I
nte
rna
t
io
na
l J
o
urna
l o
f
Adv
a
nces in Applie
d Science
s
(
I
J
AAS)
Vo
l.
15
,
No
.
1
,
Ma
r
ch
20
26
,
p
p
.
1
5
5
~
1
6
7
I
SS
N:
2252
-
8
8
1
4
,
DOI
:
1
0
.
1
1
5
9
1
/ijaas
.
v15.
i
1
.
pp
155
-
1
6
7
155
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
a
s
.
ia
esco
r
e.
co
m
Co
mpa
ra
tive a
na
ly
sis
of YO
L
O
va
ria
nts and E
f
ficie
ntNet for
detec
ting bo
ne
fr
a
ctures in
X
-
ra
y
i
ma
g
es
Sh
a
t
a
bd
i Sa
rk
er
1
,
Av
izit
Ro
y
1
,
Sh
a
ila
Sh
a
rm
in
1
,
Sh
a
k
ila
Ra
hm
a
n
1
,
J
ia
Uddi
n
2
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
A
meri
c
a
n
I
n
t
e
r
n
a
t
i
o
n
a
l
U
n
i
v
e
r
si
t
y
-
B
a
n
g
l
a
d
e
sh
,
D
h
a
k
a
,
B
a
n
g
l
a
d
e
s
h
2
D
e
p
a
r
t
me
n
t
o
f
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
B
i
g
D
a
t
a
,
C
o
l
l
e
g
e
o
f
E
n
d
i
c
o
t
t
,
W
o
o
s
o
n
g
U
n
i
v
e
r
si
t
y
,
D
a
e
j
e
o
n
,
R
e
p
u
b
l
i
c
o
f
K
o
r
e
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
29
,
2
0
2
5
R
ev
is
ed
No
v
21
,
2
0
2
5
Acc
ep
ted
J
an
1
,
2
0
2
6
A
b
o
n
e
fra
c
tu
re
is
a
se
rio
u
s
m
e
d
ica
l
p
ro
b
lem
,
a
n
d
a
c
c
u
ra
te
a
n
d
p
ro
m
p
t
d
iag
n
o
sis
is
c
ru
c
ial
fo
r
o
p
ti
m
a
l
tr
e
a
tme
n
t.
Th
is
stu
d
y
h
i
g
h
li
g
h
ts
th
e
p
ro
g
re
ss
o
f
a
u
t
o
m
a
ti
c
b
o
n
e
fra
c
tu
re
d
e
tec
ti
o
n
u
si
n
g
d
e
e
p
lea
rn
in
g
(DL
)
m
o
d
e
ls.
A
d
a
tas
e
t
c
o
n
tain
i
n
g
1
7
d
iffere
n
t
fra
c
tu
re
c
las
se
s w
a
s u
se
d
to
train
a
n
d
e
v
a
lu
a
te
th
e
m
o
d
e
ls.
Th
e
d
a
tas
e
t
h
a
d
c
la
ss
imb
a
lan
c
e
a
n
d
m
in
o
r
fra
c
tu
re
d
e
tec
ti
o
n
c
h
a
ll
e
n
g
e
s.
Ex
te
n
siv
e
p
re
p
r
o
c
e
ss
in
g
,
in
c
l
u
d
i
n
g
d
a
ta
a
u
g
m
e
n
t
a
ti
o
n
a
n
d
re
siz
in
g
,
h
a
s
b
e
e
n
a
p
p
l
ied
to
so
lv
e
th
e
se
p
ro
b
lem
s,
wh
ich
h
a
s
h
e
lp
e
d
to
in
c
re
a
se
th
e
ro
b
u
stn
e
ss
o
f
t
h
e
m
o
d
e
l.
S
e
v
e
n
sta
te
-
of
-
t
h
e
-
a
rt
m
o
d
e
ls
—
y
o
u
o
n
l
y
lo
o
k
o
n
c
e
(
YO
LO
)
v
8
,
YO
LOv
9
,
YO
LOv
1
0
,
YO
LOv
1
1
,
Eff
icie
n
tNe
tB0
,
De
n
se
Ne
t1
6
9
a
n
d
Re
sN
e
t5
0
—
a
re
train
e
d
a
n
d
e
v
a
lu
a
ted
.
