g
/
n
eu
r
a
l
n
e
tw
o
r
k
s
c
lass
i
f
i
er
h
as
s
h
o
wn
en
c
o
u
r
a
g
i
n
g
o
u
tc
o
m
es
.
T
h
e
s
y
s
t
em
p
r
e
-
p
r
o
c
ess
es
m
a
m
m
o
g
r
am
im
ag
es,
e
x
t
r
ac
ts
f
e
at
u
r
es
u
s
i
n
g
a
n
e
n
s
em
b
le
m
o
d
el
,
a
n
d
th
e
n
u
s
es
a
n
e
u
r
al
n
et
wo
r
k
f
o
r
c
lass
i
f
i
ca
t
io
n
.
I
t
ac
h
i
e
v
e
d
a
n
8
8
%
ac
cu
r
ac
y
a
n
d
a
n
AUC
o
f
0
.
8
8
,
d
em
o
n
s
t
r
ati
n
g
its
p
o
te
n
t
ial
.
H
o
we
v
er
,
t
h
e
s
y
s
tem
'
s
p
er
f
o
r
m
a
n
ce
is
b
ase
d
o
n
ex
is
ti
n
g
d
ata
s
h
o
u
ld
n
o
t
g
en
er
ali
ze
o
r
d
i
v
e
r
s
e
d
at
ase
ts
.
F
u
r
t
h
e
r
v
a
li
d
at
io
n
a
n
d
v
al
id
ati
o
n
ar
e
n
e
ed
ed
t
o
e
n
s
u
r
e
i
ts
e
f
f
ec
ti
v
e
n
ess
i
n
c
li
n
ic
al
s
etti
n
g
s
an
d
p
atie
n
t
o
u
tc
o
m
es
.
T
h
e
s
y
s
te
m
'
s
a
c
ce
s
s
i
b
il
it
y
in
r
es
o
u
r
ce
-
c
o
n
s
t
r
ain
e
d
h
ea
l
th
ca
r
e
s
e
tti
n
g
s
m
ay
b
e
l
im
i
te
d
b
y
co
m
p
u
tati
o
n
a
l
r
es
o
u
r
ce
s
a
n
d
t
im
e
.
C
o
n
ti
n
u
al
u
p
d
ates
a
n
d
m
ain
te
n
a
n
c
e
a
r
e
n
ec
ess
a
r
y
t
o
e
n
s
u
r
e
ac
c
u
r
a
cy
a
n
d
r
el
e
v
a
n
c
e.
E
t
h
ica
l
co
n
s
i
d
er
ati
o
n
s
r
e
g
ar
d
i
n
g
p
atie
n
t
p
r
iv
ac
y
a
n
d
c
o
n
s
en
t
an
d
s
y
s
te
m
in
te
g
r
ati
o
n
ar
e
als
o
cr
u
c
ial
.
T
h
an
k
s
to
r
ec
en
t
d
e
v
elo
p
m
e
n
ts
in
m
ed
ical
tech
n
o
lo
g
y
,
a
m
o
r
e
r
ap
id
a
n
d
ac
c
u
r
ate
m
eth
o
d
o
f
d
iag
n
o
s
in
g
b
r
ea
s
t
ca
n
ce
r
is
n
o
w
r
e
q
u
ir
ed
[
2
0
]
.
R
esear
ch
er
s
d
ev
elo
p
ed
a
co
m
p
u
ter
-
m
o
n
ito
r
ed
d
iag
n
o
s
tic
s
y
s
tem
th
at
u
s
e
s
h
is
to
lo
g
y
p
ictu
r
es
v
ia
th
e
u
s
e
o
f
ML
an
d
i
m
ag
e
p
r
o
ce
s
s
in
g
.
T
h
e
s
tu
d
y
f
o
cu
s
ed
o
n
in
v
asiv
e
d
u
ctal
ca
r
cin
o
m
a,
u
s
in
g
C
NN
s
lik
e
VGG1
9
.
T
h
e
r
esu
lts
s
h
o
wed
a
s
ig
n
if
ican
t
im
p
r
o
v
em
en
t
in
F
1
-
s
co
r
e
an
d
ac
cu
r
ac
y
,
with
th
e
Den
s
eNe
t
m
o
d
el
ac
h
iev
in
g
an
ac
cu
r
ac
y
o
f
8
6
.
9
7
%.
Ho
wev
er
,
th
e
s
tu
d
y
'
s
g
en
er
aliza
b
ilit
y
to
o
th
er
ty
p
es
o
f
b
r
ea
s
t
ca
n
ce
r
an
d
th
e
r
elian
ce
o
n
p
r
e
-
tr
ai
n
ed
C
NN
ar
ch
itectu
r
es
m
ay
li
m
it
its
ad
ap
tab
ilit
y
.
Fu
r
th
er
v
alid
atio
n
an
d
test
in
g
in
clin
ical
tr
ials
ar
e
n
ee
d
ed
to
ass
es
s
th
e
s
y
s
tem
's
r
ea
l
-
wo
r
ld
ap
p
licab
ilit
y
an
d
r
eliab
ilit
y
.
E
th
ical
co
n
s
id
er
at
io
n
s
s
u
ch
as
p
atien
t
p
r
iv
ac
y
,
d
ata
s
ec
u
r
ity
,
a
n
d
in
teg
r
atio
n
with
ex
is
tin
g
h
ea
lth
ca
r
e
s
y
s
tem
s
n
ee
d
to
b
e
ad
d
r
ess
ed
b
ef
o
r
e
p
r
a
ctica
l im
p
lem
en
tatio
n
.
Usi
n
g
th
er
m
o
g
r
ap
h
y
an
d
d
ee
p
lear
n
in
g
m
o
d
els
f
o
r
ca
teg
o
r
iz
atio
n
,
E
m
am
et
a
l.
[
2
1
]
talk
s
ab
o
u
t
h
o
w
im
p
o
r
tan
t
it
is
to
id
en
tify
b
r
ea
s
t
ca
n
ce
r
ea
r
ly
.
W
ith
a
clas
s
if
icatio
n
ac
cu
r
ac
y
o
f
9
9
.
9
7
%
o
n
th
e
test
s
et,
it
p
r
esen
ts
th
e
im
p
r
o
v
e
d
Den
s
e
Net
m
o
d
el
(
L
év
y
f
lig
h
t
an
d
r
an
d
o
m
o
p
p
o
s
itio
n
-
b
ased
lear
n
in
g
-
im
p
r
o
v
ed
co
ati
o
p
tim
izatio
n
alg
o
r
ith
m
-
Den
s
eNe
t1
2
1
-
b
r
ea
s
t
ca
n
ce
r
(
L
FR
-
C
O
A
-
Den
s
eNe
t1
2
1
-
BC
)
)
.
T
h
e
m
o
d
el'
s
ef
f
ec
tiv
en
ess
was
ev
alu
ated
in
r
ea
l
-
wo
r
ld
m
ed
ical
s
ce
n
a
r
io
s
,
s
h
o
wca
s
in
g
its
ef
f
icac
y
in
b
r
ea
s
t
ca
n
ce
r
class
if
icatio
n
.
Ho
wev
er
,
th
e
a
r
ticle
lack
s
d
etailed
i
n
s
ig
h
t
i
n
to
th
e
s
p
ec
if
ic
ch
alle
n
g
es
f
a
ce
d
b
y
C
NNs
with
h
y
p
er
p
ar
am
eter
s
,
t
h
e
n
atu
r
e
o
f
ab
n
o
r
m
alities
d
etec
ted
b
y
th
er
m
o
g
r
ap
h
y
,
an
d
th
e
co
r
r
elatio
n
with
b
r
ea
s
t
ca
n
ce
r
.
T
h
e
co
m
p
ar
is
o
n
with
estab
lis
h
ed
m
o
d
els
an
d
alg
o
r
ith
m
s
co
u
ld
b
e
n
ef
it
f
r
o
m
a
m
o
r
e
in
-
d
e
p
th
a
n
aly
s
is
,
in
clu
d
in
g
p
o
ten
tial
b
iases
o
r
l
im
itatio
n
s
.
Fu
tu
r
e
s
tu
d
ies
s
h
o
u
ld
ad
d
r
ess
th
e
in
ter
p
r
eta
b
ilit
y
o
f
t
h
e
Den
s
eNe
t
m
o
d
el
an
d
th
e
p
r
ac
tical
f
ea
s
ib
ilit
y
o
f
im
p
lem
en
ti
n
g
th
e
L
FR
-
C
OA
alg
o
r
ith
m
in
clin
ical
s
ettin
g
s
.
