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
.
3
,
J
u
n
e
2
0
2
6
,
p
p
.
2567
~
2
5
8
0
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
3
.
p
p
2
5
6
7
-
2
5
8
0
2567
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Dia
g
no
sing
t
uber
culo
sis
f
ro
m X
-
ra
y
ima
g
ing
bas
ed on
co
ntras
t
limited ada
ptive
histo
g
ra
m equa
liz
a
tion
Ng
uy
en
T
ro
ng
Vinh
1
,
L
a
m
T
ha
nh
H
ien
1
,
H
a
M
a
nh
T
o
a
n
2
,
Do
Na
ng
T
o
a
n
2
1
F
a
c
u
l
t
y
o
f
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
La
c
H
o
n
g
U
n
i
v
e
r
si
t
y
,
B
i
e
n
H
o
a
,
V
i
e
t
n
a
m
2
I
n
st
i
t
u
t
e
o
f
I
n
f
o
r
ma
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
V
i
e
t
n
a
m
A
c
a
d
e
my
o
f
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
H
a
n
o
i
,
V
i
e
t
n
a
m
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
2
2
,
2
0
2
5
R
ev
is
ed
Feb
1
1
,
2
0
2
6
Acc
ep
ted
Ma
r
5
,
2
0
2
6
Tu
b
e
rc
u
l
o
sis
is
a
se
rio
u
s
t
h
re
a
t,
a
n
d
o
n
e
o
f
t
h
e
e
ffe
c
ti
v
e
d
a
ta
ty
p
e
s
f
o
r
d
iag
n
o
sin
g
t
u
b
e
rc
u
l
o
sis is
c
h
e
st X
-
ra
y
d
a
ta.
In
t
h
is p
a
p
e
r,
we
h
y
p
o
th
e
siz
e
th
e
e
ffe
c
t
o
f
ima
g
e
e
n
h
a
n
c
e
m
e
n
t
o
n
th
e
e
ffe
c
ti
v
e
n
e
ss
o
f
d
e
e
p
lea
rn
in
g
m
o
d
e
ls
i
n
th
e
p
ro
b
lem
o
f
d
ia
g
n
o
sin
g
p
u
lmo
n
a
ry
tu
b
e
rc
u
lo
sis
fr
o
m
c
h
e
st
X
-
ra
y
ima
g
e
s.
To
c
larify
t
h
e
h
y
p
o
th
e
sis,
we
h
a
v
e
d
e
sig
n
e
d
a
d
a
ta
p
ro
c
e
ss
in
g
p
r
o
c
e
ss
with
a
n
ima
g
e
e
n
h
a
n
c
e
m
e
n
t
ste
p
u
si
n
g
th
e
c
o
n
tras
t
li
m
it
e
d
a
d
a
p
ti
v
e
h
isto
g
ra
m
e
q
u
a
li
z
a
ti
o
n
(
CLAHE
)
tec
h
n
iq
u
e
to
e
n
h
a
n
c
e
th
e
q
u
a
li
ty
o
f
i
n
p
u
t
c
h
e
st
X
-
ra
y
d
a
ta,
a
n
d
th
e
e
x
p
e
rime
n
ts
we
re
c
o
n
d
u
c
ted
wit
h
a
sta
n
d
a
rd
d
a
tas
e
t
th
a
t
wa
s
p
u
b
li
sh
e
d
o
n
th
e
Ka
g
g
le
sy
ste
m
.
Th
e
e
v
a
lu
a
ti
o
n
is
p
e
rfo
rm
e
d
c
o
m
p
re
h
e
n
siv
e
l
y
wit
h
p
o
p
u
lar
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
a
rc
h
it
e
c
tu
re
s
,
in
c
lu
d
in
g
De
n
se
Ne
t2
0
1
,
De
n
se
N
e
t1
2
1
,
Eff
icie
n
tNe
tB0
,
a
n
d
M
o
b
il
e
Ne
tV2
,
c
o
m
p
a
re
d
in
two
sc
e
n
a
ri
o
s
with
a
n
d
with
o
u
t
th
e
ima
g
e
e
n
h
a
n
c
e
m
e
n
t
ste
p
.
Ex
p
e
rime
n
ts
h
a
v
e
sh
o
w
n
th
a
t
th
e
ima
g
e
e
n
h
a
n
c
e
m
e
n
t
ste
p
e
ffe
c
ti
v
e
ly
imp
ro
v
e
s
th
e
c
las
sifica
ti
o
n
p
e
rfo
rm
a
n
c
e
o
f
a
ll
m
o
d
e
ls,
c
lea
rly
th
ro
u
g
h
imp
o
rtan
t
sc
o
r
e
s
su
c
h
a
s
a
re
a
u
n
d
e
r
c
u
r
v
e
(
AUC
)
,
a
c
c
u
ra
c
y
,
F
1
-
sc
o
re
,
p
re
c
isio
n
,
a
n
d
re
c
a
ll
.
T
h
e
b
e
st
re
s
u
lt
tes
ted
is
t
h
e
Eff
icie
n
tNe
tB
0
m
o
d
e
l
wit
h
0
.
9
2
5
9
2
6
a
c
c
u
ra
c
y
sc
o
re
,
0
.
9
7
0
7
3
2
AU
C
sc
o
re
,
0
.
9
0
4
7
6
2
p
re
c
isio
n
sc
o
re
,
0
.
9
5
re
c
a
ll
sc
o
re
,
a
n
d
0
.
9
2
6
8
2
9
F
1
-
sc
o
re
.
I
n
a
d
d
it
i
o
n
,
q
u
a
li
tativ
e
a
n
a
ly
sis
u
sin
g
g
ra
d
ien
t
-
we
ig
h
ted
c
las
s
a
c
ti
v
a
ti
o
n
m
a
p
p
in
g
(
G
ra
d
-
CAM
)
sh
o
ws
t
h
a
t
th
e
re
su
lt
in
g
m
o
d
e
ls
h
a
v
e
sh
o
wn
a
fo
c
u
s
o
n
th
e
l
u
n
g
re
g
i
o
n
,
re
fl
e
c
ti
n
g
th
e
in
terp
re
tab
il
it
y
a
n
d
s
u
it
a
b
i
li
ty
fo
r
ra
d
io
l
o
g
ist
e
x
p
e
rti
se
.
K
ey
w
o
r
d
s
:
C
h
est X
-
r
ay
Dee
p
lear
n
in
g
I
m
ag
e
en
h
an
ce
m
en
t
Neu
r
al
n
etwo
r
k
T
u
b
er
cu
l
o
s
is
d
iag
n
o
s
is
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
:
Do
Nan
g
T
o
a
n
I
n
s
titu
te
o
f
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
,
Vietn
a
m
Aca
d
em
y
o
f
S
cien
ce
an
d
T
ec
h
n
o
lo
g
y
1
8
Ho
an
g
Qu
o
c
Viet,
C
au
Gia
y
,
Han
o
i1
0
0
7
2
,
Vietn
am
E
m
ail: d
n
to
an
@
io
it.a
c.
v
n
1.
I
NT
RO
D
UCT
I
O
N
On
e
o
f
th
e
m
o
s
t
s
er
io
u
s
th
r
ea
ts
to
h
u
m
a
n
h
ea
lth
is
tu
b
er
cu
lo
s
is
,
ca
u
s
ed
b
y
th
e
b
ac
ter
iu
m
m
y
co
b
ac
ter
iu
m
tu
b
er
c
u
lo
s
is
.
T
u
b
er
cu
lo
s
is
r
em
ain
s
o
n
e
o
f
th
e
lead
in
g
ca
u
s
es
o
f
d
e
ath
f
r
o
m
in
f
ec
tio
u
s
d
is
ea
s
es,
esp
ec
ially
in
d
ev
elo
p
in
g
co
u
n
tr
ies.
Mo
r
e
s
p
ec
if
ic
ally
,
r
ep
o
r
ts
o
n
tu
b
e
r
cu
lo
s
is
s
h
o
w
th
at
th
er
e
a
r
e
m
o
r
e
th
an
4
,
0
0
0
d
ea
th
s
an
d
ab
o
u
t 3
0
,
0
0
0
n
ew
ca
s
es e
v
er
y
d
a
y
,
b
ased
o
n
d
ata
r
elea
s
ed
o
n
W
o
r
ld
T
u
b
er
cu
lo
s
is
Day
2
0
2
2
[
1
]
.
T
r
ea
tm
en
t
o
f
t
u
b
er
cu
lo
s
is
d
ep
e
n
d
s
lar
g
ely
o
n
wh
en
th
e
d
is
ea
s
e
is
d
etec
ted
.
Mo
r
e
s
p
ec
if
ically
,
m
an
y
d
if
f
er
en
t
tu
b
er
c
u
lo
s
is
tr
ea
tm
en
t
r
e
g
im
en
s
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
[
2
]
,
s
u
ch
as
th
e
s
ix
-
m
o
n
th
o
r
n
in
e
-
m
o
n
th
r
if
am
p
in
,
is
o
n
ia
zid
,
p
y
r
az
i
n
am
id
e,
a
n
d
eth
am
b
u
to
l
(
R
I
PE
)
r
e
g
im
en
o
r
th
e
f
o
u
r
-
m
o
n
t
h
r
if
ap
en
tin
e
-
m
o
x
if
lo
x
ac
in
r
eg
i
m
en
.
E
ar
ly
d
etec
tio
n
an
d
ti
m
ely
tr
ea
tm
en
t
ar
e
im
p
o
r
tan
t
in
p
r
ev
en
tin
g
th
e
s
p
r
ea
d
o
f
t
u
b
er
c
u
lo
s
is
an
d
m
in
im
izin
g
its
p
u
b
lic
h
ea
lth
c
o
n
s
e
q
u
en
ce
s
.
I
n
t
h
e
m
e
d
i
c
a
l
p
r
a
c
t
i
c
e
o
f
d
i
a
g
n
o
s
i
n
g
p
u
l
m
o
n
a
r
y
t
u
b
e
r
c
u
l
o
s
i
s
,
c
h
e
s
t
X
-
r
a
y
i
m
a
g
es
a
r
e
a
c
o
m
m
o
n
a
n
d
e
f
f
e
c
t
i
v
e
t
y
p
e
o
f
d
a
t
a
f
o
r
d
i
a
g
n
o
s
i
s
.
C
h
es
t
X
-
r
a
y
d
a
t
a
is
lo
w
-
c
o
s
t
a
n
d
h
a
s
t
h
e
a
b
i
l
i
t
y
to
r
e
f
l
e
c
t
a
b
n
o
r
m
a
l
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
.
3
,
J
u
n
e
2
0
2
6
:
2
5
6
7
-
2
5
8
0
2568
c
h
a
r
a
c
t
e
r
is
t
i
cs
o
n
t
h
e
l
u
n
g
s
,
s
o
i
t
i
s
wi
d
e
l
y
u
s
e
d
f
o
r
s
c
r
e
e
n
i
n
g
p
u
l
m
o
n
a
r
y
t
u
b
e
r
c
u
l
o
s
i
s
.
H
o
w
e
v
e
r
,
t
h
e
i
n
t
e
r
p
r
e
t
a
ti
o
n
a
n
d
a
n
a
l
y
s
i
s
o
f
X
-
r
a
y
i
m
a
g
e
s
r
e
q
u
i
r
e
h
i
g
h
e
x
p
e
r
t
i
s
e
.
T
h
is
r
e
q
u
i
r
e
s
h
i
g
h
l
y
t
r
a
in
e
d
a
n
d
e
x
p
e
r
i
e
n
c
e
d
r
a
d
i
o
l
o
g
i
s
ts
,
w
h
i
c
h
is
al
w
a
y
s
a
s
c
a
r
c
e
g
r
o
u
p
i
n
t
h
e
m
e
d
i
c
a
l
te
a
m
.
T
h
e
r
e
f
o
r
e
,
t
h
e
r
e
i
s
a
n
e
ed
t
o
a
p
p
l
y
c
o
m
p
u
t
e
r
v
i
s
i
o
n
a
n
d
a
r
t
i
f
i
c
i
a
l
i
n
t
e
ll
i
g
e
n
c
e
t
e
c
h
n
o
l
o
g
i
es
t
o
b
u
i
l
d
e
f
f
e
ct
i
v
e
a
u
t
o
m
a
t
e
d
p
u
l
m
o
n
a
r
y
t
u
b
e
r
c
u
l
o
s
is
d
i
a
g
n
o
s
is
s
y
s
te
m
s
.
T
h
i
s
is
al
s
o
t
h
e
m
o
t
i
v
a
t
i
o
n
f
o
r
a
p
p
l
y
i
n
g
m
a
n
y
t
r
a
d
i
ti
o
n
a
l
m
et
h
o
d
s
,
s
u
c
h
as
r
e
s
e
a
r
c
h
[
3
]
u
s
i
n
g
t
r
a
d
i
t
i
o
n
a
l
n
e
u
r
al
n
e
t
w
o
r
k
c
o
m
b
i
n
e
d
w
i
t
h
a
g
e
n
et
i
c
al
g
o
r
it
h
m
,
r
e
s
e
a
r
c
h
[
4
]
u
s
i
n
g
a
n
id
e
n
t
i
f
i
c
a
ti
o
n
t
r
e
e
i
n
t
u
b
e
r
c
u
l
o
s
is
d
i
a
g
n
o
s
is
p
r
o
b
l
e
m
s
.