P
re
c
isio
n
,
re
c
a
ll
,
F
1
-
sc
o
re
,
a
n
d
m
e
a
n
a
v
e
ra
g
e
p
re
c
isio
n
(
m
AP)
we
re
u
se
d
to
e
v
a
lu
a
te
th
e
p
e
rfo
rm
a
n
c
e
o
f
th
e
m
o
d
e
ls.
Am
o
n
g
a
ll
m
o
d
e
ls,
YO
L
Ov
1
1
lea
d
s
th
e
o
th
e
rs
b
y
a
c
h
iev
i
n
g
t
h
e
h
ig
h
e
st
p
re
c
isio
n
,
m
AP,
a
n
d
p
re
c
is
io
n
-
re
c
a
ll
b
a
lan
c
e
.
YO
LOv
1
1
a
d
d
s
a
rc
h
i
tec
tu
ra
l
imp
ro
v
e
m
e
n
ts
su
c
h
a
s
a
d
e
e
p
b
a
c
k
b
o
n
e
n
e
two
rk
a
n
d
h
y
b
ri
d
f
e
a
tu
re
fu
si
o
n
,
wh
ich
m
a
k
e
th
e
m
o
d
e
l
m
o
r
e
re
li
a
b
le
in
d
iffere
n
t
ty
p
e
s
o
f
fra
c
t
u
re
d
e
tec
ti
o
n
.
It
is
c
a
p
a
b
le o
f
re
d
u
c
in
g
fa
lse
d
e
tec
ti
o
n
s
a
n
d
m
a
in
tain
i
n
g
sta
b
l
e
m
e
m
o
ry
u
sa
g
e
c
o
n
siste
n
c
y
e
v
e
n
u
n
d
e
r
d
iffere
n
t
ima
g
i
n
g
c
o
n
d
it
i
o
n
s.
O
v
e
ra
ll
,
YO
LOv
1
1
sh
o
we
d
p
r
o
m
isin
g
re
su
l
ts
a
n
d
h
i
g
h
l
ig
h
ted
th
e
p
o
ten
ti
a
l
o
f
AI
-
p
o
we
re
d
d
iag
n
o
stic
t
o
o
ls
t
o
imp
ro
v
e
c
li
n
ica
l
p
r
o
c
e
ss
e
s
a
n
d
p
a
ti
e
n
t
c
a
r
e
.
As
fu
t
u
re
wo
r
k
,
th
e
a
p
p
li
c
a
ti
o
n
field
o
f
th
e
m
o
d
e
l
c
a
n
b
e
e
x
ten
d
e
d
to
la
rg
e
r
m
e
d
ica
l
ima
g
i
n
g
tas
k
s
,
a
n
d
it
c
a
n
b
e
fu
rth
e
r
re
fi
n
e
d
f
o
r
e
ff
e
c
ti
v
e
u
se
i
n
re
so
u
rc
e
-
li
m
it
e
d
e
n
v
iro
n
m
e
n
ts.
K
ey
w
o
r
d
s
:
B
o
n
e
f
r
ac
tu
r
e
C
o
m
p
u
ter
v
is
io
n
E
f
f
icien
t
N
et
X
-
r
ay
im
ag
e
YOL
Ov
1
1
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
:
J
ia
Ud
d
in
Dep
ar
tm
en
t o
f
Ar
tific
ial
I
n
tellig
en
ce
an
d
B
ig
Data
,
C
o
lleg
e
o
f
E
n
d
ic
o
tt,
W
o
o
s
o
n
g
U
n
iv
er
s
ity
Dae
jeo
n
,
3
4
6
0
6
,
R
ep
u
b
lic
o
f
Ko
r
ea
E
m
ail:
jia.
u
d
d
i
n
@
wsu
.
ac
.
k
r
1.