An
in
tellig
en
t
m
eth
o
d
f
o
r
an
aly
zin
g
b
r
ea
s
t
ca
n
ce
r
im
ag
es
h
as
b
ee
n
cr
ea
ted
[
2
2
]
em
p
l
o
y
in
g
ML
m
o
d
els
f
o
r
tr
a
n
s
f
er
lea
r
n
in
g
a
n
d
e
n
s
em
b
le
s
tack
in
g
.
T
h
e
s
y
s
tem
in
co
r
p
o
r
ates
tr
an
s
f
er
lear
n
in
g
m
o
d
els
lik
e
as
I
n
ce
p
tio
n
V3
,
VGG1
9
,
an
d
V
GG1
6
,
as
well
as
en
s
em
b
le
m
u
lti
-
lay
er
p
e
r
ce
p
tr
o
n
(
MLP
)
a
n
d
SVM
m
o
d
els,
f
o
r
th
e
p
u
r
p
o
s
e
o
f
ev
alu
atin
g
u
ltra
s
o
u
n
d
b
r
ea
s
t
ca
n
ce
r
p
ictu
r
es.
W
ith
ac
cu
r
ac
y
v
alu
e
o
f
0
.
8
5
8
an
d
an
AUC
o
f
0
.
9
4
7
,
th
e
s
u
g
g
ested
tech
n
iq
u
e
(
I
n
ce
p
tio
n
V
3
+Stac
k
in
g
)
s
u
r
p
ass
es
cu
r
r
en
t
b
r
ea
s
t
ca
n
ce
r
d
iag
n
o
s
tic
m
eth
o
d
s
.
Data
g
ath
er
in
g
,
p
r
e
-
p
r
o
ce
s
s
in
g
,
tr
an
s
f
er
lear
n
in
g
,
ML
m
o
d
e
l
en
s
em
b
le
s
tack
in
g
,
an
d
p
e
r
f
o
r
m
an
ce
ass
ess
m
en
t
ar
e
all
p
ar
t
o
f
th
e
s
y
s
tem
.
T
h
e
r
esu
lts
m
ay
n
ee
d
m
o
r
e
test
in
g
an
d
co
n
f
ir
m
atio
n
o
n
b
ig
g
er
d
atasets
to
r
ep
r
esen
t
clin
ical
u
s
e
in
th
e
r
ea
l
wo
r
l
d
.
I
m
p
o
r
ta
n
t
f
o
r
b
u
ild
in
g
co
n
f
i
d
en
ce
an
d
ac
ce
p
tab
ilit
y
i
n
h
ea
l
th
ca
r
e
s
ettin
g
s
,
th
e
r
esear
ch
d
o
es
n
o
t
ad
d
r
ess
th
e
q
u
esti
o
n
o
f
A
I
/ML
m
o
d
els'
in
ter
p
r
etab
ilit
y
.
T
h
e
f
in
d
i
n
g
s
a
r
e
n
o
t
s
p
ec
if
ically
ad
d
r
ess
ed
in
ter
m
s
o
f
t
h
eir
tr
an
s
f
er
ab
ilit
y
to
o
th
er
h
ea
lth
c
ar
e
co
n
tex
ts
o
r
g
eo
g
r
a
p
h
ical
ar
ea
s
.
R
esear
ch
o
n
h
ea
lth
ca
r
e
AI
an
d
ML
s
y
s
tem
s
ca
n
b
en
e
f
it f
r
o
m
ad
d
r
ess
in
g
p
o
s
s
ib
le
p
r
iv
ac
y
,
s
ec
u
r
ity
,
an
d
eth
ical
is
s
u
es.
B
o
u
d
o
u
h
an
d
B
o
u
ak
k
az
[
2
3
]
d
ev
elo
p
e
d
a
m
o
d
el
f
o
r
d
etec
ti
n
g
b
r
ea
s
t
tu
m
o
r
s
b
y
c
o
m
b
in
i
n
g
d
ata
f
r
o
m
th
r
ee
d
if
f
er
en
t
d
atab
ases
u
s
in
g
p
r
e
-
p
r
o
ce
s
s
in
g
f
ilter
s
,
tr
an
s
f
e
r
lear
n
in
g
,
d
ata
au
g
m
en
tatio
n
,
an
d
g
lo
b
al
p
o
o
lin
g
m
eth
o
d
s
.
Af
ter
u
n
d
er
g
o
in
g
te
s
tin
g
an
d
m
o
d
if
icatio
n
,
two
o
f
th
e
s
ev
en
p
r
e
-
tr
ain
ed
C
NN
s
—
R
es
Net5
0
V2
an
d
I
n
ce
p
tio
n
V3
—
ac
h
ie
v
ed
th
e
b
est
ac
cu
r
ac
y
r
ates.
B
r
ea
s
t
tu
m
o
r
i
d
en
tific
atio
n
was
s
u
cc
ess
f
u
lly
ac
co
m
p
lis
h
ed
u
s
in
g
th
e
s
tr
ateg
y
,
wh
ic
h
b
e
g
an
with
f
ilter
s
elec
tio
n
an
d
co
n
tin
u
ed
with
d
atab
ase
g
a
th
er
in
g
an
d
m
o
d
el
f
in
e
-
tu
n
in
g
.
T
h
e
r
esear
ch
als
o
u
s
ed
th
e
d
ataset
to
f
in
d
a
p
p
r
o
p
r
iate
h
y
p
e
r
-
p
ar
a
m
eter
s
f
o
r
ea
ch
m
o
d
el.
T
h
e
s
tu
d
y
'
s
lim
itatio
n
s
in
clu
d
e
an
i
n
ab
ilit
y
to
g
en
er
alize
th
e
r
esu
lts
d
u
e
to
th
e
lack
o
f
atten
tio
n
g
iv
en
to
th
e
u
n
iq
u
e
d
if
f
icu
lties
o
f
ea
ch
d
atab
ase.
T
h
e
r
esear
ch
d
o
esn
'
t
p
r
o
v
id
e
lig
h
t
o
n
th
e
co
m
p
u
ti
n
g
r
eso
u
r
ce
s
n
ee
d
ed
f
o
r
r
ea
l
-
wo
r
ld
clin
ical
s
itu
atio
n
s
,
an
d
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
m
ay
ch
an
g
e
d
ep
e
n
d
in
g
o
n
th
e
v
ar
iety
an
d
q
u
alit
y
o
f
th
e
d
ata
ac
q
u
ir
e
d
.
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
B
r
ea
s
t c
a
n
ce
r
d
etec
tio
n
u
s
in
g
r
esid
u
a
l D
en
s
eNets
in
d
ee
p
lea
r
n
in
g
(
N
a
g
a
n
a
n
d
in
i G
u
r
u
r
a
j
a
r
a
o
)
1637
A
n
u
m
b
e
r
o
f
r
ec
e
n
t
r
esear
ch
[
6
]
h
a
v
e
s
h
o
wn
th
at
C
AD
s
y
s
tem
s
b
u
ilt
u
s
in
g
d
ee
p
lear
n
in
g
tech
n
iq
u
es
b
ased
o
n
tr
an
s
f
er
lear
n
in
g
ar
e
ef
f
ec
tiv
e
in
d
etec
tin
g
a
n
d
an
al
y
zin
g
d
is
ea
s
es
at
an
ea
r
ly
s
tag
e.
I
n
o
r
d
e
r
to
s
av
e
tim
e,
d
ee
p
lear
n
i
n
g
-
b
ased
co
m
p
u
ter
v
is
io
n
j
o
b
s
g
en
e
r
ally
m
ak
e
u
s
e
o
f
p
r
e
-
tr
ai
n
ed
m
o
d
e
ls
.
W
ith
an
ac
cu
r
ac
y
o
f
8
4
.
0
7
%,
Xce
p
tio
n
o
u
tp
er
f
o
r
m
ed
s
ix
o
th
er
tr
a
n
s
f
er
lear
n
in
g
m
o
d
els
u
s
ed
to
class
if
y
tu
m
o
r
s
in
r
esear
ch
u
s
in
g
th
e
B
r
ea
k
His
d
ataset.
B
alan
ce
d
a
cc
u
r
ac
y
(
B
AC
)
,
a
n
e
w
m
etr
ic
in
tr
o
d
u
ce
d
b
y
Dar
k
Net5
3
,
ac
h
ie
v
ed
th
e
m
ax
im
u
m
ac
cu
r
ac
y
o
f
8
7
.
1
7
%.
T
h
e
s
tu
d
y
'
s
o
v
er
ar
ch
i
n
g
g
o
al
is
to
p
r
o
v
i
d
e
d
o
cto
r
s
with
m
o
r
e
ac
c
u
r
ate
illn
ess
class
if
icatio
n
to
o
ls
.
B
u
t
it
d
o
esn
'
t
lo
o
k
at
th
in
g
s
lik
e
p
o
s
s
ib
le
b
iases
,
lim
its
,
g
en
er
aliza
b
ilit
y
,
eth
ical
is
s
u
es,
in
ter
p
r
etab
ilit
y
,
s
ca
lab
ilit
y
,
o
r
ac
tu
al
th
er
ap
e
u
tic
u
s
e.