I
n
2
0
1
7
,
L
a
k
h
a
n
i
a
n
d
S
u
n
d
a
r
a
m
[
5
]
u
s
e
d
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
m
o
d
e
l
s
s
u
c
h
a
s
A
l
e
x
N
e
t
a
n
d
G
o
o
g
l
eNe
t
t
o
c
la
s
s
i
f
y
t
u
b
e
r
c
u
l
o
s
is
o
n
ch
e
s
t
X
-
r
a
y
,
a
c
h
i
e
v
i
n
g
h
i
g
h
a
c
cu
r
a
c
y
.
T
h
e
a
u
t
h
o
r
s
u
s
e
d
f
o
u
r
d
e
-
i
d
e
n
t
i
f
i
e
d
d
a
t
as
e
ts
,
a
n
d
a
c
h
i
e
v
e
d
t
h
e
b
e
s
t
a
r
e
a
u
n
d
e
r
c
u
r
v
e
(
A
UC
)
s
c
o
r
e
w
ith
0
.
9
9
a
l
s
o
i
n
2
0
1
7
,
R
a
j
p
u
r
k
a
r
e
t
a
l
.
[
6
]
p
r
e
s
e
n
t
e
d
t
h
e
C
h
e
X
N
e
t
m
o
d
e
l
.
C
h
eX
N
et
i
s
a
D
e
n
s
eN
e
t
1
2
1
m
o
d
e
l
f
o
r
d
e
t
e
c
t
i
n
g
p
n
e
u
m
o
n
i
a
f
r
o
m
c
h
e
s
t
X
-
r
a
y
d
a
t
a
l
e
a
r
n
e
d
o
n
t
h
e
C
h
e
s
t
X
-
r
a
y
1
4
,
t
h
e
d
a
ta
s
e
t
c
o
n
t
ai
n
i
n
g
o
v
e
r
1
0
0
,
0
0
0
i
m
a
g
e
s
.
T
h
e
a
u
t
h
o
r
s
n
o
t
e
d
t
h
a
t
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
o
u
t
p
e
r
f
o
r
m
e
d
h
u
m
a
n
r
a
d
i
o
l
o
g
i
s
ts
i
n
d
et
e
c
ti
n
g
p
n
e
u
m
o
n
i
a
f
r
o
m
c
h
e
s
t r
a
d
i
o
g
r
a
p
h
s
w
i
t
h
t
h
e
F
1
m
et
r
i
c
.
Ne
x
t
,
i
n
2
0
1
7
,
H
o
o
d
a
e
t
a
l
.
[
7
]
p
u
b
l
i
s
h
e
d
a
s
t
u
d
y
p
r
o
p
o
s
i
n
g
a
d
e
e
p
l
e
a
r
n
i
n
g
-
b
a
s
ed
t
u
b
e
r
c
u
l
o
s
is
d
e
te
c
t
i
o
n
s
y
s
t
e
m
.
I
n
t
h
e
s
t
u
d
y
,
t
h
e
a
u
t
h
o
r
s
u
s
e
d
a
n
e
u
r
a
l
n
e
tw
o
r
k
c
o
n
s
is
ti
n
g
o
f
s
e
v
e
n
c
o
n
v
o
l
u
t
i
o
n
al
l
a
y
e
r
s
a
n
d
t
h
r
e
e
f
u
l
l
y
c
o
n
n
e
c
te
d
l
a
y
e
r
s
.
E
x
p
e
r
i
m
e
n
ts
s
h
o
w
ed
t
h
a
t
t
h
e
b
e
s
t
ca
s
e
ac
h
i
e
v
e
d
0
.
8
2
0
9
f
o
r
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
a
n
d
0
.
9
4
7
3
f
o
r
o
v
e
r
a
l
l
a
c
c
u
r
a
c
y
.
I
n
2
0
1
8
,
R
a
j
p
u
r
k
a
r
e
t
a
l
.
[
8
]
c
o
m
p
a
r
e
d
t
h
e
e
f
f
e
c
t
i
v
e
n
e
s
s
o
f
t
h
e
C
h
e
X
Ne
t
m
o
d
e
l
wi
t
h
h
u
m
an
r
a
d
i
o
l
o
g
i
s
ts
i
n
d
i
a
g
n
o
s
i
n
g
a
v
a
r
i
e
t
y
o
f
p
a
t
h
o
l
o
g
i
es
o
n
c
h
e
s
t
r
a
d
i
o
g
r
a
p
h
s
.
R
a
d
i
o
l
o
g
is
ts
h
a
v
e
a
n
a
v
e
r
a
g
e
ex
p
e
r
i
e
n
c
e
o
f
t
w
e
l
v
e
y
e
a
r
s
,
a
n
d
t
h
r
e
e
s
e
n
i
o
r
r
a
d
i
o
l
o
g
y
r
e
s
i
d
e
n
ts
.
T
h
e
y
c
a
m
e
f
r
o
m
t
h
r
e
e
d
i
f
f
e
r
e
n
t
a
c
a
d
e
m
ic
i
n
s
t
i
t
u
t
i
o
n
s
.
I
n
2
0
2
0
,
C
h
a
n
d
r
a
e
t
a
l
.
[
9
]
a
p
p
l
i
e
d
a
h
i
e
r
a
r
c
h
i
c
a
l
f
ea
t
u
r
e
e
x
t
r
a
c
t
i
o
n
m
e
t
h
o
d
t
o
d
e
t
e
c
t
p
u
l
m
o
n
a
r
y
t
u
b
e
r
c
u
l
o
s
i
s
a
b
n
o
r
m
a
li
t
i
es
i
n
c
h
es
t
X
-
r
a
y
i
m
a
g
e
s
,
t
h
r
o
u
g
h
w
h
i
c
h
t
h
e
a
u
t
h
o
r
s
i
m
p
r
o
v
e
d
t
h
e
a
c
c
u
r
a
c
y
o
f
t
h
e
d
i
a
g
n
o
s
t
i
c
s
y
s
te
m
.
I
n
t
h
e
e
x
p
e
r
i
m
e
n
t
,
t
h
e
a
u
t
h
o
r
s
a
c
h
i
e
v
e
d
a
n
a
c
c
u
r
a
c
y
o
f
a
p
p
r
o
x
i
m
a
t
e
l
y
0
.
9
5
6
a
n
d
a
n
A
UC
o
f
a
p
p
r
o
x
i
m
a
t
e
l
y
0
.
9
5
o
n
t
h
e
M
o
n
t
g
o
m
e
r
y
d
a
t
as
e
t
,
a
n
d
s
i
m
il
a
r
l
y
0
.
9
9
4
a
n
d
0
.
9
9
o
n
t
h
e
S
h
e
n
z
h
e
n
d
a
t
a
s
et
.
I
n
2
0
1
9
,
P
a
s
a
e
t
a
l
.
[
1
0
]
p
r
o
p
o
s
e
d
a
s
i
m
p
l
e
n
e
t
w
o
r
k
t
o
d
i
a
g
n
o
s
e
t
u
b
e
r
c
u
l
o
s
is
f
a
s
t
e
r
.
I
n
t
h
e
e
x
p
e
r
i
m
e
n
t
,
t
h
e
a
u
t
h
o
r
s
al
s
o
d
is
c
u
s
s
e
d
t
h
a
t
t
h
e
i
r
a
r
c
h
i
te
c
t
u
r
e
c
a
n
p
r
es
e
r
v
e
t
h
e
a
c
c
u
r
a
c
y
s
c
o
r
e
v
e
r
s
u
s
p
r
e
v
i
o
u
s
s
t
u
d
i
e
s
.
I
n
t
h
a
t
y
e
a
r
,
M
e
r
a
j
e
t
a
l
.
[
1
1
]
a
l
s
o
p
r
e
s
e
n
t
e
d
r
es
e
a
r
c
h
u
s
i
n
g
f
o
u
r
d
i
f
f
e
r
e
n
t
c
o
n
v
o
l
u
t
i
o
n
a
l
n
eu
r
a
l
n
e
t
w
o
r
k
s
w
i
t
h
M
o
n
t
g
o
m
e
r
y
a
n
d
S
h
e
n
z
h
e
n
d
a
t
a
s
et
s
t
o
d
ia
g
n
o
s
e
t
u
b
e
r
c
u
l
o
s
is
e
f
f
e
c
ti
v
e
l
y
.
I
n
2
0
2
3
,
T
o
a
n
e
t
a
l
[
1
2
]
p
u
b
l
i
s
h
e
d
t
h
e
i
r
w
o
r
k
d
e
s
i
g
n
i
n
g
s
e
v
e
r
a
l
t
r
ai
n
i
n
g
s
t
r
a
t
e
g
i
es
wi
t
h
s
e
v
e
r
al
m
o
d
e
l
s
.
T
h
ei
r
s
t
u
d
y
a
i
m
e
d
t
o
e
v
a
l
u
a
te
a
n
d
d
i
s
c
u
s
s
t
u
b
e
r
c
u
l
o
s
is
d
i
a
g
n
o
s
i
s
i
n
V
i
et
n
a
m
e
s
e
d
a
t
a
,
as
w
el
l
a
s
t
h
e
i
m
p
a
ct
o
f
V
i
et
n
a
m
e
s
e
d
a
t
a
w
it
h
a
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
.
I
n
2
0
2
2
,
W
o
n
g
et
a
l
.
[
1
3
]
i
n
tr
o
d
u
ce
d
a
s
elf
-
atten
tio
n
n
et
wo
r
k
s
p
ec
if
ically
d
esig
n
e
d
f
o
r
d
etec
tin
g
p
u
lm
o
n
a
r
y
t
u
b
er
c
u
lo
s
is
.
T
h
e
a
u
th
o
r
s
also
c
o
n
d
u
cted
a
n
e
x
p
l
ain
ab
ilit
y
ev
alu
atio
n
.
T
h
e
s
tu
d
y
u
s
ed
ch
est
X‑
r
a
y
d
ata
f
r
o
m
m
u
ltip
le
co
u
n
tr
ies
an
d
th
e
ex
p
e
r
im
en
tal
r
esu
lts
s
h
o
wed
th
at
th
e
m
o
d
el
ac
h
ie
v
ed
an
ac
cu
r
ac
y
o
f
9
9
.
8
6
%.
I
n
2
0
2
4
,
Vin
h
et
a
l
.
[
1
4
]
f
o
c
u
s
ed
o
n
th
e
p
r
o
b
lem
o
f
au
t
o
m
atic
d
ia
g
n
o
s
is
o
f
p
u
l
m
o
n
ar
y
tu
b
e
r
cu
lo
s
is
in
th
e
co
n
tex
t
o
f
Vietn
am
.
T
h
e
au
th
o
r
s
b
u
ilt
a
cu
s
to
m
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
co
m
b
in
ed
with
a
h
ea
t
-
m
ap
p
r
ed
ictiv
e
v
is
u
aliza
tio
n
m
ec
h
an
is
m
.
T
h
e
s
tu
d
y
u
s
ed
a
lar
g
e
X
-
r
a
y
d
ataset
o
f
Vietn
a
m
ese
p
atien
ts
an
d
th
e
ex
p
er
im
en
tal
r
esu
lts
r
ec
o
r
d
e
d
th
at
th
eir
m
o
d
el
ac
h
ie
v
ed
b
o
th
s
en
s
itiv
ity
s
co
r
e
an
d
s
p
ec
if
icity
s
co
r
e
o
n
th
e
test
s
et
ab
o
v
e
0
.
9
,
an
d
th
e
h
i
g
h
l
ig
h
ted
r
eg
io
n
s
in
th
e
h
ea
t
-
m
ap
wer
e
s
im
ilar
to
th
e
s
y
m
p
to
m
s
an
n
o
tated
b
y
d
o
cto
r
s
.
I
n
2
0
2
4
,
Patel
et
a
l
.
[
1
5
]
s
tu
d
ie
d
th
e
p
r
o
b
lem
o
f
s
cr
ee
n
in
g
f
o
r
p
u
lm
o
n
ar
y
t
u
b
er
cu
lo
s
is
f
r
o
m
X
-
r
ay
im
ag
es
b
y
p
r
o
p
o
s
in
g
a
s
elf
-
s
u
p
er
v
is
ed
lear
n
in
g
n
etwo
r
k
w
ith
th
e
ab
ilit
y
to
ex
p
lain
,
to
r
ed
u
ce
th
e
co
s
t
an
d
d
if
f
icu
lty
i
n
d
ata
co
llectio
n
.
T
h
e
r
esu
lts
in
d
icate
d
th
at
th
eir
s
y
s
tem
ac
h
iev
ed
9
8
.
1
4
%
ac
c
u
r
ac
y
s
co
r
e
,
th
e
r
eb
y
s
tr
en
g
th
en
in
g
th
e
p
o
ten
tial
f
o
r
p
r
ac
tical
ap
p
licatio
n
in
co
m
m
u
n
ity
tu
b
er
cu
l
o
s
is
s
c
r
ee
n
in
g
.
I
n
2
0
2
4
,
B
is
ta
et
a
l
.