I
NT
RO
D
UCT
I
O
N
E
v
er
y
d
ay
,
co
u
n
tles
s
p
eo
p
le
s
u
f
f
er
f
r
o
m
b
o
n
e
f
r
ac
tu
r
es,
wh
ich
ar
e
a
co
m
m
o
n
b
u
t
s
er
io
u
s
h
ea
lth
p
r
o
b
lem
.
Acc
u
r
ate
f
r
ac
tu
r
e
d
i
ag
n
o
s
is
at
th
e
r
ig
h
t
tim
e
is
cr
itical,
as
in
co
r
r
ec
t
o
r
late
d
ia
g
n
o
s
is
ca
n
lead
t
o
lo
n
g
-
ter
m
m
o
b
ilit
y
p
r
o
b
lem
s
o
r
p
o
o
r
r
ec
o
v
e
r
y
.
C
u
r
r
en
tly
,
X
-
r
ay
is
th
e
m
o
s
t
co
m
m
o
n
m
eth
o
d
o
f
f
r
ac
tu
r
e
d
etec
tio
n
.
Ho
wev
er
,
it
h
as
s
ev
er
al
lim
itatio
n
s
.
Fra
ctu
r
e
a
p
p
ea
r
an
ce
ca
n
v
ar
y
f
r
o
m
p
at
ien
t
to
p
atien
t
;
th
e
d
iag
n
o
s
tic
v
alu
e
d
ep
en
d
s
en
tir
ely
o
n
th
e
s
k
ill
o
f
th
e
r
ad
io
lo
g
is
t,
an
d
m
an
y
tim
es
it
m
ay
f
ail
to
d
etec
t
s
m
all
o
r
s
u
b
tle
h
air
lin
e
f
r
ac
tu
r
es.
T
h
is
is
wh
er
e
ar
tific
ial
in
tellig
en
ce
(
AI
)
p
lay
s
an
im
p
o
r
ta
n
t
r
o
le.
T
h
e
y
o
u
o
n
ly
l
o
o
k
o
n
ce
(
YOL
O)
f
am
ily
o
f
d
ee
p
lear
n
in
g
(
DL
)
-
b
ased
o
b
ject
d
etec
tio
n
m
eth
o
d
s
is
p
ar
ticu
lar
ly
p
o
p
u
lar
,
as
it
i
s
ab
le
to
d
etec
t
f
r
ac
tu
r
es
with
h
ig
h
ac
cu
r
ac
y
in
r
ea
l
-
tim
e
[
1
]
.
C
h
an
n
el
-
wis
e
f
u
s
io
n
a
n
d
s
p
atial
-
wis
e
g
r
o
u
p
atten
tio
n
(
C
FS
G
)
U
-
Ne
t
-
l
ik
e
m
o
d
els
h
a
v
e
f
u
r
t
h
e
r
im
p
r
o
v
e
d
th
e
a
cc
u
r
a
cy
o
f
r
i
b
f
r
a
ct
u
r
e
d
e
tect
io
n
in
c
o
m
p
u
t
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
5
,
No
.
1
,
Ma
r
c
h
2
0
2
6
:
155
-
1
6
7
156
to
m
o
g
r
a
p
h
y
(
CT
)
s
ca
n
s
b
y
a
d
d
i
n
g
ch
a
n
n
el
-
b
ase
d
an
d
s
p
at
ial
a
tte
n
t
io
n
p
r
o
ce
s
s
es
[
2]
.
C
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
-
b
ased
m
eth
o
d
s
h
av
e
s
h
o
wn
o
u
ts
tan
d
in
g
p
er
f
o
r
m
a
n
ce
in
m
ed
ical
im
ag
e
an
aly
s
is
in
ter
m
s
o
f
class
if
icatio
n
,
s
eg
m
en
tatio
n
,
an
d
ab
n
o
r
m
ality
d
etec
tio
n
[
3
]
.
Ad
v
an
ce
d
co
m
p
u
ter
-
ai
d
e
d
d
iag
n
o
s
is
(
C
AD)
s
y
s
tem
s
h
av
e
m
ad
e
f
r
ac
tu
r
e
d
etec
tio
n
m
o
r
e
ef
f
icie
n
t
b
y
r
e
d
u
cin
g
t
h
e
wo
r
k
lo
ad
o
f
r
ad
i
o
lo
g
is
ts
an
d
r
ed
u
cin
g
h
u
m
an
e
r
r
o
r
[
4
]
.