T
h
e
m
o
d
els'
ef
f
icac
y
m
ay
n
o
t
r
ef
lect
th
eir
ac
tu
al
p
er
f
o
r
m
an
ce
in
t
h
e
r
ea
l w
o
r
l
d
if
th
ey
d
o
n
o
t
u
s
e
au
g
m
e
n
tatio
n
an
d
p
r
ep
r
o
ce
s
s
in
g
ap
p
r
o
ac
h
es.
E
ar
ly
b
r
ea
s
t
ca
n
ce
r
id
en
tific
atio
n
f
r
o
m
u
ltra
s
o
u
n
d
p
ictu
r
es
m
ay
b
e
ac
h
iev
ed
with
th
e
u
s
e
o
f
tr
an
s
f
er
lear
n
in
g
m
o
d
els
lik
e
Mo
b
ile
NetV2
,
R
es
Net5
0
,
an
d
VGG1
6
wh
en
p
air
ed
with
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
[
2
4
]
.
T
h
e
ef
f
icien
cy
o
f
th
e
s
u
g
g
ested
tech
n
i
q
u
e
u
s
in
g
VGG1
6
was
s
h
o
wn
b
y
its
h
ig
h
Ma
tth
ews
co
r
r
elatio
n
co
ef
f
icien
t
(
MCC
)
,
Kap
p
a
c
o
ef
f
icien
t,
an
d
AUC
,
as
well
a
s
it
s
r
em
ar
k
ab
le
F1
-
s
co
r
e
o
f
9
9
.
0
%.
A
s
tr
o
n
g
p
er
f
o
r
m
a
n
ce
o
v
er
n
u
m
er
o
u
s
v
alid
atio
n
s
ets
was
s
h
o
wn
b
y
a
n
av
er
a
g
e
F1
-
s
co
r
e
o
f
9
6
%,
wh
ich
was
ac
h
iev
ed
b
y
cr
o
s
s
-
v
alid
atio
n
u
s
in
g
th
e
k
-
f
o
l
d
ap
p
r
o
ac
h
.
T
o
im
p
r
o
v
e
th
e
v
is
ib
ilit
y
o
f
th
e
m
o
d
el'
s
d
ec
is
io
n
-
m
ak
in
g
p
r
o
ce
s
s
,
v
is
u
aliza
tio
n
to
o
ls
s
u
ch
as
g
r
a
d
ien
t
-
weig
h
ted
class
ac
tiv
atio
n
m
ap
p
i
n
g
(
Gr
ad
-
C
AM
)
an
d
in
ter
p
r
etab
ilit
y
to
o
ls
lik
e
lo
c
a
l
in
ter
p
r
etab
le
m
o
d
el
-
ag
n
o
s
tic
ex
p
lan
atio
n
s
(
LIME
)
wer
e
u
s
ed
.
B
o
o
ts
tr
ap
p
in
g
an
d
th
e
n
o
r
m
al
a
p
p
r
o
x
im
atio
n
in
ter
v
al
co
n
f
id
en
ce
in
ter
v
al
s
p
r
o
v
e
d
th
at
th
e
p
r
o
ce
d
u
r
e
was
co
n
s
is
ten
t
an
d
r
eliab
le
wh
en
esti
m
atin
g
p
er
f
o
r
m
an
ce
.
T
h
e
m
eth
o
d
'
s
ap
p
licab
ilit
y
to
o
th
er
d
atasets
o
r
clin
ical
co
n
tex
ts
m
ay
b
e
lim
ited
,
n
ev
e
r
th
eless
,
d
u
e
t
o
th
e
s
tu
d
y
'
s
p
o
s
s
ib
le
ab
s
en
ce
o
f
e
x
ter
n
al
v
alid
atio
n
.
T
h
e
cli
n
ical
r
elev
an
ce
an
d
ef
f
ec
t o
n
p
atien
t o
u
tc
o
m
es n
ee
d
m
o
r
e
r
esear
ch
.
Usi
n
g
th
r
ee
DC
NN
ar
ch
itect
u
r
es
—
VGG
-
1
6
,
Xce
p
tio
n
,
an
d
Den
s
e
N
et2
0
1
—
a
p
r
o
p
o
s
ed
AI
s
y
s
tem
b
ased
o
n
tr
an
s
f
er
lear
n
in
g
[
2
5
]
ca
n
id
en
tify
b
r
ea
s
t
ca
n
ce
r
f
r
o
m
h
is
to
p
ath
o
lo
g
y
p
ictu
r
es.
W
ith
a
9
9
.
4
2
%
an
d
9
9
.
1
2
%
ac
c
u
r
ac
y
r
ate,
r
esp
ec
t
iv
ely
,
th
e
s
y
s
tem
o
u
tp
e
r
f
o
r
m
s
s
tate
-
of
-
th
e
-
ar
t
ap
p
r
o
ac
h
es.
S
o
m
e
lim
itatio
n
s
o
f
th
e
s
tu
d
y
in
clu
d
e
th
e
ab
s
en
ce
o
f
r
esear
ch
in
to
th
e
s
y
s
tem
'
s
u
s
e
in
ac
tu
al
clin
ical
s
ettin
g
s
,
co
n
s
id
er
atio
n
o
f
r
eg
u
lato
r
y
an
d
eth
ical
co
n
s
id
e
r
atio
n
s
,
an
d
f
u
r
t
h
er
v
alid
atio
n
o
n
v
ar
ied
d
atasets
.
New
d
ev
elo
p
m
en
ts
in
AI
an
d
p
ictu
r
e
an
aly
s
is
m
ay
b
e
b
e
y
o
n
d
th
e
s
y
s
tem
'
s
f
lex
ib
ilit
y
s
in
ce
it
u
s
es
p
r
e
-
tr
ain
ed
b
ase
m
o
d
els.
I
t
is
also
im
p
o
r
tan
t
to
r
ec
o
g
n
ize
an
d
r
eso
lv
e
an
y
is
s
u
es
th
at
m
ay
ar
is
e
with
th
e
AI
-
b
ased
ca
te
g
o
r
izatio
n
s
y
s
tem
's
in
ter
p
r
etab
ilit
y
,
r
e
p
ea
tab
ilit
y
,
an
d
tr
an
s
p
ar
e
n
cy
.
3.
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
ed
wo
r
k
f
o
r
th
is
is
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
u
s
in
g
C
NNs
an
d
r
esid
u
al
d
en
s
e
n
etwo
r
k
s
(
R
esDen
s
eNe
t
s
)
.
I
t
in
clu
d
es
t
h
e
s
tep
s
lik
e
d
ata
co
llect
io
n
,
d
ata
p
r
ep
r
o
ce
s
s
in
g
,
tr
ain
in
g
an
d
b
u
ild
in
g
a
C
NN
an
d
R
esDen
s
eNe
ts
m
o
d
els
,
a
n
d
f
in
ally
p
e
r
f
o
r
m
an
ce
ev
al
u
atio
n
.
T
h
e
o
v
er
all
s
y
s
tem
ar
ch
itectu
r
e
is
s
h
o
wn
in
Fig
u
r
e
1
.
D
a
t
a
C
o
l
l
e
c
t
i
o
n
T
r
a
i
n
u
s
i
n
g
C
N
N
B
u
i
l
d
M
o
d
e
l
s
T
r
a
i
n
u
s
i
n
g
R
e
s
D
e
n
s
e
N
e
t
s
D
a
t
a
P
r
e
p
r
o
c
e
s
s
i
n
g
E
v
a
l
u
a
t
e
P
e
r
f
o
r
m
a
n
c
e
Fig
u
r
e
1
.
Sy
s
tem
ar
c
h
itectu
r
e
3
.
1
.
Da
t
a
c
o
llect
io
n
T
h
e
d
atasets
u
s
ed
in
th
is
s
tu
d
y
in
clu
d
e
p
u
b
licly
av
ailab
le
r
ep
o
s
ito
r
ies
s
u
ch
as
DDSM,
B
r
ea
k
His
,
an
d
MI
AS,
wh
ich
ar
e
wid
ely
ad
o
p
ted
in
b
r
ea
s
t
ca
n
ce
r
r
esear
ch
f
o
r
ev
alu
atin
g
d
ee
p
lear
n
in
g
m
o
d
els
[
1
]
,
[
7
]
,
[
1
7
]
.
T
h
ese
d
atasets
co
n
tain
m
am
m
o
g
r
am
s
,
u
ltra
s
o
u
n
d
im
ag
es,
an
d
h
is
to
p
at
h
o
lo
g
ical
s
lid
es,
en
ab
lin
g
co
m
p
r
eh
e
n
s
iv
e
an
aly
s
is
ac
r
o
s
s
m
u
ltip
le
im
ag
in
g
m
o
d
alities
.