[
1
6
]
f
o
cu
s
ed
o
n
t
h
e
p
r
o
b
lem
o
f
d
etec
tin
g
p
u
lm
o
n
ar
y
tu
b
e
r
cu
lo
s
is
f
r
o
m
X
-
r
a
y
im
ag
es
u
s
in
g
t
h
e
YOL
Ov
7
ar
ch
itectu
r
e
co
m
b
i
n
ed
with
a
c
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
.
T
h
e
s
tu
d
y
was
co
n
d
u
cted
o
n
th
e
T
B
X1
1
K
d
ataset,
an
d
th
e
r
e
s
u
lts
s
h
o
wed
th
at
th
e
s
y
s
tem
ac
h
iev
ed
m
ap
o
f
0
.
5
8
7
.
I
n
2
0
2
4
,
K
o
tei
an
d
T
h
ir
u
n
a
v
u
k
ar
asu
[
1
7
]
s
tu
d
ied
th
e
p
r
o
b
lem
o
f
au
to
m
atic
d
ia
g
n
o
s
is
o
f
tu
b
er
c
u
lo
s
is
b
y
p
r
o
p
o
s
in
g
a
co
m
b
in
ed
m
o
d
el.
T
h
e
s
tu
d
y
u
s
es
th
e
T
B
X1
1
K
d
ataset
an
d
t
h
e
ex
p
e
r
im
en
tal
r
esu
lts
s
h
o
w
th
at
th
e
m
o
d
el
ac
h
iev
es
9
9
.
3
8
% a
cc
u
r
ac
y
.
W
e
also
in
clu
d
e
a
co
m
p
a
r
ativ
e
tab
le
o
f
tu
b
er
cu
lo
s
is
d
etec
tio
n
m
eth
o
d
s
th
at
co
n
s
is
ts
o
f
b
o
th
tr
ad
itio
n
al
im
ag
e
p
r
o
ce
s
s
in
g
a
n
d
C
NN
-
b
ased
m
o
d
els.
T
h
e
d
etails ar
e
in
T
ab
le
1
.
I
n
g
e
n
er
al,
d
etec
tin
g
p
u
lm
o
n
ar
y
tu
b
er
cu
lo
s
is
lesi
o
n
s
o
n
X
-
r
ay
im
a
g
es
is
ch
allen
g
i
n
g
.
On
th
e
o
n
e
h
an
d
,
th
is
co
m
es
f
r
o
m
th
e
i
m
ag
e
f
ea
tu
r
es
th
at
ar
e
o
f
ten
n
o
t
r
ea
lly
clea
r
,
s
o
m
etim
es
with
a
m
ild
lev
el
o
f
d
am
ag
e,
a
n
d
ca
n
b
e
m
i
x
ed
wit
h
an
ato
m
ical
s
tr
u
ctu
r
es
s
u
ch
a
s
r
ib
s
,
h
ea
r
t,
o
r
b
lo
o
d
v
ess
els.
On
th
e
o
th
e
r
h
an
d
,
th
e
q
u
ality
o
f
X
-
r
ay
im
ag
es
in
p
r
ac
tice
ca
n
b
e
af
f
ec
ted
b
y
d
i
f
f
er
en
t
s
h
o
o
tin
g
co
n
d
itio
n
s
d
u
e
to
d
if
f
er
en
ce
s
in
eq
u
ip
m
en
t
as
well
as
th
e
way
th
e
tech
n
ician
co
n
tr
o
ls
it.
T
h
is
lead
s
to
ce
r
tain
v
ar
iatio
n
s
r
elate
d
to
im
a
g
e
q
u
ality
,
s
u
ch
as
th
e
lev
el
o
f
co
n
tr
ast
b
etwe
en
d
am
a
g
ed
ar
ea
s
an
d
h
ea
lth
y
l
u
n
g
tis
s
u
e.
All
o
f
t
h
e
ab
o
v
e
p
r
o
b
lem
s
lead
to
ce
r
tain
d
if
f
ic
u
lties
f
o
r
d
ee
p
lear
n
in
g
m
o
d
els
wh
en
r
ec
o
g
n
izin
g
ab
n
o
r
m
alities
.
Acc
o
r
d
in
g
ly
,
a
f
ea
s
ib
le
s
o
lu
tio
n
to
th
is
p
r
o
b
lem
is
to
i
m
p
r
o
v
e
im
a
g
e
q
u
a
lity
b
ased
o
n
lo
c
al
co
n
tr
ast
en
h
an
ce
m
en
t,
th
er
eb
y
clar
if
y
in
g
th
e
d
etails
in
th
e
im
ag
e.
W
ith
th
at
g
o
al,
th
e
r
esear
ch
team
f
o
cu
s
ed
o
n
ap
p
ly
in
g
t
h
e
co
n
tr
ast
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
(
C
L
AHE
)
tech
n
iq
u
e,
wh
ic
h
h
elp
s
in
cr
ea
s
e
th
e
d
is
cr
im
in
atio
n
in
lo
w
-
co
n
tr
ast
X
-
r
ay
r
eg
i
o
n
s
wh
ile
tr
y
in
g
n
o
t
to
ca
u
s
e
n
o
is
e
am
p
lific
atio
n
in
h
ig
h
-
c
o
n
tr
ast
r
eg
io
n
s
.
T
h
is
is
ex
p
ec
ted
to
h
el
p
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
Dia
g
n
o
s
in
g
tu
b
ercu
lo
s
is
fr
o
m
X
-
r
a
y
ima
g
in
g
b
a
s
ed
o
n
co
n
tr
a
s
t limited
a
d
a
p
tive
…
(
N
g
u
ye
n
Tr
o
n
g
V
in
h
)
2569
im
p
r
o
v
e
th
e
class
if
icatio
n
p
er
f
o
r
m
an
ce
in
th
e
p
r
o
b
lem
o
f
d
ia
g
n
o
s
in
g
p
u
l
m
o
n
ar
y
tu
b
er
c
u
lo
s
is
o
n
ch
est X
-
r
ay
s
.
I
n
th
is
r
esear
c
h
co
n
tex
t,
C
L
AHE
-
b
ased
p
r
ep
r
o
ce
s
s
in
g
h
elp
s
im
p
r
o
v
e
th
e
p
e
r
f
o
r
m
an
ce
o
f
c
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
m
o
d
els wh
en
an
aly
zi
n
g
ch
est X
-
r
ay
im
ag
es.
C
L
AHE
h
elp
s
to
clar
if
y
th
e
s
tr
u
ctu
r
a
l d
etails in
th
e
lu
n
g
r
eg
io
n
t
h
at
m
ay
b
e
o
b
s
cu
r
ed
b
y
n
o
is
e
o
r
u
n
e
v
en
b
r
ig
h
t
n
ess
.
As
a
r
esu
lt,
p
ath
o
lo
g
ica
l
f
ea
tu
r
es
s
u
ch
as
tu
b
er
cu
lo
s
is
lesi
o
n
s
,
n
o
d
u
les,
o
r
lu
n
g
tis
s
u
e
ab
n
o
r
m
alities
b
ec
o
m
e
m
o
r
e
p
r
o
m
in
en
t
an
d
ar
e
ea
s
ily
r
ec
o
g
n
ized
b
y
th
e
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
m
o
d
el.
C
o
m
p
ar
e
d
with
p
r
e
p
r
o
ce
s
s
in
g
b
ased
o
n
g
lo
b
al
h
is
to
g
r
a
m
eq
u
aliza
tio
n
,
C
L
AHE
b
alan
ce
s
an
d
r
etain
s
m
o
r
e
im
p
o
r
tan
t
lo
ca
l
f
ea
tu
r
es.
At
th
e
s
am
e
tim
e,
u
n
lik
e
ed
g
e
-
en
h
an
cin
g
p
r
e
p
r
o
ce
s
s
in
g
th
at
em
p
h
asizes
th
e
b
o
u
n
d
a
r
y
,
C
L
AHE
p
r
o
v
i
d
es
m
o
r
e
d
etai
led
in
f
o
r
m
atio
n
o
n
th
e
en
tire
lu
n
g
tex
tu
r
e
to
h
elp
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
etwo
r
k
m
o
d
els.
Ou
r
s
tu
d
y
is
s
et
in
th
e
co
n
te
x
t
o
f
d
ev
elo
p
in
g
a
n
ef
f
ec
tiv
e
d
ee
p
lear
n
in
g
s
o
lu
tio
n
f
o
r
d
iag
n
o
s
in
g
p
u
lm
o
n
a
r
y
tu
b
e
r
cu
lo
s
is
f
r
o
m
ch
est
X
-
r
ay
d
ata.
I
n
th
e
s
tu
d
y
,
we
will
ev
alu
ate
th
e
r
o
le
o
f
im
ag
e
en
h
an
ce
m
en
t,
an
d
s
p
ec
if
ically
C
L
AHE
,
in
b
o
o
s
tin
g
th
e
d
iag
n
o
s
tic
q
u
ality
o
f
d
ee
p
lear
n
in
g
m
o
d
els.
T
o
co
n
d
u
ct
a
co
m
p
r
eh
e
n
s
iv
e
ev
al
u
atio
n
,
w
e
will
co
m
p
a
r
e
th
e
p
e
r
f
o
r
m
a
n
ce
o
f
f
o
u
r
p
o
p
u
lar
d
ee
p
lea
r
n
in
g
ar
ch
itectu
r
es,
in
clu
d
in
g
Den
s
eNe
t2
0
1
,
De
n
s
eNe
t1
2
1
,
E
f
f
icien
tNetB
0
,
an
d
Mo
b
ileNetV2
,
o
n
th
e
t
ask
o
f
class
if
y
in
g
p
u
lm
o
n
a
r
y
tu
b
er
cu
lo
s
is
f
r
o
m
ch
est
X
-
r
ay
s
,
in
two
s
ce
n
ar
io
s
:
with
an
d
with
o
u
t
im
ag
e
en
h
a
n
ce
m
en
t.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
o
u
r
r
esear
ch
in
clu
d
e:
i)
p
r
o
p
o
s
in
g
a
s
im
p
le
an
d
e
f
f
icien
t
p
r
o
ce
s
s
in
g
p
r
o
ce
d
u
r
e
th
at
ca
n
b
e
ea
s
ily
in
teg
r
ated
in
to
tu
b
e
r
cu
lo
s
is
d
iag
n
o
s
is
s
u
p
p
o
r
t
s
y
s
tem
s
u
s
in
g
ch
est
X
-
r
ay
im
ag
es
;
ii)
clar
if
y
in
g
th
e
h
y
p
o
th
esis
o
f
im
ag
e
q
u
ality
e
n
h
an
ce
m
e
n
t
th
r
o
u
g
h
ev
alu
atin
g
th
e
class
if
icatio
n
p
e
r
f
o
r
m
an
ce
o
n
f
o
u
r
d
if
f
er
en
t
p
o
p
u
lar
d
ee
p
lear
n
in
g
m
o
d
els,
n
am
ely
Den
s
eNe
t2
0
1
,
Den
s
eNe
t1
2
1
,
E
f
f
icien
tNetB
0
,
an
d
Mo
b
ileNetV2
;
an
d
iii)
d
is
cu
s
s
in
g
th
e
q
u
alitativ
e
ev
alu
atio
n
th
r
o
u
g
h
g
r
ad
ien
t
-
weig
h
ted
class
ac
tiv
atio
n
m
a
p
p
i
n
g
(
Gr
a
d
-
C
AM
)
o
f
th
e
co
n
ce
n
tr
atio
n
r
e
g
io
n
s
o
f
d
ee
p
lear
n
in
g
m
o
d
els wh
e
n
ap
p
ly
in
g
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
T
ab
le
1
.
So
m
e
t
u
b
er
c
u
lo
s
is
d
iag
n
o
s
is
s
tu
d
ies
A
u
t
h
o
r
s
Y
e
a
r
M
o
d
e
l
El
v
e
r
e
n
a
n
d
Y
u
mu
ş
a
k
[
3
]
2
0
1
1
Tr
a
d
i
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
a
n
d
g
e
n
e
t
i
c
a
l
g
o
r
i
t
h
m
D
o
n
g
a
r
d
i
v
e
e
t
a
l
.
[
4
]
2
0
1
1
I
d
e
n
t
i
f
i
c
a
t
i
o
n
t
r
e
e
La
k
h
a
n
i
a
n
d
S
u
n
d
a
r
a
m
[
5
]
2
0
1
7
A
l
e
x
N
e
t
a
n
d
G
o
o
g
l
e
N
e
t
R
a
j
p
u
r
k
a
r
e
t
a
l
.
[
6
]
2
0
1
7
C
h
e
X
N
e
t
H
o
o
d
a
e
t
a
l
.
[
7
]
2
0
1
7
A
sp
e
c
i
f
i
c
C
N
N
R
a
j
p
u
r
k
a
r
e
t
a
l
.
[
8
]
2
0
1
8
C
h
e
X
N
e
X
t
C
h
a
n
d
r
a
e
t
a
l
.
[
9
]
2
0
2
0
H
i
e
r
a
r
c
h
i
c
a
l
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
sc
h
e
m
e
P
a
sa
e
t
a
l
.
[
1
0
]
2
0
1
9
A
si
m
p
l
e
C
N
N
M
e
r
a
j
e
t
a
l
.
[
1
1
]
2
0
1
9
G
o
o
g
l
e
N
e
t
,
R
e
sN
e
t
5
0
,
V
G
G
-
1
6
,
a
n
d
V
G
G
-
19.
To
a
n
e
t
a
l
.