Ho
wev
er
,
d
esp
ite
s
ig
n
if
ican
t
p
r
o
g
r
ess
,
s
ev
er
al
ch
allen
g
es
r
em
ain
-
e
s
p
ec
ially
en
s
u
r
in
g
co
n
s
is
ten
tly
r
eliab
l
e
p
er
f
o
r
m
a
n
ce
o
f
AI
m
o
d
els
ac
r
o
s
s
p
atie
n
t
h
eter
o
g
en
eity
a
n
d
c
h
an
g
in
g
im
ag
in
g
m
o
d
alities
is
a
m
ajo
r
p
r
o
b
lem
[
5
]
.
Alth
o
u
g
h
th
e
u
s
e
o
f
AI
in
m
ed
ical
im
ag
in
g
h
as
g
r
o
wn
r
ap
id
ly
,
s
ev
er
al
c
h
allen
g
es
r
em
ai
n
in
u
s
in
g
C
AD
s
y
s
tem
s
f
o
r
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
.
On
e
o
f
th
ese
is
th
e
u
n
e
v
en
q
u
ality
o
f
th
e
d
at
aset.
Var
iatio
n
s
in
X
-
r
ay
im
ag
e
clar
ity
,
a
n
n
o
tati
o
n
ac
cu
r
ac
y
–
th
ese
f
ac
to
r
s
af
f
ec
t
m
o
d
el
p
er
f
o
r
m
a
n
ce
[
6
]
.
I
n
p
ar
ticu
lar
,
p
o
o
r
q
u
ality
X
-
r
ay
im
a
g
es p
r
ev
e
n
t
ef
f
ec
tiv
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
,
r
e
s
u
ltin
g
in
lo
s
s
o
f
in
f
o
r
m
atio
n
,
an
d
s
u
b
tle
f
r
ac
tu
r
es
ar
e
o
f
ten
m
is
class
if
ied
[
7
]
.
An
o
th
er
m
ajo
r
p
r
o
b
lem
is
class
i
m
b
alan
ce
in
f
r
ac
tu
r
e
d
atasets
.
I
n
th
is
,
th
e
m
o
d
els
p
er
f
o
r
m
well
i
n
d
etec
tin
g
c
o
m
m
o
n
o
r
f
r
e
q
u
en
t
f
r
ac
tu
r
es,
b
u
t
s
h
o
w
wea
k
n
ess
in
d
etec
ti
n
g
r
ar
e
o
r
co
m
p
le
x
f
r
ac
tu
r
es
[
8
]
.
L
im
itatio
n
s
in
g
e
n
er
aliza
tio
n
ab
ilit
y
ar
e
also
m
ajo
r
ch
allen
g
es,
as m
o
s
t A
I
m
o
d
els ar
e
tr
ain
ed
o
n
lim
ited
,
s
im
ilar
d
ataset
s
,
r
ed
u
cin
g
th
eir
ef
f
ec
tiv
en
ess
in
r
ea
l
-
life
,
d
iv
er
s
e
m
ed
ical
im
ag
es
[
9
]
.
Ma
n
y
AI
-
b
ased
f
r
ac
tu
r
e
d
etec
tio
n
m
o
d
els
lack
tr
an
s
p
ar
e
n
cy
,
m
a
k
in
g
it
d
if
f
icu
lt
f
o
r
m
ed
ical
p
r
o
f
ess
io
n
a
ls
to
in
ter
p
r
et
an
d
tr
u
s
t
th
e
m
o
d
el
o
u
tp
u
t
[
1
0
]
.
T
o
s
o
lv
e
th
is
p
r
o
b
lem
,
r
esear
c
h
er
s
ar
e
tr
y
in
g
to
in
ter
p
r
et
h
o
w
AI
d
ec
is
io
n
s
ar
e
b
ein
g
m
a
d
e
u
s
in
g
atten
tio
n
-
b
ased
v
is
u
aliza
tio
n
m
eth
o
d
s
s
u
ch
as
Gr
ad
-
C
AM
[
1
1
]
.