T
h
e
d
ata
co
llectio
n
d
etails ab
o
u
t th
e
d
atasets
an
d
im
ag
e
ty
p
es
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
3
2
-
1
6
4
5
1638
i)
Data
s
ets:
m
ak
e
u
s
e
o
f
o
p
en
-
s
o
u
r
ce
d
atasets
lik
e
DDSM,
B
r
e
ak
His
,
an
d
MI
AS d
ataset.
ii)
I
m
ag
e
ty
p
es:
in
clu
d
e
m
am
m
o
g
r
am
s
,
u
ltra
s
o
u
n
d
im
a
g
es,
an
d
h
is
to
p
ath
o
lo
g
ical
s
lid
e
s
to
en
s
u
r
e
a
co
m
p
r
eh
e
n
s
iv
e
an
aly
s
is
ac
r
o
s
s
d
if
f
er
en
t im
a
g
in
g
m
o
d
alities
.
3
.
2
.
Da
t
a
prepro
ce
s
s
ing
Data
p
r
ep
r
o
ce
s
s
in
g
in
clu
d
es
im
ag
e
r
esizin
g
,
n
o
r
m
aliza
tio
n
,
an
d
au
g
m
en
tatio
n
tech
n
iq
u
es
t
o
im
p
r
o
v
e
m
o
d
el
p
er
f
o
r
m
an
ce
.
I
m
a
g
e
r
e
s
izin
g
en
s
u
r
es
u
n
if
o
r
m
in
p
u
t
d
im
en
s
io
n
s
s
u
itab
le
f
o
r
C
NN
ar
ch
itectu
r
es,
wh
ile
n
o
r
m
aliza
tio
n
s
tan
d
ar
d
izes
p
ix
el
in
ten
s
ity
v
al
u
es
to
ac
ce
le
r
ate
co
n
v
er
g
e
n
ce
d
u
r
in
g
tr
ain
in
g
[
1
]
,
[
1
5
]
.
Dat
a
au
g
m
en
tatio
n
tech
n
iq
u
es
s
u
ch
as
f
lip
p
in
g
,
r
o
tatio
n
,
an
d
zo
o
m
in
g
ar
e
ap
p
lied
to
in
cr
ea
s
e
d
ataset
d
iv
er
s
ity
an
d
r
ed
u
ce
o
v
er
f
itti
n
g
,
wh
ic
h
h
as
b
ee
n
s
h
o
wn
to
s
ig
n
if
ica
n
tly
en
h
an
ce
C
NN
p
er
f
o
r
m
an
ce
in
m
ed
ical
im
ag
e
an
aly
s
is
[
1
3
]
,
[
2
0
]
.
i)
R
esizin
g
:
g
et
all
th
e
im
ag
es
cr
o
p
p
e
d
to
t
h
e
s
am
e
s
ize
with
(
2
0
0
×
6
0
0
p
ix
els)
s
o
th
e
n
e
u
r
al
n
etwo
r
k
'
s
in
p
u
t la
y
er
m
ay
u
s
e
th
e
m
.
ii)
No
r
m
aliza
tio
n
:
im
p
r
o
v
in
g
co
n
v
er
g
e
n
ce
s
p
ee
d
d
u
r
in
g
tr
ain
i
n
g
m
ay
b
e
ac
h
ie
v
ed
b
y
s
tan
d
ar
d
izin
g
p
i
x
el
v
alu
es to
a
r
an
g
e,
s
u
ch
as
0
to
1
.
iii)
Data
au
g
m
en
tatio
n
:
im
p
r
o
v
e
t
h
e
m
o
d
el'
s
r
esil
ien
ce
b
y
u
s
in
g
d
ata
au
g
m
e
n
tatio
n
m
eth
o
d
s
lik
e
f
lip
p
in
g
,
cr
o
p
p
in
g
,
zo
o
m
in
g
,
an
d
r
o
tati
n
g
to
ar
tific
ially
e
x
p
an
d
th
e
tr
ain
in
g
s
et.
3
.
3
.
M
o
del a
rc
hite
ct
ure
(
co
n
v
o
lutio
na
l neura
l net
wo
rk
)
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
e
m
p
lo
y
s
a
C
NN
ar
c
h
itectu
r
e
t
o
au
t
o
m
a
tically
ex
tr
ac
t
m
e
an
in
g
f
u
l f
ea
t
u
r
es
f
r
o
m
b
r
ea
s
t c
an
ce
r
m
ed
ical
im
ag
es.
C
NNs
ar
e
h
ig
h
ly
ef
f
ec
tiv
e
in
m
ed
ical
im
ag
in
g
task
s
d
u
e
to
th
eir
ab
ilit
y
to
lear
n
s
p
atial
h
ier
ar
ch
ies
o
f
f
ea
tu
r
e
s
[
1
]
,
[
3
]
.
C
o
n
v
o
lu
tio
n
al
lay
er
s
ca
p
tu
r
e
lo
w
-
an
d
h
ig
h
-
le
v
el
f
ea
tu
r
es,
wh
ile
ac
tiv
atio
n
f
u
n
ctio
n
s
s
u
ch
as R
eL
U
in
tr
o
d
u
ce
n
o
n
-
lin
ea
r
ity
a
n
d
im
p
r
o
v
e
lear
n
in
g
ca
p
ab
ilit
y
[
7
]
.
Po
o
lin
g
lay
er
s
r
ed
u
ce
s
p
atial
d
im
e
n
s
io
n
s
an
d
co
m
p
u
tatio
n
al
co
m
p
lex
ity
,
an
d
f
u
lly
co
n
n
ec
ted
lay
e
r
s
p
e
r
f
o
r
m
class
if
icatio
n
b
ased
o
n
e
x
tr
ac
ted
f
ea
tu
r
es [
8
]
.
i)
C
o
n
v
o
lu
tio
n
al
lay
e
r
s
:
u
s
e
m
u
ltip
le
co
n
v
o
lu
tio
n
al
lay
er
s
with
in
cr
ea
s
in
g
f
ilter
s
izes
to
lear
n
h
ier
ar
ch
ical
f
ea
tu
r
es.
ii)
Activ
atio
n
f
u
n
ctio
n
s
:
af
ter
e
v
er
y
co
n
v
o
lu
tio
n
al
lay
er
,
ad
d
n
o
n
-
lin
ea
r
ity
b
y
u
s
in
g
r
ec
tifi
ed
lin
ea
r
u
n
it
(
R
eL
U)
ac
tiv
atio
n
f
u
n
ctio
n
s
.
iii)
Po
o
lin
g
lay
er
s
: m
o
s
t c
r
itical
c
h
ar
ac
ter
is
tics
wh
ile
r
ed
u
cin
g
s
p
atial
d
im
en
s
io
n
s
,
u
s
e
m
ax
p
o
o
lin
g
lay
er
s
.
iv
)
Fu
lly
co
n
n
ec
ted
lay
er
s
: f
o
r
ca
teg
o
r
izatio
n
,
a
d
d
f
u
lly
lin
k
e
d
l
ay
er
s
to
th
e
n
etwo
r
k
'
s
en
d
.
v)
Ou
tp
u
t la
y
er
: u
s
e
So
f
tMa
x
lay
er
f
o
r
m
u
lti
-
class
class
if
icatio
n
o
r
s
ig
m
o
i
d
lay
er
f
o
r
b
in
ar
y
c
lass
if
icatio
n
.
R
esDen
s
eNe
ts
ar
ch
itectu
r
e:
i)
R
esid
u
al
b
lo
ck
s
:
im
p
lem
en
t
r
esid
u
al
b
lo
ck
s
with
id
en
tity
s
h
o
r
tcu
t
c
o
n
n
ec
tio
n
s
to
e
n
ab
le
th
e
tr
ain
in
g
o
f
v
er
y
d
ee
p
n
etwo
r
k
s
.
ii)
Dee
p
ar
ch
itectu
r
e:
u
s
e
R
esNet
with
Den
s
N
et2
0
1
m
o
d
els
an
d
f
in
e
-
tu
n
e
th
em
o
n
th
e
b
r
ea
s
t
ca
n
ce
r
d
ataset.
iii)
T
r
an
s
f
er
lear
n
in
g
:
ad
ju
s
t
th
e
last
f
ew
lay
er
s
o
f
th
e
p
r
e
-
tr
ain
ed
R
esNet
wh
ile
m
ain
tain
in
g
th
e
f
ir
s
t
lay
er
s
u
n
ch
an
g
ed
in
o
r
d
e
r
to
m
a
k
e
u
s
e
o
f
th
e
f
ea
tu
r
es th
at
h
av
e
b
e
en
lear
n
t.
3
.
4
.