[
1
2
]
2
0
2
3
R
e
sN
e
t
5
0
,
R
e
sN
e
t
3
4
,
A
l
e
x
N
e
t
,
a
n
d
D
e
n
s
e
N
e
t
1
2
1
W
o
n
g
e
t
a
l
.
[
1
3
]
2
0
2
2
TB
N
e
t
V
i
n
h
e
t
a
l
.
[
1
4
]
2
0
2
4
A
c
u
s
t
o
m C
N
N
P
a
t
e
l
e
t
a
l
.
[
1
5
]
2
0
2
4
A
sel
f
-
su
p
e
r
v
i
s
e
d
l
e
a
r
n
i
n
g
n
e
t
w
o
r
k
B
i
s
t
a
e
t
a
l
.
[
1
6
]
2
0
2
4
Y
O
LO
v
7
K
o
t
e
i
e
t
a
l
.
[
1
7
]
2
0
2
4
D
a
t
a
e
f
f
i
c
i
e
n
t
i
m
a
g
e
t
r
a
n
sf
o
r
mer
a
n
d
R
e
sN
e
t
16
2.
M
E
T
H
O
D
2
.
1
.
Da
t
a
s
et
I
n
th
is
r
esear
ch
,
we
u
s
ed
th
e
Pu
lm
o
n
ar
y
ch
est
X
-
r
ay
ab
n
o
r
m
alities
d
atase
t
[
1
8
]
p
u
b
licly
a
v
ailab
le
o
n
th
e
Kag
g
le
p
latf
o
r
m
.
T
h
e
d
ata
s
et
was
b
u
ilt
b
ased
o
n
d
ata
f
r
o
m
Mo
n
tg
o
m
e
r
y
C
o
u
n
ty
an
d
Sh
en
zh
en
Ho
s
p
ital.
T
h
is
d
ataset
was
b
u
ilt
as
o
n
e
o
f
th
e
e
f
f
o
r
ts
to
s
er
v
e
th
e
r
esear
ch
o
n
class
if
y
in
g
tu
b
er
cu
lo
s
is
f
r
o
m
ch
est
X
-
r
a
y
d
ata.
T
h
e
im
ag
e
d
ataset
c
o
n
s
is
t
s
o
f
8
0
0
im
ag
es,
o
f
wh
i
ch
6
6
2
im
ag
es
ar
e
f
r
o
m
th
e
Sh
en
zh
e
n
s
et
a
n
d
1
3
8
im
a
g
es
ar
e
f
r
o
m
th
e
M
o
n
tg
o
m
e
r
y
s
et.
B
esid
es,
th
e
lab
el
d
is
tr
ib
u
tio
n
in
t
h
e
Sh
e
n
zh
en
s
et
in
clu
d
es
3
2
6
n
o
r
m
al
im
ag
es
a
n
d
3
3
6
tu
b
er
c
u
lo
s
is
im
ag
es.
T
h
e
lab
el
d
is
tr
ib
u
tio
n
in
t
h
e
Mo
n
tg
o
m
er
y
s
et
in
clu
d
es
8
0
n
o
r
m
al
im
ag
es a
n
d
5
8
tu
b
e
r
cu
lo
s
is
im
ag
es.
T
h
e
lab
el
d
is
t
r
ib
u
tio
n
is
illu
s
tr
ated
in
Fig
u
r
e
1
.
Fig
u
r
e
1
.
T
h
e
lab
el
d
is
tr
ib
u
tio
n
in
d
atasets
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
.
3
,
J
u
n
e
2
0
2
6
:
2
5
6
7
-
2
5
8
0
2570
2
.
2
.
M
o
dels
Nex
t,
E
f
f
icien
tNet
[
1
9
]
was
p
r
o
p
o
s
ed
as
a
f
am
ily
o
f
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
w
ith
n
o
tab
le
im
p
r
o
v
em
e
n
ts
.
Den
s
eNe
t
is
a
well
-
k
n
o
w
n
d
ee
p
n
eu
r
al
n
et
wo
r
k
i
n
tr
o
d
u
ce
d
b
y
Hu
an
g
et
a
l
.
[
2
0
]
i
n
2
0
1
7
.
I
n
th
e
wo
r
k
,
th
e
a
u
th
o
r
s
p
r
esen
t
ed
a
d
esig
n
in
wh
ic
h
ea
ch
lay
er
is
d
ir
ec
tly
lin
k
ed
to
p
r
e
v
io
u
s
lay
er
s
.
T
h
u
s
,
th
e
in
p
u
t o
f
ea
ch
lay
er
will in
clu
d
e
n
o
t o
n
ly
th
e
o
u
t
p
u
t o
f
its
d
ir
ec
t p
r
ev
io
u
s
lay
er
,
b
u
t a
ls
o
th
e
o
u
tp
u
ts
f
r
o
m
o
th
e
r
p
r
ev
io
u
s
lay
er
s
,
as
d
escr
ib
ed
in
Fig
u
r
e
2
.
T
h
is
is
a
d
esig
n
th
at
aim
s
to
r
eu
s
e
f
ea
tu
r
es
an
d
h
an
d
le
g
r
a
d
ien
t
v
an
is
h
in
g
.
I
n
th
is
s
tu
d
y
,
we
u
s
e
two
v
er
s
io
n
s
o
f
De
n
s
eNe
t
,
in
clu
d
in
g
Den
s
eNe
t1
2
1
,
a
n
d
D
en
s
eNe
t2
0
1
.
Fig
u
r
e
2
.
A
5
-
lay
er
d
en
s
e
b
lo
c
k
[
2
0
]
I
n
th
is
s
tu
d
y
,
we
u
s
e
th
e
E
f
f
i
cien
tNetB
0
v
er
s
io
n
,
wh
ic
h
is
a
b
asic
v
er
s
io
n
an
d
d
o
es
n
o
t
r
eq
u
ir
e
to
o
m
an
y
r
eso
u
r
ce
s
.
An
o
th
e
r
ca
s
e
we
u
s
ed
is
Mo
b
ileNet
[
2
1
]
.
T
h
is
is
a
lig
h
tweig
h
t
C
NN
m
o
d
el
f
am
ily
d
esig
n
e
d
with
m
o
b
ile
d
e
v
ices
in
m
in
d
.
T
h
e
d
esig
n
u
s
es
th
e
d
ep
t
h
-
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
ti
o
n
s
m
ec
h
an
is
m
as
d
escr
ib
ed
in
Fig
u
r
e
3
,
wh
ich
allo
ws f
o
r
r
ed
u
ce
d
co
m
p
u
tatio
n
s
an
d
p
ar
a
m
eter
co
u
n
ts
wh
ile
m
ain
tain
in
g
a
ce
r
tain
lev
el
o
f
p
er
f
o
r
m
an
ce
.
T
h
e
v
er
s
io
n
we
t
ested
in
th
is
s
tu
d
y
is
Mo
b
ileNetV2
.
Fig
u
r
e
3
.
Dep
t
h
-
wis
e
s
ep
ar
ab
l
e
co
n
v
o
lu
tio
n
s
[
2
1
]
2
.
3
.
T
he
da
t
a
pro
ce
s
s
ing
pro
ce
du
re
Ou
r
d
ata
p
r
e
p
r
o
ce
s
s
in
g
p
ip
eli
n
e
is
d
escr
ib
ed
in
Fig
u
r
e
4
.
First,
th
e
im
ag
e
is
r
ea
d
as
a
g
r
ay
s
ca
le
im
ag
e.
T
h
e
im
ag
e
d
ata
is
th
e
n
n
o
r
m
alize
d
to
th
e
p
ix
el
v
al
u
e
s
in
th
e
r
a
n
g
e
[
0
,
1
]
b
y
d
iv
i
d
in
g
th
e
p
ix
el
v
alu
es
b
y
2
5
5
.
Nex
t,
th
e
im
ag
e
is
r
esized
ac
co
r
d
i
n
g
to
a
p
r
e
d
eter
m
in
ed
p
ar
a
m
eter
.
Fin
ally
,
th
e
im
ag
e
is
r
ef
o
r
m
atted
to
b
e
c
o
m
p
atib
le
with
th
e
f
o
u
r
-
d
im
e
n
s
io
n
al
ten
s
o
r
in
p
u
t
[
1
,
1
,
h
eig
h
t,
wid
th
]
o
f
th
e
d
e
ep
lear
n
in
g
lib
r
a
r
y
.
Usu
ally
,
in
th
e
p
r
ep
r
o
ce
s
s
in
g
p
r
o
ce
s
s
,
s
o
m
e
is
s
u
es
ca
n
b
e
co
n
s
id
er
ed
s
u
ch
as
lu
n
g
s
eg
m
en
tatio
n
o
r
n
o
is
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
Dia
g
n
o
s
in
g
tu
b
ercu
lo
s
is
fr
o
m
X
-
r
a
y
ima
g
in
g
b
a
s
ed
o
n
co
n
tr
a
s
t limited
a
d
a
p
tive
…
(
N
g
u
ye
n
Tr
o
n
g
V
in
h
)
2571
r
ed
u
ctio
n
.
I
n
th
is
s
tu
d
y
,
d
ata
n
o
r
m
aliza
tio
n
h
as
s
u
p
p
o
r
te
d
n
o
is
e
r
ed
u
ctio
n
,
wh
ich
h
elp
s
to
im
p
r
o
v
e
im
ag
e
q
u
ality
an
d
s
u
p
p
o
r
t th
e
m
o
d
el
to
lear
n
m
o
r
e
ef
f
ec
tiv
el
y
.
Af
ter
lo
ad
in
g
th
e
im
a
g
e,
th
e
n
ex
t
s
tep
is
to
p
er
f
o
r
m
co
n
t
r
a
s
t
en
h
an
ce
m
en
t.
T
h
e
tech
n
iq
u
e
u
s
ed
in
th
is
s
tep
is
C
L
AHE
[
2
2
]
.
T
h
is
en
h
an
ce
s
th
e
lo
ca
l
c
o
n
tr
ast
in
im
ag
es,
th
e
r
eb
y
m
ak
in
g
t
h
e
d
etails
in
ch
est
X
-
r
ay
d
ata,
f
o
r
ex
am
p
le
in
th
e
lu
n
g
r
eg
i
o
n
s
,
clea
r
er
.
T
h
is
tr
a
n
s
f
o
r
m
atio
n
is
p
er
f
o
r
m
e
d
d
ir
e
ctly
o
n
th
e
g
r
a
p
h
ics
p
r
o
ce
s
s
in
g
u
n
it (
GPU
)
with
th
e
h
elp
o
f
f
u
n
ctio
n
s
in
th
e
Ko
r
n
ia
lib
r
ar
y
[
2
3
]
.
Fig
u
r
e
4
.
T
h
e
d
ata
p
r
ep
r
o
ce
s
s
in
g
p
ip
elin
e
Nex
t,
th
e
im
ag
e
will
b
e
p
ass
ed
th
r
o
u
g
h
a
d
ata
a
u
g
m
en
tat
io
n
tr
an
s
f
o
r
m
atio
n
s
tep
to
en
h
an
ce
th
e
d
iv
er
s
ity
o
f
th
e
in
p
u
t
im
ag
e
f
o
r
th
e
tr
ain
in
g
p
h
ase
o
f
t
h
e
d
ee
p
lear
n
in
g
m
o
d
el.
Sp
ec
if
ically
,
th
e
im
ag
e
will
b
e
ap
p
lied
with
r
a
n
d
o
m
tr
an
s
f
o
r
m
atio
n
s
in
clu
d
in
g
h
o
r
izo
n
tal
f
lip
p
in
g
,
r
an
d
o
m
lef
t
-
r
i
g
h
t
r
o
ta
tio
n
,
r
an
d
o
m
s
lig
h
t
tr
an
s
latio
n
in
b
o
th
h
o
r
izo
n
tal
an
d
v
e
r
tical
d
ir
ec
tio
n
s
,
an
d
r
a
n
d
o
m
b
r
ig
h
tn
ess
ch
an
g
e
.
T
h
e
s
e
tr
an
s
f
o
r
m
atio
n
s
ar
e
p
er
f
o
r
m
e
d
n
o
n
-
u
n
if
o
r
m
ly
f
o
r
ea
c
h
im
ag
e
in
e
ac
h
b
atch
,
an
d
th
e
en
tire
p
ip
elin
e
is
e
x
ec
u
ted
o
n
th
e
GPU
to
in
cr
ea
s
e
p
r
o
ce
s
s
in
g
s
p
ee
d
.
Fi
n
ally
,
p
ix
el
v
al
u
e
is
n
o
r
m
aliz
ed
u
s
in
g
m
ea
n
an
d
s
tan
d
a
r
d
d
ev
iatio
n
ca
lcu
lated
f
r
o
m
I
m
a
g
eNe
t.
T
h
is
is
a
co
m
m
o
n
n
o
r
m
aliza
tio
n
s
tep
wh
en
f
ee
d
in
g
im
ag
es in
to
n
eu
r
al
n
et
wo
r
k
s
.
T
h
e
co
n
te
n
t
is
also
s
h
o
wn
in
p
s
eu
d
o
co
d
e
i
n
Alg
o
r
ith
m
1
.