Yet
AI
m
o
d
els
th
at
ca
n
p
r
o
p
er
l
y
a
d
ap
t
t
o
d
if
f
er
e
n
t
ty
p
es
o
f
f
r
ac
tu
r
es,
d
if
f
e
r
en
t
im
ag
i
n
g
m
o
d
alities
,
an
d
d
iv
er
s
e
p
atien
t
g
r
o
u
p
s
ar
e
s
till
a
m
ajo
r
ch
allen
g
e
[
1
2
]
.
I
n
ad
d
itio
n
,
co
m
p
u
tatio
n
al
c
o
s
t,
h
a
r
d
war
e
lim
itati
o
n
s
,
an
d
im
p
le
m
en
tatio
n
is
s
u
es
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
en
v
ir
o
n
m
en
ts
ar
e
h
in
d
e
r
in
g
th
e
wid
esp
r
ea
d
ad
o
p
tio
n
o
f
AI
tech
n
o
l
o
g
ies
[
1
3
]
.
T
h
is
s
tu
d
y
p
r
esen
ted
s
ev
e
r
al
im
p
o
r
tan
t in
n
o
v
atio
n
s
to
ad
d
r
e
s
s
th
e
id
en
tifie
d
ch
allen
g
es:
‒
First,
a
cu
r
ated
X
-
r
ay
d
ataset
c
o
n
tain
in
g
1
7
f
r
ac
t
u
r
e
ty
p
es is
u
s
ed
[
1
4
]
.
T
h
e
d
ataset
co
n
tain
s
well
-
lab
eled
im
ag
es
in
J
PG,
PNG,
an
d
W
E
B
P
f
o
r
m
ats,
co
r
r
ec
tly
class
if
ied
ac
co
r
d
i
n
g
to
f
r
ac
tu
r
e
ty
p
e.
I
t
is
cr
ea
ted
u
s
in
g
p
u
b
licly
av
ailab
le
m
e
d
ical
im
ag
es c
o
llected
f
r
o
m
v
ar
io
u
s
o
p
en
s
o
u
r
ce
s
.
‒
Seco
n
d
,
th
e
s
tu
d
y
ap
p
lied
ad
v
an
ce
d
p
r
e
-
p
r
o
ce
s
s
in
g
m
eth
o
d
s
–
s
u
ch
as
d
ata
au
g
m
en
tatio
n
an
d
s
tan
d
ar
d
izatio
n
–
to
im
p
r
o
v
e
d
a
taset q
u
ality
,
in
cr
ea
s
e
m
o
d
el
r
o
b
u
s
tn
ess
,
an
d
r
e
d
u
ce
X
-
r
a
y
i
m
ag
e
n
o
is
e.
‒
T
h
ir
d
,
a
c
o
m
p
ar
is
o
n
b
etw
ee
n
cu
ttin
g
-
e
d
g
e
DL
m
o
d
els
,
in
clu
d
in
g
Den
s
eNe
t1
6
9
,
R
esNet5
0
,
E
f
f
icien
tNetB
0
,
YOL
Ov
8
,
Y
OL
Ov
9
,
YOL
Ov
1
0
,
an
d
YOL
Ov
1
1
,
was c
o
n
d
u
cte
d.
‒
Fin
ally
,
a
co
m
p
ar
is
o
n
o
f
v
a
r
io
u
s
p
o
p
u
lar
m
o
d
els
s
h
o
wed
th
a
t
YOL
Ov
1
1
g
av
e
th
e
b
est
r
esu
lts
.
Hen
ce
,
it
is
ch
o
s
en
as
th
e
m
ain
m
o
d
el
o
f
th
is
s
tu
d
y
.