T
ra
ini
ng
I
n
th
is
s
tag
e,
d
ata
s
p
lit
in
to
t
r
ain
a
n
d
test
.
T
h
is
h
elp
s
to
e
v
alu
ate
th
e
p
er
f
o
r
m
an
ce
u
s
in
g
lo
s
s
an
d
ac
cu
r
ac
y
m
etr
ics
with
ap
p
r
o
p
r
iate
f
u
n
ctio
n
s
.
Per
f
ec
t
o
p
ti
m
izer
s
ar
e
u
s
ed
f
o
r
e
f
f
ic
ien
t
tr
ain
in
g
an
d
ap
p
ly
h
y
p
er
p
ar
am
eter
s
tu
n
in
g
with
l
ea
r
n
in
g
r
ate,
b
atch
s
ize
an
d
e
p
o
ch
co
u
n
t
f
o
r
o
p
tim
al
co
n
f
i
g
u
r
atio
n
.
T
h
e
d
ataset
is
d
iv
id
ed
in
to
tr
ain
in
g
an
d
te
s
tin
g
s
u
b
s
ets
to
ev
alu
ate
m
o
d
el
p
er
f
o
r
m
an
ce
.
Op
tim
izatio
n
tech
n
iq
u
es
s
u
ch
as
ad
ap
tiv
e
lear
n
in
g
r
ates
an
d
b
atch
-
b
ased
tr
ain
in
g
ar
e
u
s
ed
to
im
p
r
o
v
e
co
n
v
er
g
e
n
ce
an
d
s
tab
ilit
y
.
Hy
p
er
p
ar
a
m
eter
tu
n
in
g
,
in
clu
d
in
g
lear
n
in
g
r
ate,
b
atch
s
ize,
an
d
ep
o
ch
s
,
is
ess
en
tial
f
o
r
ac
h
iev
in
g
o
p
tim
al
p
er
f
o
r
m
an
ce
,
as d
e
m
o
n
s
tr
ated
in
p
r
io
r
d
ee
p
lear
n
in
g
s
tu
d
ies o
n
b
r
ea
s
t c
an
ce
r
class
if
icatio
n
[
3
]
,
[
1
3
]
.
3
.
5
.
E
v
a
lua
t
i
o
n
E
v
alu
atio
n
m
etr
ic
a
n
d
co
n
f
u
s
io
n
m
atr
ix
ar
e
th
e
two
e
v
a
lu
atio
n
p
ar
am
eter
s
c
o
n
s
id
er
e
d
f
o
r
t
h
e
p
r
o
p
o
s
ed
m
o
d
el.
E
v
alu
atio
n
m
etr
ic
h
elp
s
to
ass
ess
th
e
m
ajo
r
p
er
f
o
r
m
a
n
ce
lik
e
ac
cu
r
ac
y
an
d
lo
s
s
m
ea
s
u
r
es.
C
o
n
f
u
s
io
n
m
atr
i
x
h
elp
s
t
o
u
n
d
er
s
tan
d
t
h
e
m
o
d
el’
s
p
er
f
o
r
m
an
ce
f
o
r
v
a
r
io
u
s
class
es.
Data
p
r
ep
r
o
ce
s
s
in
g
,
s
tr
o
n
g
m
o
d
el
ar
c
h
itectu
r
e,
a
n
d
ex
h
a
u
s
tiv
e
ass
es
s
m
en
t
ar
e
th
e
th
r
ee
p
illar
s
u
p
o
n
wh
ich
th
is
s
u
g
g
ested
tech
n
iq
u
e
r
ests
,
o
u
tlin
in
g
a
f
u
l
l
-
s
tr
en
g
th
s
tr
ateg
y
f
o
r
b
u
ild
in
g
an
d
im
p
lem
en
tin
g
C
NN
an
d
R
esNet
m
o
d
els
f
o
r
b
r
ea
s
t
ca
n
ce
r
d
iag
n
o
s
is
.
Mo
d
e
l
p
er
f
o
r
m
an
ce
is
ev
al
u
ated
u
s
i
n
g
m
etr
ics
s
u
ch
as
ac
c
u
r
ac
y
,
l
o
s
s
,
an
d
co
n
f
u
s
io
n
m
atr
ix
an
al
y
s
is
.
T
h
ese
ev
al
u
atio
n
m
etr
ics
a
r
e
wid
el
y
u
s
ed
in
m
ed
ical
im
a
g
e
class
if
icatio
n
t
o
ass
ess
d
iag
n
o
s
tic
p
er
f
o
r
m
an
ce
an
d
r
eliab
ilit
y
[
2
]
,
[
8
]
.
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
B
r
ea
s
t c
a
n
ce
r
d
etec
tio
n
u
s
in
g
r
esid
u
a
l D
en
s
eNets
in
d
ee
p
lea
r
n
in
g
(
N
a
g
a
n
a
n
d
in
i G
u
r
u
r
a
j
a
r
a
o
)
1639
4.
I
M
P
E
L
E
M
E
N
T
A
T
I
O
N
4
.
1
.
Wo
r
k
ing
o
f
co
nv
o
lutio
n
a
l neura
l net
wo
rk
wit
h 3
la
y
er
ed
a
rc
hite
ct
ure
s
T
h
e
C
N
N
a
r
c
h
i
t
e
ct
u
r
e
c
o
n
s
is
ts
o
f
m
u
l
t
i
p
l
e
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
er
s
t
h
a
t
e
x
t
r
ac
t
h
i
e
r
a
r
c
h
ic
a
l
f
e
atu
r
e
s
f
r
o
m
i
n
p
u
t
i
m
a
g
e
s
.
C
o
n
v
o
l
u
t
i
o
n
o
p
e
r
a
t
i
o
n
s
h
e
l
p
d
e
te
c
t
p
a
tt
e
r
n
s
s
u
c
h
a
s
e
d
g
e
s
a
n
d
t
e
x
t
u
r
es
,
wh
i
c
h
a
r
e
c
r
u
c
i
a
l
f
o
r
i
d
e
n
t
i
f
y
i
n
g
a
b
n
o
r
m
a
l
i
t
ie
s
i
n
m
e
d
i
c
a
l
i
m
a
g
es
[
1
]
,
[
3
]
.
A
c
t
i
v
a
t
io
n
f
u
n
c
t
i
o
n
s
s
u
c
h
a
s
R
e
L
U
i
n
t
r
o
d
u
c
e
n
o
n
-
l
i
n
e
a
r
i
t
y
,
e
n
a
b
l
i
n
g
t
h
e
n
e
tw
o
r
k
t
o
le
a
r
n
c
o
m
p
l
e
x
r
e
p
r
e
s
e
n
ta
t
i
o
n
s
[
7
]
.
P
o
o
l
i
n
g
l
a
y
e
r
s
r
e
d
u
c
e
s
p
a
ti
a
l
d
i
m
e
n
s
i
o
n
s
a
n
d
c
o
m
p
u
t
a
t
i
o
n
a
l
c
o
s
t
w
h
il
e
p
r
ese
r
v
i
n
g
i
m
p
o
r
t
a
n
t
f
ea
t
u
r
e
s
[
8
]
.
F
u
l
l
y
c
o
n
n
e
ct
e
d
l
a
y
e
r
s
i
n
t
e
g
r
a
te
e
x
t
r
a
c
te
d
f
e
a
t
u
r
es
t
o
p
e
r
f
o
r
m
f
i
n
a
l
c
l
ass
i
f
i
c
at
i
o
n
,
t
y
p
i
c
a
ll
y
u
s
i
n
g
S
o
f
t
M
a
x
f
o
r
m
u
l
t
i
-
c
l
ass
p
r
o
b
l
e
m
s
.
Fi
g
u
r
e
2
i
l
l
u
s
t
r
a
te
s
a
C
NN
w
i
t
h
a
3
-
l
a
y
e
r
e
d
a
r
c
h
i
te
c
t
u
r
e
.
A
n
C
N
N
wi
t
h
t
h
r
ee
c
o
n
v
o
l
u
t
io
n
a
l
l
a
y
e
r
s
m
a
k
e
s
u
p
t
h
e
m
o
d
e
l'
s
a
r
c
h
i
t
e
ct
u
r
e
.
T
h
e
k
e
r
n
e
l
s
i
z
e
f
o
r
e
ac
h
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
is
s
et
t
o
3
,
w
it
h
c
h
a
n
n
e
l
s
iz
e
s
o
f
7
,
5
,
a
n
d
3
,
c
o
r
r
e
s
p
o
n
d
i
n
g
l
y
.
Fig
u
r
e
2
.