Usi
n
g
th
ese
tr
an
s
f
o
r
m
atio
n
s
is
ex
p
ec
ted
to
h
elp
th
e
m
o
d
el
lear
n
m
o
r
e
g
en
er
al
f
ea
tu
r
es
an
d
im
p
r
o
v
e
its
ab
ilit
y
to
r
esis
t
o
v
er
f
itti
n
g
.
I
n
th
e
e
x
p
er
im
en
t,
in
o
r
d
er
to
d
em
o
n
s
tr
ate
th
e
h
y
p
o
th
esis
o
f
th
e
r
o
le
o
f
i
m
ag
e
en
h
an
ce
m
e
n
t,
th
e
ch
est
X
-
r
ay
d
ata
p
r
o
ce
s
s
in
g
alg
o
r
ith
m
was
d
esig
n
ed
to
en
ab
le
o
r
d
is
ab
le
co
n
tr
ast
en
h
an
ce
m
en
t
with
C
L
AHE
.
T
h
u
s
,
in
th
e
e
x
p
er
im
en
ts
,
we
will
p
er
f
o
r
m
two
o
p
tio
n
s
o
f
d
ata
p
r
o
ce
s
s
in
g
: w
ith
o
r
wit
h
o
u
t u
s
in
g
th
e
co
n
tr
ast en
h
an
c
em
en
t step
.
Alg
o
r
ith
m
1
.
T
h
e
d
ata
p
r
o
ce
s
s
in
g
p
r
o
ce
d
u
r
e
Input:
image
file path file_path
Output:
image
tensor img_tensor
Process:
1
:
gray_image = load_gray_image(file_path)
2
:
img_tensor = convert_tensor(gray_image)
3
:
normalize_0_1(img_tensor)
4
:
resize(img_tensor)
5
:
enhance_contrast(img_tensor)
6
:
augment_transform(img_tensor)
7
:
resize(img_tensor)
8
:
normalize_mean_stddev(img_tensor)
2
.
4
.
E
x
perim
ent
c
o
nfig
ura
t
i
o
n
I
n
th
is
s
tu
d
y
,
C
L
AHE
was
s
et
with
clip
L
im
it a
s
2
an
d
tileGr
i
d
Size
as 8
×8
.
T
h
is
en
s
u
r
es th
at
th
e
ce
lls
ar
e
s
m
all
en
o
u
g
h
to
h
i
g
h
lig
h
t
th
e
lu
n
g
tex
tu
r
e
d
etails
an
d
ab
n
o
r
m
al
ar
ea
s
,
b
u
t
s
till
m
ain
tain
th
e
co
n
tin
u
ity
o
f
th
e
lu
n
g
tis
s
u
e
with
o
u
t
tear
i
n
g
it
to
o
m
u
ch
.
T
h
e
clip
L
im
it
v
alu
e
o
f
2
h
elp
s
to
av
o
id
o
v
er
-
e
n
h
an
cin
g
th
e
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
.
3
,
J
u
n
e
2
0
2
6
:
2
5
6
7
-
2
5
8
0
2572
co
n
tr
ast,
wh
ich
ca
n
ca
u
s
e
an
i
m
b
alan
ce
b
etwe
en
lig
h
t
a
n
d
d
ar
k
ar
ea
s
an
d
m
a
k
e
th
e
im
ag
e
lo
o
k
ar
tific
ial
an
d
u
n
n
atu
r
al.
T
h
er
ef
o
r
e,
s
ettin
g
th
ese
p
ar
am
eter
s
co
n
tr
ib
u
tes
to
s
u
p
p
o
r
tin
g
th
e
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
m
o
d
el
to
lear
n
an
d
r
ec
o
g
n
ize
m
o
r
e
ac
cu
r
ately
.
I
n
th
is
r
esear
ch
,
p
r
ep
r
o
ce
s
s
ed
ch
est
X
-
r
ay
d
ata
is
f
e
d
in
to
m
o
d
els
to
p
er
f
o
r
m
p
u
lm
o
n
ar
y
tu
b
er
cu
lo
s
is
d
iag
n
o
s
is
class
if
icat
io
n
.
As
s
h
o
wn
in
Fig
u
r
e
5
,
th
e
p
r
o
ce
s
s
in
clu
d
es
two
o
p
tio
n
s
A
an
d
B
.
Op
tio
n
A
u
s
es
th
e
o
r
ig
in
al
im
ag
e
with
o
u
t
c
o
n
tr
a
s
t
en
h
an
ce
m
en
t
a
n
d
o
p
tio
n
B
u
s
es
th
e
im
ag
e
th
at
h
as
b
ee
n
co
n
tr
ast
en
h
a
n
ce
d
u
s
in
g
th
e
C
L
AHE
m
eth
o
d
.
I
t a
llo
ws
u
s
to
ass
ess
th
e
ef
f
ec
tiv
en
ess
an
d
v
alid
ate
o
u
r
r
esear
ch
h
y
p
o
th
esis
o
n
t
h
e
ef
f
ec
t
o
f
c
o
n
tr
ast
en
h
an
ce
m
en
t
u
s
in
g
th
e
C
L
AHE
m
eth
o
d
o
n
th
e
p
u
lm
o
n
ar
y
tu
b
er
cu
lo
s
is
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
o
f
m
o
d
els.
Fig
u
r
e
5
.
T
h
e
tr
ain
in
g
wo
r
k
f
lo
w
I
n
th
e
ex
p
e
r
im
en
t,
th
e
s
ele
cted
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
m
o
d
els
in
clu
d
e
Den
s
eNe
t1
2
1
,
Den
s
eNe
t2
0
1
,
E
f
f
icien
tNetB
0
,
an
d
Mo
b
ileNetV2
.
T
h
ese
ar
e
all
p
o
p
u
lar
ar
c
h
itectu
r
es
in
im
ag
e
r
ec
o
g
n
itio
n
p
r
o
b
lem
s
a
n
d
a
r
e
s
u
itab
le
f
o
r
m
ed
ical
a
p
p
licatio
n
s
.
T
h
e
ex
p
er
im
en
tal
r
esu
lts
ar
e
ex
p
ec
ted
to
s
h
o
w
th
at
in
teg
r
atin
g
co
n
tr
ast
en
h
an
ce
m
en
t
u
s
in
g
th
e
C
L
AHE
m
eth
o
d
ca
n
im
p
r
o
v
e
t
h
e
p
u
lm
o
n
ar
y
tu
b
er
cu
lo
s
is
r
ec
o
g
n
itio
n
p
e
r
f
o
r
m
an
ce
o
f
d
e
ep
lear
n
i
n
g
m
o
d
els
b
y
h
ig
h
lig
h
tin
g
th
e
tex
tu
r
e
r
eg
io
n
s
in
th
e
lu
n
g
s
.
I
n
ad
d
itio
n
,
th
e
v
alid
atio
n
o
n
m
an
y
d
if
f
er
e
n
t
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
ar
ch
itectu
r
es
h
elp
s
co
n
f
ir
m
th
e
s
ca
lab
ilit
y
an
d
s
tab
ilit
y
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
in
th
e
p
ap
e
r
,
cr
ea
tin
g
a
b
asis
f
o
r
p
r
ac
tical
ap
p
licatio
n
in
p
u
lm
o
n
ar
y
tu
b
er
cu
lo
s
is
d
iag
n
o
s
is
s
u
p
p
o
r
t
s
y
s
tem
s
f
r
o
m
ch
est X
-
r
ay
im
a
g
e
d
ata.
Du
r
in
g
th
e
tr
ai
n
in
g
o
f
th
e
class
if
icatio
n
m
o
d
el,
we
p
er
f
o
r
m
two
s
tep
s
f
o
r
tr
a
n
s
f
er
lear
n
in
g
.
W
e
u
s
e
a
p
r
e
-
tr
ain
ed
m
o
d
el
with
I
m
ag
eNe
t
p
r
o
v
id
ed
in
th
e
Py
T
o
r
c
h
lib
r
ar
y
[
2
4
]
.
W
e
r
ep
lace
th
e
last
cla
s
s
if
ica
tio
n
lay
er
to
s
u
it
o
u
r
p
r
o
b
lem
.
I
n
th
e
f
ir
s
t
two
e
p
o
ch
s
,
o
n
ly
t
h
e
last
p
ar
t
o
f
th
e
m
o
d
el
is
tr
ain
ed
,
w
h
ile
th
e
r
em
ain
in
g
lay
er
s
r
etain
th
eir
i
n
itial
weig
h
ts
.
Fro
m
th
e
th
ir
d
ep
o
ch
,
th
e
en
tire
m
o
d
el
is
tr
ain
ed
,
allo
win
g
all
m
o
d
el
p
ar
a
m
eter
s
to
b
e
u
p
d
ate
d
ev
er
y
ep
o
c
h
.
I
n
ter
m
s
o
f
o
p
tim
izatio
n
,
a
r
c
h
itectu
r
es
ar
e
t
r
ain
ed
u
s
in
g
t
h
e
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
co
m
b
in
ed
with
th
e
Ad
am
alg
o
r
ith
m
[
2
5
]
.
T
h
e
lear
n
in
g
r
ate
is
ad
ju
s
ted
u
s
in
g
th
e
o
n
e
cy
cle
m
ec
h
an
is
m
[
2
6
]
.
T
h
is
s
ettin
g
is
ex
p
ec
ted
to
h
elp
th
e
lear
n
i
n
g
r
ate
ch
a
n
g
e
r
ea
s
o
n
ab
ly
d
u
r
in
g
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
,
av
o
id
f
allin
g
in
to
l
o
ca
l
ex
tr
em
es,
an
d
h
elp
t
h
e
m
o
d
el
co
n
v
er
g
e
f
aster
.
I
n
a
d
d
itio
n
,
tr
ain
i
n
g
is
also
s
u
p
p
o
r
te
d
b
y
th
e
g
r
ad
ien
t
ac
cu
m
u
latio
n
tech
n
iq
u
e,
wh
ic
h
h
elp
s
th
e
m
o
d
el
tr
ain
m
o
r
e
e
f
f
icien
tly
u
n
d
er
lim
ited
GPU
m
em
o
r
y
c
o
n
d
itio
n
s
.
I
n
th
is
s
tu
d
y
,
to
ev
alu
ate
th
e
tr
ain
in
g
r
esu
lts
o
f
th
e
d
ee
p
lear
n
in
g
m
o
d
el
f
o
r
class
if
y
in
g
tu
b
er
cu
lo
s
is
f
r
o
m
ch
est
X
-
r
ay
im
ag
es,
we
u
s
ed
th
e
m
o
d
el
p
er
f
o
r
m
an
ce
ev
alu
atio
n
in
d
ex
es,
in
clu
d
i
n
g
AUC,
ac
cu
r
ac
y
,
F1
-
s
co
r
e
,
p
r
ec
is
io
n
,
an
d
r
ec
a
ll
.
I
n
a
d
d
itio
n
,
to
b
etter
u
n
d
er
s
tan
d
h
o
w
th
e
m
o
d
el
m
ak
es
p
r
ed
ictio
n
s
,
th
e
Gr
ad
-
C
AM
tech
n
iq
u
e
[
2
7
]
h
e
lp
s
g
en
er
ate
h
ea
tm
ap
s
f
r
o
m
t
h
e
X
-
r
ay
im
ag
es.
Gr
a
d
-
C
AM
will
ca
lcu
late
th
e
g
r
ad
ien
t
o
f
th
e
last
class
if
icati
o
n
lay
er
o
n
th
e
s
p
atial
f
ea
tu
r
e
asp
ec
t
to
d
eter
m
in
e
th
e
p
r
o
b
a
b
ilit
y
m
ap
to
ass
ess
th
e
m
o
d
el'
s
co
n
ce
n
tr
atio
n
o
n
th
e
im
ag
e.
I
n
th
is
s
tu
d
y
,
Gr
ad
-
C
AM
h
ea
tm
ap
s
ca
n
h
elp
u
s
ch
ec
k
wh
eth
er
th
e
m
o
d
el
r
ea
lly
p
ay
s
atten
tio
n
to
th
e
lu
n
g
r
eg
io
n
s
o
r
lesi
o
n
s
th
at
ar
e
ch
ar
ac
ter
is
tic
o
f
tu
b
er
cu
lo
s
is
,
th
er
eb
y
in
cr
ea
s
in
g
r
eliab
ilit
y
an
d
aim
i
n
g
to
s
u
p
p
o
r
t c
lin
ical
v
alid
atio
n
.
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
Dia
g
n
o
s
in
g
tu
b
ercu
lo
s
is
fr
o
m
X
-
r
a
y
ima
g
in
g
b
a
s
ed
o
n
co
n
tr
a
s
t limited
a
d
a
p
tive
…
(
N
g
u
ye
n
Tr
o
n
g
V
in
h
)
2573
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
illu
s
tr
ates
th
e
ex
p
er
im
en
tal
r
esu
lts
o
f
f
o
u
r
m
o
d
els,
in
clu
d
in
g
De
n
s
eNe
t1
2
1
,
Den
s
eNe
t2
0
1
,
Mo
b
ileNetV2
,
an
d
E
f
f
icien
tNetB
0
,
tr
ain
ed
with
two
d
if
f
er
en
t
tr
ain
in
g
s
ce
n
ar
io
s
,
with
o
r
with
o
u
t c
o
n
tr
ast en
h
an
ce
m
e
n
t.