YOL
Ov
1
1
p
r
o
v
e
d
to
b
e
m
o
r
e
ef
f
ec
tiv
e
t
h
an
o
t
h
er
m
o
d
els
in
ter
m
s
o
f
d
etec
tio
n
ac
cu
r
ac
y
,
f
ast wo
r
k
in
g
ab
ilit
y
,
a
n
d
o
v
er
al
l stab
ilit
y
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
DL
an
d
AI
h
a
v
e
m
ad
e
s
ig
n
if
ican
t
im
p
r
o
v
em
en
ts
in
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
in
m
ed
ical
i
m
ag
in
g
in
r
ec
en
t
y
ea
r
s
.
E
a
r
lier
X
-
r
ay
a
n
aly
s
is
d
ep
en
d
ed
en
tire
ly
o
n
th
e
ex
p
er
ien
ce
o
f
r
ad
io
lo
g
is
ts
.
T
h
is
wo
u
ld
h
av
e
allo
wed
f
o
r
h
u
m
an
er
r
o
r
s
,
d
is
cr
ep
an
cies,
an
d
d
elay
s
in
d
iag
n
o
s
is
.
B
u
t
n
o
w
it
is
b
ec
o
m
in
g
p
o
s
s
ib
le
to
d
etec
t
f
r
ac
tu
r
es
m
u
c
h
f
aster
an
d
m
o
r
e
ac
cu
r
ately
u
s
in
g
C
NN,
m
ac
h
in
e
lear
n
in
g
m
eth
o
d
s
,
a
n
d
o
b
ject
d
etec
tio
n
m
o
d
els
lik
e
YOL
O
o
r
E
f
f
icien
tNet.
E
ar
ly
s
tu
d
ies
h
a
v
e
s
h
o
wn
th
at
u
s
in
g
m
ac
h
in
e
lear
n
i
n
g
m
et
h
o
d
s
s
u
ch
as
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
,
d
ec
is
io
n
tr
ee
,
an
d
n
aiv
e
B
ay
es
h
a
s
y
ie
ld
ed
ab
o
u
t
6
4
%
to
9
2
%
a
cc
u
r
ac
y
i
n
f
r
ac
tu
r
e
class
if
icatio
n
[
1
5
]
.
T
h
ese
m
eth
o
d
s
wer
e
h
elp
f
u
l,
b
u
t
th
ey
wer
e
n
o
t
s
ca
lab
le
a
n
d
m
ain
ly
r
elied
o
n
h
an
d
c
r
af
te
d
f
ea
tu
r
es.
T
h
e
d
ev
elo
p
m
e
n
t
o
f
d
ee
p
C
NNs
g
r
ea
tly
in
cr
ea
s
ed
t
h
e
ac
cu
r
ac
y
o
f
d
etec
tio
n
.
Nar
r
ativ
e
r
ev
iews
co
n
f
ir
m
ed
th
at
C
NN
ar
ch
itectu
r
es su
ch
a
s
R
es
Net,
I
n
ce
p
tio
n
V3
,
an
d
F
aster
R
-
C
N
N
co
n
s
is
ten
tly
s
u
r
p
ass
ed
co
n
v
en
tio
n
al
d
iag
n
o
s
tic
tech
n
iq
u
es
[
1
6
]
.
Fu
r
th
er
,
C
NN
-
b
ased
m
o
d
els
s
u
ch
as
R
esNet
-
1
8
h
av
e
ac
h
iev
ed
8
9
.
8
%
v
alid
atio
n
ac
cu
r
ac
y
with
tr
a
n
s
f
er
lear
n
in
g
in
MA
T
L
AB
o
n
X
-
r
ay
im
a
g
e
class
if
icatio
n
in
to
f
r
ac
t
u
r
e
d
o
r
n
o
n
-
f
r
ac
tu
r
ed
ca
teg
o
r
ies,
alth
o
u
g
h
th
ey
d
i
d
n
o
t
p
e
r
f
o
r
m
well
i
n
m
u
lti
-
class
f
r
ac
tu
r
e
d
etec
tio
n
an
d
s
m
all
f
r
ac
tu
r
e
lo
ca
lizatio
n
p
r
o
b
lem
s
[
1
7
]
.