C
NN
with
3
lay
er
e
d
a
r
ch
itectu
r
e
s
T
h
e
k
ey
c
o
n
ce
p
ts
o
f
C
NNs a
r
e
as f
o
llo
ws:
i)
C
o
n
v
o
lu
tio
n
al
l
a
y
er
s
:
c
o
n
v
o
lu
tio
n
al
lay
er
s
u
s
e
a
co
llectio
n
o
f
f
ilter
s
(
also
k
n
o
wn
as
k
e
r
n
els)
to
p
r
o
ce
s
s
th
e
in
p
u
t
p
ictu
r
e.
T
h
e
f
ilter
is
ap
p
lied
t
o
th
e
im
a
g
es
b
y
s
lid
in
g
it
s
p
atially
an
d
at
ea
ch
lo
ca
tio
n
,
f
in
d
in
g
th
e
d
o
t
p
r
o
d
u
ct
o
f
th
e
in
p
u
t
v
alu
es
an
d
th
e
f
ilter
elem
en
ts
.
T
h
is
p
r
o
ce
d
u
r
e
g
en
e
r
ates
a
f
e
atu
r
e
m
a
p
th
at
ac
ce
n
tu
ates
ce
r
tain
ch
ar
ac
ter
i
s
tics
o
f
th
e
in
p
u
t,
s
u
ch
as
ed
g
es,
tex
tu
r
es,
o
r
s
p
ec
if
ic
f
o
r
m
s
.
T
h
e
u
s
e
o
f
m
an
y
f
ilter
s
en
ab
les th
e
n
etwo
r
k
to
ac
q
u
ir
e
d
i
v
er
s
e
p
r
o
p
er
tie
s
th
at
ar
e
cr
u
cial
f
o
r
th
e
g
i
v
en
jo
b
.
ii)
Activ
atio
n
f
u
n
ctio
n
s
:
f
o
llo
wi
n
g
th
e
co
n
v
o
lu
tio
n
p
r
o
ce
d
u
r
e,
th
e
m
o
d
el
in
tr
o
d
u
ce
s
n
o
n
-
lin
ea
r
ity
b
y
ap
p
ly
in
g
a
n
ac
tiv
atio
n
f
u
n
cti
o
n
,
u
s
u
ally
th
e
R
eL
U.
T
h
e
p
r
esen
ce
o
f
n
o
n
-
lin
ea
r
ity
is
ess
en
tial
f
o
r
ac
q
u
ir
in
g
a
d
ee
p
u
n
d
er
s
tan
d
i
n
g
o
f
in
tr
icate
p
atter
n
s
with
in
th
e
d
ata.
T
o
h
elp
th
e
n
etwo
r
k
u
n
d
er
s
tan
d
t
h
e
co
n
n
ec
tio
n
s
b
etwe
en
d
if
f
er
e
n
t
f
ea
tu
r
es,
t
h
e
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
m
a
k
es
s
u
r
e
th
at
an
y
n
eg
ativ
e
p
ix
e
l
v
alu
es in
th
e
f
ea
tu
r
e
m
ap
ar
e
s
et
to
ze
r
o
.
iii)
Po
o
lin
g
l
ay
er
s
:
b
y
p
r
o
g
r
ess
iv
ely
r
ed
u
cin
g
th
e
s
p
atial
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
e
m
ap
s
,
th
e
c
o
m
p
u
tatio
n
al
co
s
t
an
d
n
u
m
b
er
o
f
p
ar
a
m
eter
s
in
th
e
n
etwo
r
k
m
a
y
b
e
r
e
d
u
ce
d
v
ia
th
e
u
s
e
o
f
p
o
o
lin
g
lay
er
s
.
Als
o
,
th
is
m
ak
es
it
ea
s
ier
to
g
u
ar
an
tee
t
h
at
th
e
r
e
p
r
esen
tatio
n
w
o
n
'
t
ch
an
g
e
e
v
en
i
f
th
e
i
n
p
u
t
is
s
lig
h
tly
tr
an
s
lated
.
On
e
co
m
m
o
n
m
eth
o
d
in
n
eu
r
al
n
etwo
r
k
s
f
o
r
p
o
o
lin
g
d
ata
is
m
ax
p
o
o
lin
g
,
w
h
ich
in
v
o
l
v
e
s
s
elec
tin
g
th
e
h
ig
h
est
v
alu
e
f
r
o
m
ea
ch
f
ea
tu
r
e
m
ap
r
e
g
io
n
.
An
o
th
er
co
m
m
o
n
m
eth
o
d
is
av
er
a
g
e
p
o
o
lin
g
,
wh
ic
h
in
v
o
lv
es c
alcu
latin
g
t
h
e
av
er
a
g
e
v
alu
e.
iv
)
Fu
lly
co
n
n
ec
ted
lay
er
s
:
t
h
e
n
etwo
r
k
ty
p
ically
in
clu
d
es
o
n
e
o
r
m
o
r
e
d
en
s
e
lay
er
s
—
f
u
lly
co
u
p
led
lay
er
s
—
f
o
llo
win
g
a
s
eq
u
en
ce
o
f
co
n
v
o
lu
tio
n
al
an
d
p
o
o
lin
g
lay
er
s
.
T
h
ese
lay
er
s
f
u
n
ctio
n
o
n
in
p
u
t
d
at
a
th
at
h
as
b
ee
n
f
latten
ed
,
tr
ea
tin
g
th
e
in
p
u
t
as
a
s
in
g
u
lar
v
ec
to
r
.
Fu
lly
co
n
n
ec
ted
lay
er
s
ac
q
u
ir
e
co
m
p
r
eh
e
n
s
iv
e
p
atter
n
s
in
th
e
d
ata
b
y
in
teg
r
atin
g
ch
ar
ac
ter
is
tics
r
etr
iev
ed
b
y
th
e
co
n
v
o
lu
ti
o
n
al
lay
er
s
to
p
r
o
v
id
e
f
in
al
p
r
ed
ictio
n
s
.
v)
Ou
tp
u
t
l
ay
er
:
d
ep
e
n
d
in
g
o
n
t
h
e
task
at
h
an
d
,
th
e
o
u
tp
u
t
la
y
er
is
d
eter
m
in
ed
.
I
t
is
co
m
m
o
n
p
r
ac
tice
to
u
tili
ze
th
e
So
f
tMa
x
ac
tiv
atio
n
f
u
n
ctio
n
to
cr
ea
te
a
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
ac
r
o
s
s
all
o
f
th
e
class
es
in
a
class
if
icatio
n
is
s
u
e.
4
.
2
.
Res
idu
a
l
DenseNet
s
Dee
p
n
eu
r
al
n
etwo
r
k
s
o
f
ten
s
u
f
f
er
f
r
o
m
th
e
v
a
n
is
h
in
g
g
r
a
d
i
en
t
p
r
o
b
lem
,
wh
ich
lim
its
th
eir
ab
ilit
y
to
lear
n
ef
f
ec
tiv
ely
as
d
ep
th
in
cr
ea
s
es.
R
es
id
u
al
lear
n
in
g
ad
d
r
ess
es
th
is
is
s
u
e
b
y
in
tr
o
d
u
ci
n
g
s
h
o
r
tcu
t
co
n
n
ec
tio
n
s
th
at
allo
w
g
r
a
d
ien
ts
to
f
lo
w
d
ir
ec
tly
th
r
o
u
g
h
th
e
n
etwo
r
k
[
5
]
.
Den
s
e
co
n
n
ec
ti
o
n
s
f
u
r
t
h
er
en
h
an
ce
lear
n
in
g
b
y
en
a
b
lin
g
ea
c
h
la
y
er
to
r
ec
ei
v
e
in
p
u
ts
f
r
o
m
all
p
r
ec
ed
in
g
lay
er
s
,
im
p
r
o
v
i
n
g
f
ea
tu
r
e
r
e
u
s
e
an
d
r
ed
u
cin
g
r
ed
u
n
d
an
c
y
[
6
]
.
T
h
e
co
m
b
in
atio
n
o
f
th
ese
two
m
ec
h
an
is
m
s
in
R
esDen
s
eN
et
en
ab
les
ef
f
icien
t
tr
ain
in
g
o
f
v
er
y
d
ee
p
n
etwo
r
k
s
an
d
h
as
b
ee
n
s
h
o
w
n
to
o
u
tp
er
f
o
r
m
tr
a
d
itio
n
al
C
NN
ar
ch
itectu
r
es
in
m
ed
ical
im
ag
e
class
if
icatio
n
task
s
[
7
]
,
[
2
1
]
.
Fig
u
r
e
3
illu
s
tr
ates
a
R
esDen
s
e
Nets
m
o
d
el
ar
ch
itectu
r
e.
F
ig
u
r
e
3
illu
s
tr
ates a
R
e
s
Den
s
eNe
t
s
m
o
d
el
ar
ch
itectu
r
e.