W
e
ca
ll
th
e
o
p
tio
n
with
o
u
t c
o
n
tr
ast en
h
an
ce
m
en
t o
p
tio
n
A
an
d
th
e
o
p
tio
n
with
co
n
tr
ast en
h
an
ce
m
en
t o
p
tio
n
B
.
T
h
u
s
,
o
p
tio
n
B
is
an
im
p
r
o
v
em
en
t o
v
er
o
p
tio
n
A.
3
.
1
.
Cho
o
s
ing
pa
ra
m
et
er
s
o
f
co
ntr
a
s
t
enha
ncem
ent
T
o
s
u
p
p
o
r
t
th
e
r
o
b
u
s
tn
ess
o
f
p
ar
am
eter
c
h
o
ice,
we
co
n
d
u
ct
s
o
m
e
p
r
elim
in
a
r
y
ex
p
e
r
im
en
ts
to
ev
alu
ate
s
o
m
e
p
a
r
am
eter
c
h
o
i
ce
s
.
Sp
ec
if
ically
,
with
r
ef
er
e
n
ce
to
th
e
s
tu
d
y
[
2
8
]
,
s
o
m
e
p
ar
am
eter
ch
o
ices
ar
e
co
n
d
u
cte
d
in
clu
d
i
n
g
(
clip
L
i
m
it
=
2
,
tileGr
id
Size
=
4
×4
)
,
(
c
lip
L
im
it
=
2
,
tileGr
id
Size
=
8
×
8
)
,
(
clip
L
im
it
=
4
,
tileGr
id
Size
=
1
6
×1
6
)
,
a
n
d
(
cli
p
L
im
it
=
1
0
,
tileGr
id
Size
=
3
2
×3
2
)
.
Fo
r
ea
c
h
p
ar
a
m
eter
ch
o
i
ce
,
we
will
co
n
d
u
ct
ex
p
er
im
en
ts
with
f
o
u
r
m
o
d
el
s
,
in
clu
d
in
g
Den
s
eNe
t1
2
1
,
D
en
s
eNe
t2
0
1
,
Mo
b
ileNetV2
,
an
d
E
f
f
icien
tNetB
0
.
Fo
r
ea
ch
p
ar
am
eter
s
elec
tio
n
ca
s
e,
we
av
er
ag
e
th
e
s
co
r
es
f
o
r
ac
cu
r
ac
y
,
AUC,
F1
-
s
co
r
e
,
p
r
ec
is
io
n
,
an
d
r
ec
al
l
ac
r
o
s
s
all
f
o
u
r
m
o
d
els.
Deta
ils
ar
e
s
h
o
wn
in
T
ab
le
2
.
T
ab
le
2
.
T
h
e
av
er
a
g
e
p
er
f
o
r
m
an
ce
o
f
p
ar
am
eter
c
h
o
ices
(
c
l
i
p
L
i
mi
t
,
t
i
l
e
G
r
i
d
S
i
z
e
)
A
c
c
u
r
a
c
y
AUC
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
(
2
,
4
×
4
)
0
.
8
8
2
7
1
6
0
.
9
6
1
8
9
0
.
8
9
6
4
6
9
0
.
8
7
5
0
.
8
8
1
5
0
5
(
2
,
8
×
8
)
0
.
9
1
3
5
8
0
.
9
7
6
0
6
7
0
.
9
2
5
2
1
7
0
.
9
0
.
9
1
0
9
2
3
(
4
,
1
6
×
1
6
)
0
.
8
7
6
5
4
3
0
.
9
3
4
7
5
6
0
.
8
8
0
2
9
5
0
.
8
6
8
7
5
0
.
8
7
4
1
6
9
(
1
0
,
3
2
×
3
2
)
0
.
8
7
6
5
4
3
0
.
9
4
7
5
6
1
0
.
8
7
1
9
5
1
0
.
8
8
1
2
5
0
.
8
7
6
0
0
5
Fro
m
T
ab
le
2
,
we
ca
n
s
ee
th
at
th
e
av
er
ag
e
s
co
r
es
o
f
th
e
ca
s
e
o
f
ch
o
o
s
in
g
p
ar
am
eter
s
clip
L
im
it
as
2
an
d
tileGr
id
Size
as
8
×
8
h
av
e
a
ce
r
tain
s
u
p
er
io
r
ity
.
Sp
ec
i
f
ically
,
th
e
a
v
er
ag
e
AUC
v
a
lu
e
o
f
th
e
ca
s
e
o
f
ch
o
o
s
in
g
p
ar
am
eter
s
clip
L
im
i
t
as
2
an
d
tileGr
id
Size
as
8
×8
r
ea
ch
es
th
e
h
ig
h
est
p
o
s
itio
n
with
a
v
alu
e
o
f
0
.
9
7
6
0
6
7
,
th
e
av
er
ag
e
AUC
v
alu
es
o
f
th
e
o
th
er
ca
s
es
a
r
e
all
s
lig
h
tly
lo
wer
wi
th
0
.
9
6
1
8
9
o
f
th
e
ca
s
e
clip
L
im
it
=
2
an
d
tileGr
id
Size
=
4
×4
,
0
.
9
3
4
7
5
6
o
f
t
h
e
ca
s
e
clip
L
im
it
=
4
an
d
tileGr
id
Size
=
1
6
×1
6
,
0
.
9
4
7
5
6
1
o
f
th
e
ca
s
e
clip
L
im
it
=
1
0
an
d
tileGr
id
Size
=
3
2
×3
2
.
C
o
n
s
id
er
in
g
an
o
th
e
r
m
etr
ic
ca
s
e,
th
e
av
er
ag
e
p
r
ec
is
io
n
v
alu
e
o
f
th
e
ca
s
e
with
p
ar
a
m
eter
s
elec
tio
n
clip
L
im
it
=
2
an
d
tileGr
id
Size
=
8
×8
r
ea
ch
ed
t
h
e
h
ig
h
est
p
o
s
itio
n
with
th
e
v
al
u
e
o
f
0
.
9
2
5
2
1
7
,
t
h
e
av
er
a
g
e
p
r
ec
is
io
n
v
alu
es
o
f
th
e
o
th
er
s
elec
tio
n
ca
s
es
we
r
e
all
s
lig
h
tly
lo
wer
with
0
.
8
9
6
4
6
9
o
f
th
e
ca
s
e
with
clip
L
im
it
=
2
an
d
tileGr
id
Size
=
4
×4
,
0
.
8
8
0
2
9
5
o
f
t
h
e
ca
s
e
with
clip
L
im
it
=
4
an
d
tileGr
id
Size
=
1
6
×1
6
,
0
.
8
7
1
9
5
1
o
f
th
e
ca
s
e
with
clip
L
im
it
=
1
0
an
d
tileGr
id
Siz
e
=
3
2
×3
2
.
T
h
u
s
,
b
ased
o
n
th
e
p
r
elim
in
a
r
y
test
in
g
to
ev
alu
ate
s
o
m
e
p
ar
a
m
eter
ch
o
ices,
th
e
ch
o
ice
with
clip
L
im
it
=
2
an
d
tileGr
id
Size
=
8
×8
ac
h
ie
v
ed
th
e
h
ig
h
est
test
s
co
r
e.
T
h
is
is
also
co
n
s
is
ten
t
with
th
e
p
ar
am
eter
s
elec
tio
n
co
n
s
id
er
atio
n
s
p
r
esen
ted
ab
o
v
e.
3
.
2
.
Q
ua
ntit
a
t
iv
e
c
o
m
pa
riso
n bet
wee
n t
ra
ini
ng
wit
h a
nd
wit
ho
ut
co
ntr
a
s
t
enha
ncem
ent
Fig
u
r
e
6
clea
r
ly
d
em
o
n
s
tr
at
es
th
at
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
b
r
in
g
s
co
n
s
is
ten
t
an
d
s
u
b
s
tan
tial
im
p
r
o
v
em
en
ts
th
r
o
u
g
h
all
k
ey
p
e
r
f
o
r
m
an
ce
m
etr
i
cs
co
m
p
ar
ed
to
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
.
Sp
ec
ially
,
th
e
av
er
a
g
e
ac
c
u
r
ac
y
in
c
r
ea
s
es
f
r
o
m
0
.
8
8
5
8
0
2
to
0
.
9
1
3
5
8
0
,
wh
ile
th
e
av
er
ag
e
AUC
im
p
r
o
v
es
f
r
o
m
0
.
9
5
0
7
6
2
to
0
.
9
7
6
0
6
7
,
p
o
in
tin
g
o
u
t
a
co
m
p
r
eh
e
n
s
iv
ely
en
h
an
ce
d
ca
p
ab
ilit
y
t
o
class
if
y
b
etwe
en
n
o
r
m
al
an
d
tu
b
er
cu
lo
s
is
d
ata.
Fu
r
th
e
r
m
o
r
e,
th
e
a
v
er
ag
e
p
r
ec
is
io
n
in
cr
ea
s
es
f
r
o
m
0
.
8
7
7
2
5
0
to
0
.
9
2
5
2
1
7
,
an
d
th
e
av
er
ag
e
r
ec
all
im
p
r
o
v
es
f
r
o
m
0
.
8
9
3
7
5
to
0
.
9
.
T
h
is
d
escr
ib
es
th
at
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
n
o
t
o
n
ly
r
ed
u
ce
s
f
alse
p
o
s
itiv
es
b
u
t
also
m
in
im
izes
f
alse
n
eg
ativ
es.
T
h
is
is
im
p
o
r
tan
t
b
ec
au
s
e
m
i
s
s
in
g
a
r
ea
l
ca
s
e
o
f
tu
b
er
cu
lo
s
is
ca
n
ca
u
s
e
s
er
io
u
s
p
u
b
lic
h
ea
lth
co
n
s
eq
u
en
ce
s
.
T
h
e
r
is
e
in
F1
-
s
co
r
e
f
r
o
m
0
.
8
8
5
3
3
2
to
0
.
9
1
0
9
2
3
h
elp
s
u
s
co
n
f
ir
m
th
at
m
o
d
els
tr
ain
ed
u
n
d
e
r
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
ca
n
ac
h
iev
e
a
m
o
r
e
b
alan
ce
d
p
er
f
o
r
m
an
ce
,
n
o
t
r
e
q
u
ir
in
g
a
t
r
ad
e
-
o
f
f
b
etwe
en
s
en
s
itiv
ity
a
n
d
s
p
ec
if
icity
.
I
t
also
r
e
f
lects
o
u
r
h
y
p
o
t
h
esis
f
o
r
o
u
r
p
r
o
p
o
s
ed
s
tu
d
y
.
W
e
co
n
tin
u
e
to
s
ee
th
e
av
er
a
g
e
p
e
r
f
o
r
m
an
ce
m
etr
ics
r
e
g
ar
d
i
n
g
t
r
u
e
p
o
s
itiv
e,
f
alse
n
eg
ativ
e,
tr
u
e
n
eg
ativ
e
,
an
d
f
a
ls
e
p
o
s
itiv
e
b
etwe
en
m
o
d
els tr
ain
ed
u
s
in
g
o
p
tio
n
s
A
an
d
B
in
Fig
u
r
e
7
.
I
n
Fig
u
r
e
7
,
th
e
av
er
a
g
e
p
er
f
o
r
m
an
ce
m
etr
ics
f
r
o
m
m
o
d
el
s
tr
ain
ed
u
s
in
g
o
p
tio
n
B
s
h
o
w
s
lig
h
tly
b
etter
p
er
f
o
r
m
an
ce
th
an
o
p
tio
n
A
ac
r
o
s
s
all
s
co
r
es,
in
clu
d
i
n
g
tr
u
e
p
o
s
itiv
e,
f
alse
n
eg
ativ
e
,
tr
u
e
n
e
g
ativ
e,
an
d
f
alse
p
o
s
itiv
e.
T
h
er
ef
o
r
e,
it
clea
r
ly
d
em
o
n
s
tr
ates
th
at
o
p
tio
n
B
b
r
i
n
g
s
m
o
r
e
ac
cu
r
ac
y
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
tu
b
er
c
u
lo
s
is
an
d
n
o
r
m
al
s
am
p
les.
Sp
ec
if
ically
,
th
e
av
er
ag
e
s
co
r
e
o
f
tr
u
e
p
o
s
itiv
e
r
ate
in
cr
ea
s
ed
s
lig
h
tly
f
r
o
m
4
4
.
1
%
to
4
4
.
4
%,
an
d
th
e
av
e
r
ag
e
s
co
r
e
o
f
f
alse
n
eg
ativ
e
r
ate
d
ec
r
ea
s
ed
f
r
o
m
5
.
2
%
to
4
.
9
%.
T
h
ese
s
co
r
es
in
d
icate
th
at
th
e
m
o
d
els
in
o
p
tio
n
B
wer
e
b
etter
at
d
etec
tin
g
r
ea
l
tu
b
er
cu
lo
s
is
ca
s
e
s
.
On
th
e
o
th
er
h
an
d
,
th
e
av
er
a
g
e
s
co
r
e
o
f
f
alse
p
o
s
itiv
e
r
ate
d
ec
r
ea
s
e
d
f
r
o
m
6
.
2
%
to
3
.
7
%,
wh
ile
t
h
e
av
er
a
g
e
s
co
r
e
o
f
tr
u
e
n
eg
ativ
e
r
ate
in
cr
ea
s
ed
f
r
o
m
4
4
.
4
%
to
4
6
.
9
%.
T
h
ese
s
co
r
es
in
d
icate
th
at
th
e
m
o
d
els
in
o
p
tio
n
B
wer
e
b
etter
at
d
is
tin
g
u
is
h
in
g
b
etw
ee
n
tu
b
er
cu
l
o
s
is
an
d
n
o
r
m
al
im
ag
es.
Alth
o
u
g
h
th
e
im
p
r
o
v
em
en
ts
ar
e
s
lig
h
t,
u
s
in
g
o
p
tio
n
B
is
g
en
er
ally
p
r
ef
er
r
ed
.
T
h
is
is
b
ec
au
s
e
it
s
h
o
ws
co
m
p
r
eh
e
n
s
iv
e
im
p
r
o
v
e
m
en
t,
r
ed
u
cin
g
b
o
th
f
alse a
lar
m
s
an
d
m
is
s
ed
d
etec
tio
n
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
.
3
,
J
u
n
e
2
0
2
6
:
2
5
6
7
-
2
5
8
0
2574
Fig
u
r
e
6
.
Av
e
r
ag
e
p
e
r
f
o
r
m
an
c
e
m
etr
ics b
etwe
en
m
o
d
els tr
ai
n
ed
u
s
in
g
o
p
tio
n
s
A
an
d
B
Fig
u
r
e
7
.
Av
e
r
ag
e
s
co
r
es o
f
tr
u
e
p
o
s
itiv
e,
f
alse n
eg
ativ
e
,
tr
u
e
n
eg
ativ
e,
a
n
d
f
alse p
o
s
itiv
e
b
etwe
en
m
o
d
els
tr
ain
ed
u
s
in
g
o
p
tio
n
s
A
an
d
B
3
.
3
.
Ana
ly
s
is
G
ra
d
-
CAM hea
t
m
a
ps
o
f
t
he
t
ra
ini
ng
s
t
ra
t
e
g
y
us
ing
o
ptio
n B
W
e
co
n
s
id
er
th
e
Gr
ad
-
C
AM
h
ea
tm
ap
s
o
f
s
o
m
e
co
r
r
ec
tly
r
e
co
g
n
ized
ca
s
es
b
y
d
ee
p
lea
r
n
i
n
g
m
o
d
els
tr
ain
ed
with
th
e
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
in
Fig
u
r
e
8
.
I
n
w
h
ich
,
th
e
f
ir
s
t
lin
e
b
elo
n
g
s
t
o
th
e
E
f
f
icien
tNetB
0
m
o
d
el,
th
e
s
ec
o
n
d
lin
e
b
elo
n
g
s
to
th
e
Den
s
eNe
t2
0
1
m
o
d
el,
th
e
th
ir
d
lin
e
b
elo
n
g
s
to
th
e
Den
s
eNe
t1
2
1
m
o
d
el,
an
d
th
e
f
o
u
r
th
lin
e
is
th
e
r
esu
l
ts
o
f
th
e
Mo
b
ileNetV2
m
o
d
el.
T
h
e
an
aly
s
is
o
f
th
e
o
b
tain
ed
Gr
ad
-
C
AM
s
h
o
ws
th
at
d
ee
p
lear
n
i
n
g
m
o
d
els
d
ia
g
n
o
s
in
g
th
ese
co
r
r
ec
tly
r
ec
o
g
n
ized
ca
s
es
o
f
ten
f
o
cu
s
o
n
t
h
e
lu
n
g
ar
ea
,
th
at
is
,
th
e
an
ato
m
ical
ar
ea
d
ir
ec
tly
r
elate
d
to
p
u
lm
o
n
ar
y
tu
b
er
cu
l
o
s
is
.
T
h
is
r
ef
lects
th
at
th
e
d
ee
p
lear
n
in
g
m
o
d
els
h
av
e
lear
n
ed
im
a
g
e
f
ea
tu
r
es
t
h
at
ar
e
r
elate
d
to
th
e
s
em
an
ti
cs
o
f
p
u
lm
o
n
a
r
y
tu
b
e
r
cu
lo
s
is
.
T
h
is
co
n
tr
ib
u
tes
to
s
tr
en
g
th
en
in
g
th
e
ex
p
lain
ab
ilit
y
an
d
clin
ical
p
lau
s
ib
ilit
y
o
f
th
e
m
o
d
el
an
d
s
u
p
p
o
r
tin
g
ex
p
er
t
s
in
ev
alu
atin
g
th
e
m
o
d
el
r
esu
lts
.
Fig
u
r
e
8
.
Gr
a
d
-
C
AM
v
is
u
aliza
tio
n
f
r
o
m
m
o
d
els tr
ain
ed
u
s
in
g
o
p
tio
n
B
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
Dia
g
n
o
s
in
g
tu
b
ercu
lo
s
is
fr
o
m
X
-
r
a
y
ima
g
in
g
b
a
s
ed
o
n
co
n
tr
a
s
t limited
a
d
a
p
tive
…
(
N
g
u
ye
n
Tr
o
n
g
V
in
h
)
2575
3
.
4
.
Q
ua
ntit
a
t
iv
e
c
o
m
pa
riso
n a
m
o
ng
m
o
dels
Nex
t,
we
wo
u
ld
s
ee
th
e
q
u
an
titativ
e
d
iag
n
o
s
tic
p
er
f
o
r
m
a
n
ce
m
etr
ics.
All
s
co
r
es we
r
e
as
s
es
s
ed
o
n
th
e
Pu
lm
o
n
ar
y
ch
est
X
-
r
ay
ab
n
o
r
m
alities
d
ataset,
a
lab
eled
t
u
b
er
cu
lo
s
is
d
ataset
b
u
ilt
b
ased
o
n
d
ata
f
r
o
m
th
e
Sh
en
zh
en
d
ataset
an
d
th
e
M
o
n
tg
o
m
er
y
d
ataset.
T
ab
le
3
s
h
o
ws
th
e
av
er
ag
e
p
e
r
f
o
r
m
an
ce
o
f
th
e
m
o
d
els
in
th
e
ex
p
er
im
en
t,
with
th
e
av
er
a
g
e
ac
cu
r
ac
y
s
co
r
e,
th
e
av
er
ag
e
AUC
s
co
r
e,
th
e
av
er
a
g
e
F1
-
s
co
r
e,
th
e
av
er
a
g
e
p
r
ec
is
io
n
s
co
r
e,
a
n
d
th
e
a
v
er
a
g
e
r
ec
all
s
co
r
e.
T
ab
le
3
s
h
o
ws
th
at
E
f
f
icien
t
NetB0
h
as
th
e
b
est
o
v
er
all
p
e
r
f
o
r
m
a
n
ce
.
E
f
f
icien
tNetB
0
r
ea
ch
es
s
o
m
e
h
ig
h
est
s
co
r
es.
T
h
e
av
er
a
g
e
ac
cu
r
ac
y
r
ea
ch
es
0
.
9
1
3
5
8
0
with
0
.
9
0
1
2
3
5
in
th
e
ca
s
e
with
o
u
t
C
L
AHE
-
b
ased
co
n
tr
ast
p
r
e
p
r
o
ce
s
s
in
g
a
n
d
0
.
9
2
5
9
2
6
in
th
e
ca
s
e
with
C
L
AHE
-
b
ased
co
n
tr
ast
p
r
e
p
r
o
ce
s
s
in
g
,
s
o
u
s
in
g
C
L
AHE
o
u
tp
er
f
o
r
m
s
0
.
0
2
4
6
9
1
.
T
h
e
a
v
er
ag
e
F1
-
s
co
r
e
r
ea
ch
es
0
.
9
1
3
4
1
5
with
0
.
9
i
n
th
e
ca
s
e
with
o
u
t
C
L
AHE
-
b
ased
co
n
tr
ast
p
r
ep
r
o
ce
s
s
in
g
an
d
0
.
9
2
6
8
2
9
in
th
e
ca
s
e
with
C
L
AHE
-
b
ased
co
n
tr
ast
p
r
ep
r
o
ce
s
s
in
g
,
s
o
u
s
in
g
C
L
AHE
o
u
tp
er
f
o
r
m
s
0
.
0
2
6
8
2
9
.
T
h
e
a
v
er
ag
e
r
ec
all
r
ea
ch
es
0
.
9
2
5
with
0
.
9
i
n
th
e
ca
s
e
with
o
u
t
C
L
AHE
-
b
ased
co
n
tr
ast
p
r
ep
r
o
ce
s
s
in
g
an
d
0
.
9
5
in
th
e
ca
s
e
with
C
L
AHE
-
b
ased
co
n
tr
as
t
p
r
ep
r
o
ce
s
s
in
g
,
s
o
u
s
in
g
C
L
AHE
o
u
tp
er
f
o
r
m
s
0
.
0
5
.
T
h
ey
in
d
icate
a
h
ig
h
er
ab
i
lity
to
co
r
r
ec
tly
d
iag
n
o
s
e
tu
b
er
cu
lo
s
is
ca
s
es
th
an
th
e
o
th
er
ar
ch
itectu
r
es
in
th
is
s
tu
d
y
.
Mo
b
ileNetV2
s
tan
d
s
o
u
t
with
th
e
h
ig
h
est
av
er
ag
e
AUC
o
f
0
.
9
7
3
1
7
1
an
d
th
e
h
ig
h
est
av
e
r
ag
e
p
r
ec
is
io
n
o
f
0
.
9
2
5
7
7
,
in
d
icatin
g
th
at
t
h
is
m
o
d
el
is
p
ar
ticu
la
r
ly
s
tr
o
n
g
in
r
ed
u
ci
n
g
f
alse
p
o
s
itiv
es.
T
h
e
AUC
in
cr
ea
s
es
f
r
o
m
0
.
9
6
4
0
2
4
to
0
.
9
8
2
3
1
7
,
an
d
t
h
e
p
r
ec
is
io
n
in
c
r
ea
s
es
f
r
o
m
0
.
8
8
0
9
5
2
to
0
.
9
7
0
5
8
8
with
C
L
AHE
-
b
ased
co
n
tr
ast
p
r
e
p
r
o
ce
s
s
in
g
.
Me
an
wh
ile,
Den
s
eNe
t2
0
1
h
as
th
e
l
o
west
p
er
f
o
r
m
a
n
ce
in
th
is
s
tu
d
y
,
in
d
icatin
g
a
lo
wer
class
if
icatio
n
ab
ilit
y
th
a
n
th
e
o
t
h
er
m
o
d
els,
an
d
th
is
m
ay
b
e
f
r
o
m
th
e
r
estricte
d
d
ataset.
Ov
er
all,
E
f
f
icien
tNetB
0
is
th
e
m
o
s
t p
r
o
m
in
en
t m
o
d
el
am
o
n
g
o
u
r
ex
p
er
i
m
en
ts
.
T
ab
le
3
.
T
h
e
av
er
a
g
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
els
M
o
d
e
l
A
c
c
u
r
a
c
y
AUC
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
Ef
f
i
c
i
e
n
t
N
e
t
B
0
0
.
9
1
3
5
8
0
.
9
6
7
3
7
8
0
.
9
0
2
3
8
1
0
.
9
2
5
0
.
9
1
3
4
1
5
D
e
n
seN
e
t
2
0
1
0
.
8
8
2
7
1
6
0
.
9
6
1
8
9
0
.
8
8
6
5
3
8
0
.
8
7
5
0
.
8
8
0
6
9
6
D
e
n
seN
e
t
1
2
1
0
.
9
0
1
2
3
5
0
.
9
5
1
2
2
0
.
8
9
0
2
4
4
0
.
9
1
2
5
0
.
9
0
1
2
3
5
M
o
b
i
l
e
N
e
t
V
2
0
.
9
0
1
2
3
5
0
.
9
7
3
1
7
1
0
.
9
2
5
7
7
0
.
8
7
5
0
.
8
9
7
1
6
5
E
s
s
en
tially
,
th
e
m
o
d
els
all
ac
h
iev
ed
f
air
ly
g
o
o
d
a
v
er
ag
e
s
co
r
es.
I
t
ca
n
b
e
p
a
r
tly
e
x
p
lain
ed
b
y
h
o
w
s
p
ec
if
ic
ar
ch
itectu
r
al
f
ea
tu
r
es
r
elate
to
tu
b
er
c
u
lo
s
is
im
ag
e
ch
ar
ac
ter
is
tics
.
E
f
f
icien
tNet,
with
its
co
m
p
o
u
n
d
s
ca
lin
g
s
tr
ateg
y
,
ca
n
h
elp
ex
p
l
o
it
s
u
b
tle
f
ea
tu
r
es
s
u
c
h
as
lesi
o
n
tex
tu
r
e
in
th
e
lu
n
g
r
eg
io
n
.
B
esid
es,
Den
s
eNe
t
u
tili
ze
s
f
ea
tu
r
e
r
eu
s
e
m
ec
h
an
i
s
m
s
th
r
o
u
g
h
d
e
n
s
e
co
n
n
ec
ti
o
n
s
,
wh
ich
m
ay
b
e
g
o
o
d
f
o
r
i
d
e
n
tify
in
g
t
h
e
lesi
o
n
s
'
ch
ar
ac
ter
is
tics
an
d
co
n
tr
ast d
is
tr
ib
u
tio
n
.