T
h
e
k
e
y
co
n
c
ep
ts
o
f
R
esNets
ar
e
as f
o
llo
ws
:
i)
Van
is
h
in
g
-
g
r
a
d
ien
t
-
p
r
o
b
lem
:
i
n
a
d
ee
p
n
eu
r
al
n
etwo
r
k
,
k
n
o
wn
as
a
v
a
n
is
h
in
g
-
g
r
ad
ien
t
-
p
r
o
b
lem
,
th
e
p
r
o
p
a
g
atio
n
o
f
g
r
ad
ie
n
ts
n
ee
d
ed
to
u
p
d
ate
weig
h
ts
s
lo
ws
o
r
s
to
p
s
th
e
lea
r
n
in
g
p
r
o
ce
s
s
.
Dee
p
n
e
u
r
al
n
etwo
r
k
s
tak
e
t
r
ain
d
i
f
f
icu
lt
y
to
tr
ai
n
,
wh
ich
lead
s
in
d
if
f
icu
lt
f
o
r
v
a
n
is
h
in
g
g
r
ad
ien
t
p
r
o
b
le
m
,
as
g
r
ad
ien
ts
n
o
t
p
r
o
p
ag
ate
to
ea
r
lier
lay
er
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
3
2
-
1
6
4
5
1640
ii)
R
esid
u
al
l
ea
r
n
in
g
:
R
esDen
s
eNe
ts
lear
n
in
g
in
s
h
o
r
t
ca
lled
as
R
esDen
s
eNe
t
s
.
R
esDen
s
e
Nets
f
o
cu
s
o
n
lear
n
in
g
th
e
r
esid
u
al
f
u
n
ctio
n
in
s
tead
o
f
lear
n
in
g
a
d
ir
ec
t
m
ap
f
r
o
m
in
p
u
t
to
o
u
tp
u
t
an
d
th
is
tells
th
e
d
if
f
er
en
ce
b
etwe
en
in
p
u
t
a
n
d
d
esire
d
o
u
tp
u
t.
T
h
e
d
esire
d
m
ap
p
in
g
in
m
at
h
em
atica
lly
d
en
o
te
as
(
)
as
s
h
o
wn
in
(
1
)
a
n
d
r
ewr
ite
th
e
m
as
(
2
)
f
o
r
b
etter
u
n
d
er
s
tan
d
in
g
.
I
n
R
esDen
s
eNe
ts
,
th
e
s
t
ac
k
lay
er
s
f
it
a
r
esid
u
al
m
ap
p
in
g
as
(
1
)
.
(
)
=
(
)
−
(
1
)
W
h
ich
we
ca
n
also
r
ewr
ite
as
(
2
)
.
(
)
=
(
)
+
(
2
)
W
h
er
e
x
is
th
e
in
p
u
ts
,
(
)
is
d
esire
d
o
u
tp
u
t,
(
)
is
f
itti
n
g
th
e
s
tack
lay
er
s
.
iii)
R
esid
u
al
b
lo
ck
s
:
th
e
b
ac
k
b
o
n
e
o
f
R
esNet
ar
e
a
r
esid
u
al
b
l
o
ck
.
T
h
e
co
n
v
o
lu
tio
n
al
la
y
er
s
in
ea
ch
b
lo
c
k
ar
e
ac
tiv
ated
u
s
in
g
R
eL
U
an
d
b
atch
n
o
r
m
aliza
tio
n
.
Af
ter
th
at,
y
o
u
'
ll
s
ee
a
s
im
ilar
s
h
o
r
tcu
t
co
n
n
ec
tio
n
th
at
jo
in
s
th
e
b
lo
ck
'
s
in
p
u
t
a
n
d
o
u
t
p
u
t.
De
p
en
d
in
g
o
n
wh
et
h
er
th
e
in
p
u
t
a
n
d
o
u
tp
u
t
d
im
e
n
s
io
n
s
ar
e
s
am
e
o
r
n
o
t,
th
e
s
h
o
r
tcu
t
co
n
n
ec
tio
n
m
ay
tak
e
th
e
f
o
r
m
o
f
an
id
en
tity
m
ap
p
in
g
o
r
a
lin
ea
r
p
r
o
jectio
n
u
s
in
g
1
×
1
co
n
v
o
lu
tio
n
s
.
iv
)
I
d
en
tity
s
h
o
r
tcu
t
co
n
n
ec
tio
n
s
:
th
e
id
en
tity
s
h
o
r
tcu
t
c
o
n
n
ec
ti
o
n
s
p
r
o
v
id
e
th
e
d
ir
ec
t
p
ass
ag
e
o
f
g
r
ad
ien
ts
ac
r
o
s
s
th
e
n
etwo
r
k
,
s
k
ip
p
in
g
o
n
e
o
r
m
o
r
e
le
v
els.
B
ec
au
s
e
o
f
th
is
,
th
e
v
an
is
h
in
g
-
g
r
a
d
ien
t
-
p
r
o
b
lem
is
n
o
lo
n
g
er
a
c
o
n
ce
r
n
,
an
d
f
ar
d
ee
p
er
n
etwo
r
k
s
m
ay
b
e
co
n
s
tr
u
cte
d
.
T
h
ese
co
n
n
ec
tio
n
s
g
u
a
r
an
te
e
th
at
ev
en
if
s
o
m
e
lev
els d
eter
io
r
ate,
th
e
n
e
two
r
k
m
ay
n
ev
er
t
h
eless
f
u
n
ctio
n
ad
e
q
u
ately
b
y
u
s
in
g
i
d
en
tity
m
ap
p
i
n
g
s
.
v)
R
esid
u
al
co
n
n
ec
tio
n
:
m
ak
e
s
tead
y
g
r
ad
ie
n
t
f
lo
w
p
o
s
s
ib
le,
allo
win
g
f
o
r
th
e
tr
ai
n
in
g
o
f
v
er
y
d
ee
p
n
etwo
r
k
s
; th
is
will m
itig
ate
th
e
v
an
is
h
in
g
g
r
ad
ien
t iss
u
e.
v
i)
Den
s
e
co
n
n
ec
tio
n
:
m
a
k
e
it
ea
s
ier
to
r
eu
s
e
f
ea
tu
r
es
b
y
en
h
an
cin
g
in
f
o
r
m
atio
n
f
lo
w
an
d
g
r
a
d
ien
t
p
r
o
p
a
g
atio
n
v
ia
f
ee
d
-
f
o
r
war
d
co
n
n
ec
tio
n
s
b
etwe
en
all
o
f
th
e
lay
er
s
.
v
ii)
Hy
b
r
id
ar
c
h
itectu
r
e:
co
m
b
in
es
th
e
ad
v
an
ta
g
es
o
f
r
esid
u
al
an
d
d
e
n
s
e
co
n
n
ec
tio
n
s
t
o
lear
n
m
o
r
e
co
m
p
r
eh
e
n
s
iv
e
an
d
d
is
cr
im
in
a
tiv
e
f
ea
tu
r
es f
r
o
m
b
r
ea
s
t c
an
ce
r
im
ag
es.
Fig
u
r
e
3
.
R
esDen
s
eNe
ts
m
o
d
el
ar
ch
itectu
r
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
B
r
ea
s
t c
a
n
ce
r
d
etec
tio
n
u
s
in
g
r
esid
u
a
l D
en
s
eNets
in
d
ee
p
lea
r
n
in
g
(
N
a
g
a
n
a
n
d
in
i G
u
r
u
r
a
j
a
r
a
o
)
1641
4
.
3
.
Alg
o
rit
hm
:
ResDens
eNe
t
a
lg
o
rit
h
m
T
h
e
p
r
o
p
o
s
ed
R
esDen
s
eNe
t
alg
o
r
ith
m
in
te
g
r
ates
th
e
ad
v
an
tag
es
o
f
r
esid
u
al
lear
n
in
g
an
d
d
e
n
s
e
co
n
n
ec
tiv
ity
to
im
p
r
o
v
e
f
ea
t
u
r
e
p
r
o
p
a
g
atio
n
an
d
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
in
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
.
R
esid
u
al
co
n
n
ec
tio
n
s
h
elp
o
v
er
co
m
e
t
h
e
v
an
is
h
in
g
g
r
a
d
ien
t
p
r
o
b
lem
in
d
ee
p
n
et
wo
r
k
s
,
wh
ile
d
en
s
e
co
n
n
ec
tio
n
s
en
h
a
n
ce
f
ea
tu
r
e
r
eu
s
e
b
y
allo
win
g
ea
ch
lay
er
to
r
ec
eiv
e
in
p
u
ts
f
r
o
m
all
p
r
ec
ed
in
g
lay
er
s
.
T
h
is
h
y
b
r
id
ar
c
h
itectu
r
e
e
n
ab
les
th
e
n
etwo
r
k
to
lea
r
n
c
o
m
p
lex
p
atter
n
s
f
r
o
m
m
ed
ic
al
im
ag
es
s
u
ch
as
m
am
m
o
g
r
a
m
s
,
u
ltra
s
o
u
n
d
s
ca
n
s
,
an
d
h
is
to
p
ath
o
lo
g
ical
im
ag
es
m
o
r
e
ef
f
ec
tiv
ely
.