Alth
o
u
g
h
all
th
e
test
ed
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
m
o
d
els
p
er
f
o
r
m
e
d
well
in
d
etec
tin
g
p
u
lm
o
n
ar
y
tu
b
er
cu
lo
s
is
o
n
ch
est
X
-
r
ay
d
ata,
E
f
f
icien
tNetB
0
o
u
tp
e
r
f
o
r
m
e
d
s
o
m
e
o
th
er
m
o
d
els
in
th
e
r
esear
ch
ex
p
er
im
en
ts
.
T
h
is
d
if
f
er
en
ce
ca
n
b
e
p
ar
tly
ex
p
lain
ed
b
y
th
e
E
f
f
icien
tNet
ar
ch
itectu
r
e,
in
wh
ich
th
e
s
im
u
ltan
eo
u
s
b
alan
ce
b
etwe
e
n
d
ep
t
h
,
wid
th
,
an
d
r
eso
lu
tio
n
h
elp
s
th
e
m
o
d
el
lear
n
b
etter
.
On
th
e
o
t
h
er
h
an
d
,
d
ee
p
er
m
o
d
els
s
u
c
h
as
Den
s
e
Net2
0
1
m
a
y
n
o
t b
e
o
p
tim
al
b
e
ca
u
s
e
o
f
th
e
lar
g
e
n
u
m
b
er
o
f
p
ar
am
eter
s
,
an
d
h
e
r
e
th
e
test
d
ataset
is
n
o
t
lar
g
e
in
s
ize.
I
n
ad
d
itio
n
,
th
e
d
ata
s
et
u
s
ed
was
g
en
er
ated
f
r
o
m
two
d
ata
s
o
u
r
ce
s
,
Mo
n
tg
o
m
er
y
C
o
u
n
ty
a
n
d
S
h
en
zh
en
Ho
s
p
ital.
I
n
th
eo
r
y
,
th
is
also
m
ak
es
d
ata
p
r
e
p
r
o
ce
s
s
in
g
u
s
ef
u
l
in
im
p
r
o
v
in
g
th
e
d
if
f
er
e
n
t im
ag
e
q
u
ality
b
etwe
en
th
e
d
atasets
.
W
e
tak
e
a
clo
s
er
l
o
o
k
at
th
e
E
f
f
icien
tNetB
0
m
o
d
el
with
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
as
s
h
o
wn
in
Fig
u
r
e
9
,
with
th
e
co
n
f
u
s
io
n
m
atr
i
x
o
n
th
e
lef
t
s
id
e
(
Fig
u
r
e
9
(
a)
)
an
d
th
e
R
OC
cu
r
v
e
o
n
th
e
r
ig
h
t
s
id
e
(
Fig
u
r
e
9
(
b
)
)
.
T
h
e
c
o
n
f
u
s
io
n
m
atr
ix
p
o
in
ts
o
u
t
th
e
d
et
ailed
v
alu
es
with
th
e
tr
u
e
p
o
s
itiv
e
o
f
3
8
,
th
e
tr
u
e
n
eg
ativ
es
o
f
3
7
,
th
e
f
alse
p
o
s
itiv
es
o
f
4
,
an
d
th
e
f
alse
n
e
g
ativ
es
o
f
2
.
Fro
m
th
er
e,
m
a
n
y
m
etr
ics
s
u
ch
as
ac
cu
r
ac
y
,
F1
-
s
co
r
e,
p
r
ec
is
io
n
,
an
d
r
ec
all
ca
n
b
e
ca
lcu
lated
.
B
asical
ly
,
th
e
m
o
d
el
g
iv
es
h
ig
h
an
d
b
alan
ce
d
p
er
f
o
r
m
an
ce
.
O
n
th
e
o
n
e
h
an
d
,
th
e
o
v
er
all
class
if
icatio
n
ac
c
u
r
ac
y
is
q
u
ite
g
o
o
d
with
a
v
al
u
e
o
f
0
.
9
2
5
9
2
6
.
On
th
e
o
th
er
h
an
d
,
th
e
F1
-
s
co
r
e
o
f
0
.
9
2
6
8
2
9
s
h
o
ws
th
e
b
alan
ce
in
th
e
d
iag
n
o
s
tic
ab
ilit
y
o
f
th
e
m
o
d
el.
Mo
r
e
s
p
ec
if
ically
,
th
e
r
ec
all
s
co
r
e
o
f
0
.
9
5
s
h
o
ws a
lo
w
r
ate
o
f
m
is
s
in
g
ca
s
es th
at
ar
e
ac
tu
ally
tu
b
e
r
cu
lo
s
is
,
wh
ile
th
e
p
r
ec
is
io
n
s
co
r
e
o
f
0
.
9
0
4
7
6
2
s
h
o
ws th
e
ac
cu
r
ac
y
i
n
th
e
r
a
n
g
e
o
f
ca
s
es c
lass
if
ied
as tu
b
er
cu
lo
s
is
.
Fro
m
th
e
R
OC
cu
r
v
e
o
n
th
e
r
ig
h
t
s
id
e
s
h
o
wn
in
Fig
u
r
e
9
,
it
is
s
ee
n
th
at
th
e
m
o
d
el
ac
h
iev
ed
0
.
9
7
0
7
3
2
AUC
s
co
r
e.
T
h
is
m
ea
n
s
th
e
m
o
d
el
ac
h
iev
e
d
a
g
o
o
d
a
b
ilit
y
to
d
is
cr
im
in
ate
b
et
wee
n
ch
est
X
-
r
ay
s
with
an
d
with
o
u
t
tu
b
e
r
cu
lo
s
is
.
T
h
e
h
ig
h
R
OC
cu
r
v
e,
alo
n
g
with
th
e
AUC
v
alu
e
q
u
ite
clo
s
e
to
1
,
s
h
o
ws
a
g
o
o
d
p
er
f
o
r
m
a
n
ce
o
v
er
th
e
en
tire
r
an
g
e
o
f
class
if
icatio
n
th
r
esh
o
ld
s
.
T
h
is
ca
n
b
e
b
e
n
ef
icial
wh
en
,
i
n
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
,
it m
ay
b
e
n
ec
ess
ar
y
to
ad
ju
s
t th
e
th
r
esh
o
ld
s
.
Fro
m
th
e
ev
alu
atio
n
s
in
th
e
s
co
p
e
o
f
th
is
s
tu
d
y
,
th
e
E
f
f
icien
tNetB
0
m
o
d
el
with
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
is
a
g
o
o
d
ch
o
ice
in
f
u
r
th
e
r
r
esear
c
h
o
r
en
ter
in
g
th
e
im
p
lem
en
tatio
n
o
f
p
r
ac
tical
m
ed
ical
ap
p
licatio
n
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
.
3
,
J
u
n
e
2
0
2
6
:
2
5
6
7
-
2
5
8
0
2576
(
a)
(
b
)
Fig
u
r
e
9
.
E
f
f
icien
tNetB
0
m
o
d
el
with
th
e
tr
ain
in
g
s
tr
ateg
y
u
s
in
g
o
p
tio
n
B
f
o
r
(
a)
co
n
f
u
s
io
n
m
atr
ix
an
d
(
b
)
r
ec
eiv
e
r
o
p
er
atin
g
c
h
ar
ac
te
r
is
tic
(
R
O
C
)
cu
r
v
e
T
h
is
r
esu
lt
is
s
im
ilar
to
an
o
th
er
s
tu
d
y
t
h
at
also
u
s
ed
C
L
AHE
f
o
r
p
r
ep
r
o
ce
s
s
in
g
[
2
9
]
.
T
h
e
au
th
o
r
s
u
s
ed
E
f
f
icien
tNetV2
to
d
etec
t p
u
lm
o
n
a
r
y
tu
b
er
cu
l
o
s
is
in
a
s
im
u
lated
em
er
g
en
cy
d
ep
a
r
tm
e
n
t
en
v
ir
o
n
m
en
t
an
d
ac
h
iev
ed
an
AUC
o
f
0
.
8
7
8
.
W
h
en
ass
es
s
ed
o
n
ex
ter
n
al
d
a
ta,
th
e
s
y
s
tem
m
ain
tain
ed
s
tab
le
p
er
f
o
r
m
an
ce
with
AUC
o
f
0
.
8
3
8
o
n
th
e
Mo
n
t
g
o
m
er
y
s
et
an
d
0
.
8
0
6
o
n
th
e
Sh
en
zh
en
s
et.
C
o
m
p
ar
e
d
t
o
o
u
r
s
tu
d
y
,
b
o
th
d
em
o
n
s
tr
ate
th
e
p
o
ten
tial
o
f
ap
p
ly
in
g
im
ag
e
p
r
e
p
r
o
ce
s
s
in
g
to
th
e
p
e
r
f
o
r
m
an
ce
o
f
d
ee
p
lear
n
in
g
m
o
d
els
in
p
u
lm
o
n
a
r
y
tu
b
er
cu
lo
s
is
d
etec
t
io
n
,
an
d
h
er
e
it is
C
L
AHE
-
b
ased
p
r
ep
r
o
ce
s
s
in
g
.
B
esid
es
th
ese
r
e
s
u
lts
,
we
al
s
o
au
g
m
en
t
with
ad
d
itio
n
al
r
elev
an
t
r
ec
en
t
s
tu
d
ies
th
a
t
r
ep
o
r
ted
im
p
r
o
v
em
e
n
ts
with
C
L
AHE
.
B
esid
es
th
ese
r
esu
lts
,
we
also
au
g
m
en
t
with
ad
d
itio
n
al
r
ele
v
an
t
r
ec
en
t
s
tu
d
ies
th
at
r
ep
o
r
ted
im
p
r
o
v
em
en
ts
w
ith
C
L
AHE
.
I
n
s
tu
d
y
[
2
8
]
,
th
e
au
th
o
r
s
p
r
esen
te
d
th
eir
wo
r
k
o
n
ey
e
d
etec
tio
n
in
th
er
m
al
im
ag
es th
at
u
s
ed
C
L
A
HE
to
im
p
r
o
v
e
p
er
f
o
r
m
a
n
ce
.
Acc
o
r
d
in
g
to
th
e
p
a
p
er
,
wh
en
ex
p
er
im
en
tin
g
with
m
o
d
els,
th
e
a
u
th
o
r
s
r
ea
c
h
ed
a
n
av
er
a
g
e
o
f
p
r
ec
is
io
n
as
0
.
9
9
0
2
5
0
in
th
e
ca
s
e
o
f
r
aw
th
er
m
al
im
ag
es
o
f
h
u
m
an
f
ac
es.
B
esid
es,
th
e
au
th
o
r
s
r
e
ac
h
ed
an
av
er
a
g
e
o
f
p
r
ec
is
io
n
as
0
.
9
9
8
0
0
0
i
n
th
e
ca
s
e
o
f
C
L
AHE
-
en
h
an
ce
d
th
er
m
al
im
ag
es.
T
h
er
ef
o
r
e,
ap
p
ly
in
g
im
ag
e
en
h
an
ce
m
en
t
wi
th
C
L
AHE
im
p
r
o
v
ed
th
e
av
er
ag
e
p
r
ec
is
io
n
s
co
r
e
o
f
th
e
m
o
d
els
b
y
0
.
0
0
7
7
5
0
.
I
n
s
tu
d
y
[
3
0
]
,
th
e
a
u
th
o
r
s
ap
p
lie
d
im
ag
e
q
u
ality
en
h
an
ce
m
e
n
t w
ith
C
L
AHE
to
th
e
p
r
o
b
lem
o
f
b
r
ea
s
t c
an
ce
r
d
iag
n
o
s
is
f
r
o
m
m
am
m
o
g
r
am
im
ag
es.
T
h
e
au
th
o
r
s
u
s
ed
th
e
m
o
d
e
l b
ased
o
n
VGG
-
1
6
an
d
ex
p
er
im
en
ted
o
n
a
to
tal
o
f
1
,
4
5
9
m
am
m
o
g
r
am
s
.
W
h
e
n
test
in
g
in
th
e
ca
s
e
o
f
r
aw
im
ag
es
with
ar
tef
ac
t
r
em
o
v
al,
th
e
a
u
th
o
r
s
ac
h
iev
e
d
8
8
.
4
5
%
o
f
th
e
test
ac
cu
r
ac
y
an
d
8
8
.
6
2
%
o
f
t
h
e
F1
-
s
co
r
e.
W
h
en
ap
p
ly
in
g
C
L
AHE
o
n
ce
,
th
e
au
th
o
r
s
ac
h
iev
ed
9
4
.
6
3
%
o
f
th
e
test
ac
cu
r
ac
y
an
d
9
4
.
7
4
%
o
f
th
e
F
1
-
s
co
r
e,
an
d
wh
en
ap
p
ly
in
g
C
L
AHE
twice,
th
e
au
th
o
r
s
ac
h
iev
ed
9
5
.
2
1
%
o
f
th
e
test
ac
cu
r
ac
y
an
d
9
3
.
2
8
%
o
f
th
e
F1
-
s
co
r
e.
T
h
er
ef
o
r
e,
ap
p
ly
in
g
im
a
g
e
en
h
an
ce
m
en
t
with
C
L
AHE
im
p
r
o
v
ed
t
h
e
test
ac
cu
r
ac
y
b
y
a
m
ax
im
u
m
o
f
6
.
7
6
%
Evaluation Warning : The document was created with Spire.PDF for Python.