T
h
e
s
tep
-
by
-
s
tep
wo
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
s
u
m
m
ar
ized
in
Alg
o
r
ith
m
1
,
wh
ic
h
d
escr
ib
es
t
h
e
m
aj
o
r
s
tag
es
in
v
o
lv
e
d
in
f
ea
tu
r
e
ex
tr
ac
tio
n
,
d
en
s
e
b
lo
ck
p
r
o
ce
s
s
in
g
,
r
esid
u
al
co
n
n
ec
tio
n
s
,
an
d
f
in
al
class
if
icatio
n
.
Alg
o
r
ith
m
1
p
r
esen
ts
th
e
d
etailed
p
r
o
ce
d
u
r
e
f
o
llo
wed
b
y
th
e
p
r
o
p
o
s
ed
R
esDen
s
eNe
t
m
o
d
el
f
o
r
b
r
ea
s
t
ca
n
ce
r
im
ag
e
class
if
icatio
n
.
T
h
e
in
teg
r
atio
n
o
f
r
esid
u
al
an
d
d
en
s
e
c
o
n
n
ec
tio
n
s
in
th
is
alg
o
r
ith
m
im
p
r
o
v
es
f
ea
t
u
r
e
p
r
o
p
ag
atio
n
an
d
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
,
as
s
u
p
p
o
r
ted
b
y
r
ec
e
n
t
s
tu
d
ies
o
n
h
y
b
r
id
d
ee
p
lear
n
i
n
g
a
r
ch
itectu
r
es
f
o
r
b
r
ea
s
t
ca
n
ce
r
d
etec
tio
n
[
7
]
,
[
2
3
]
.
Alg
o
r
ith
m
1
.
R
esid
u
al
Den
s
eNe
t
Step
1
:
I
n
p
u
t la
y
e
r
:
I
n
p
u
t: X
(
a
b
atch
o
f
im
ag
es,
e.
g
.
,
s
ize
2
0
0
×
6
0
0
×
3
f
o
r
R
GB
im
ag
es)
Step
2
:
I
n
itial c
o
n
v
o
l
u
tio
n
:
Ap
p
ly
a
co
n
v
o
l
u
tio
n
with
a
f
ilter
s
ize
o
f
7
×7
,
s
tr
id
e
2
,
an
d
p
ad
d
in
g
,
f
o
llo
wed
b
y
b
atch
n
o
r
m
aliza
tio
n
an
d
R
eL
U
ac
tiv
atio
n
.
Ap
p
ly
a
m
a
x
p
o
o
lin
g
o
p
er
atio
n
with
a
f
ilter
s
ize
o
f
3
×3
an
d
s
tr
id
e
2
.
Step
3
:
Den
s
e
b
lo
ck
1
:
Nu
m
b
er
o
f
lay
er
s
: 6
Fo
r
ea
ch
lay
er
1
in
t
h
e
d
en
s
e
b
lo
ck
:
Ap
p
ly
in
g
th
e
b
atch
n
o
r
m
aliza
t
io
n
,
R
eL
U
ac
tiv
atio
n
,
an
d
co
n
v
lay
er
with
1
×
1
co
n
v
o
lu
tio
n
wh
ich
is
a
b
o
ttlen
ec
k
lay
er
.
Ap
p
ly
b
atch
n
o
r
m
aliza
tio
n
,
R
eL
U
ac
tiv
atio
n
,
an
d
a
3
×3
co
n
v
o
lu
tio
n
.
C
o
n
ca
ten
ate
o
u
tp
u
t
o
f
th
is
lay
er
with
th
e
in
p
u
t to
th
e
d
en
s
e
b
lo
ck
to
f
o
r
m
th
e
in
p
u
t to
th
e
n
ex
t la
y
er
.
Step
4
:
T
r
an
s
itio
n
lay
er
1
:
Ap
p
ly
in
g
th
e
b
atch
n
o
r
m
aliza
t
io
n
,
R
eL
U
ac
tiv
atio
n
,
an
d
co
n
v
lay
er
with
1
×
1
co
n
v
o
lu
tio
n
.
Ap
p
ly
in
g
th
e
av
er
a
g
e
p
o
o
lin
g
with
2
×2
lay
er
with
2
s
tr
id
es.
Step
5
:
Den
s
e
b
lo
ck
2
:
Nu
m
b
er
o
f
lay
er
s
: 1
2
R
ep
ea
t th
e
p
r
o
ce
s
s
d
escr
ib
ed
i
n
d
en
s
e
b
lo
c
k
1
.
Step
6
:
T
r
an
s
itio
n
lay
er
2
:
Ap
p
ly
in
g
th
e
b
atch
n
o
r
m
aliza
t
io
n
,
R
eL
U
ac
tiv
atio
n
,
an
d
co
n
v
lay
er
with
1
×
1
co
n
v
o
lu
tio
n
.
Ap
p
ly
in
g
th
e
av
er
a
g
e
p
o
o
lin
g
with
2
×2
lay
er
with
2
s
tr
id
es.
Step
7
:
Den
s
e
b
lo
ck
3
:
Nu
m
b
er
o
f
lay
er
s
: 4
8
R
ep
ea
t th
e
p
r
o
ce
s
s
d
escr
ib
ed
i
n
d
en
s
e
b
lo
c
k
1
.
Step
8
:
T
r
an
s
itio
n
lay
er
3
:
Ap
p
ly
in
g
th
e
b
atch
n
o
r
m
aliza
t
io
n
,
R
eL
U
ac
tiv
atio
n
,
an
d
co
n
v
lay
er
with
1
×
1
co
n
v
o
lu
tio
n
.
Ap
p
ly
in
g
th
e
av
er
a
g
e
p
o
o
lin
g
with
2
×2
lay
er
with
2
s
tr
id
es.
Step
9
:
Den
s
e
b
lo
ck
4
:
Nu
m
b
er
o
f
lay
er
s
: 3
2
R
ep
ea
t th
e
p
r
o
ce
s
s
d
escr
ib
ed
i
n
d
en
s
e
b
lo
c
k
1
.
Step
1
0
:
R
esid
u
al
co
n
n
ec
tio
n
:
L
in
k
th
e
d
e
n
s
e
b
lo
ck
s
'
in
p
u
ts
to
th
e
tr
an
s
itio
n
lay
e
r
'
s
o
u
tp
u
ts
v
ia
a
s
h
o
r
tcu
t lin
k
.
Step
1
1
:
Fin
al
lay
er
s
:
Ap
p
ly
a
b
atc
h
n
o
r
m
aliza
tio
n
l
ay
er
.
Ap
p
ly
a
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
lay
er
to
r
ed
u
ce
th
e
s
p
atial
d
i
m
en
s
io
n
s
to
1
×1
.
Fu
lly
co
n
n
ec
ted
lay
er
with
So
f
tMa
x
ac
tiv
atio
n
f
o
r
class
if
icatio
n
with
n
u
m
b
er
o
f
class
es 3
.
5.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
ef
f
i
ca
c
y
o
f
t
h
e
p
r
o
p
o
s
ed
R
e
s
D
e
n
s
e
N
e
t
n
e
t
wo
r
k
f
o
r
b
r
ea
s
t
c
an
c
e
r
d
e
te
c
t
i
o
n
w
a
s
e
v
a
lu
a
t
ed
u
s
i
n
g
w
id
e
l
y
a
c
c
ep
t
e
d
p
e
r
f
o
r
m
a
n
c
e
m
e
t
r
i
c
s
s
u
c
h
a
s
a
cc
u
r
a
c
y
an
d
lo
s
s
.
T
h
e
s
e
m
e
t
r
i
c
s
p
r
o
v
i
d
e
i
n
s
i
g
h
t
i
n
to
t
h
e
m
o
d
e
l’
s
a
b
i
l
i
ty
t
o
co
r
r
e
ct
l
y
c
l
a
s
s
i
f
y
b
r
ea
s
t
c
a
n
ce
r
i
m
a
g
e
s
w
h
i
l
e
m
i
n
im
i
z
i
n
g
p
r
e
d
ic
t
i
o
n
er
r
o
r
s
d
u
r
in
g
t
r
a
i
n
in
g
a
n
d
v
a
l
id
a
t
i
o
n
.
T
o
d
e
m
o
n
s
t
r
a
t
e
t
h
e
e
f
f
e
c
t
iv
e
n
e
s
s
o
f
th
e
p
r
o
p
o
s
e
d
ar
c
h
i
t
ec
t
u
r
e
,
t
h
e
p
e
r
f
o
r
m
an
c
e
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.