Als
o
,
p
ar
allel
Den
s
eNe
t
s
h
o
wed
test
ac
cu
r
a
cy
u
p
to
7
4
%
f
o
r
an
o
m
al
y
d
e
tectio
n
o
f
th
e
wr
is
t
an
d
f
o
r
ea
r
m
.
H
o
wev
er
,
th
e
p
er
f
o
r
m
an
ce
d
ec
r
ea
s
es f
o
r
c
o
m
p
l
ex
f
r
ac
tu
r
e
p
atter
n
s
[
1
8
]
.
E
f
f
icien
tNet
ar
ch
itectu
r
e
h
as
s
h
o
wn
n
o
tab
le
p
r
o
m
is
e
in
f
r
ac
tu
r
e
d
etec
tio
n
task
s
.
Fo
r
ex
am
p
le,
E
f
f
icien
tNet
-
B
4
h
as
b
ee
n
ap
p
lied
f
o
r
v
er
teb
r
al
f
r
ac
tu
r
e
a
n
d
o
s
teo
p
o
r
o
s
is
class
if
icatio
n
f
r
o
m
later
al
s
p
in
e
X
-
r
ay
s
,
ac
h
iev
i
n
g
a
r
ea
u
n
d
er
th
e
r
ec
eiv
er
o
p
er
atin
g
ch
ar
ac
t
er
is
tic
cu
r
v
e
(
AUROC
)
v
alu
e
s
o
f
0
.
9
3
an
d
0
.
8
5
,
o
u
tp
er
f
o
r
m
s
tan
d
ar
d
clin
ical
m
o
d
els,
an
d
s
p
in
al
co
r
d
f
r
ac
tu
r
e
d
etec
tio
n
u
s
in
g
E
f
f
icie
n
tNet
-
B
4
d
em
o
n
s
tr
ated
s
u
p
er
io
r
AUROC
p
er
f
o
r
m
an
ce
f
o
r
b
o
th
f
r
ac
tu
r
e
an
d
o
s
teo
p
o
r
o
s
is
id
en
tific
atio
n
[
1
9
]
.
R
ev
iews
h
av
e
also
e
m
p
h
asized
th
e
r
o
le
o
f
C
NNs,
U
-
Net
ar
ch
itectu
r
e,
an
d
tr
an
s
f
er
lea
r
n
in
g
in
i
m
p
r
o
v
i
n
g
d
iag
n
o
s
tic
ef
f
icien
cy
[
2
0
]
.
R
ec
en
t
wo
r
k
h
as
ex
p
lo
r
ed
atten
tio
n
-
en
h
an
ce
d
f
r
am
ewo
r
k
s
.
An
atten
tio
n
-
b
ased
ca
s
ca
d
e
R
-
C
NN
m
o
d
el,
f
o
r
in
s
tan
ce
,
a
ch
iev
ed
a
m
ea
n
av
er
a
g
e
p
r
ec
i
s
io
n
(
m
AP)
o
f
0
.
7
1
i
n
s
ter
n
u
m
f
r
ac
tu
r
e
d
etec
tio
n
,
im
p
r
o
v
in
g
s
en
s
itiv
ity
f
o
r
s
m
al
l
an
d
c
o
n
ce
aled
f
r
ac
t
u
r
es
[
2
1
]
.
B
esid
es,
b
etter
r
esu
lts
wer
e
o
b
tain
ed
b
y
ad
d
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
C
o
mp
a
r
a
tive
a
n
a
lysi
s
o
f YOL
O
va
r
ia
n
ts
a
n
d
E
fficien
tN
et
fo
r
d
etec
tin
g
b
o
n
e
fr
a
ctu
r
es
…
(
S
h
a
ta
b
d
i S
a
r
ke
r
)
157
atten
tio
n
tech
n
o
l
o
g
y
to
th
e
c
u
s
to
m
ized
C
NN
an
d
YOL
O
m
o
d
els.
Fo
r
ex
a
m
p
le,
th
e
Y
OL
Ov
7
m
o
d
el
with
f
o
cu
s
o
n
th
e
Fra
cAtlas
d
ataset
ac
h
iev
ed
8
6
.
2
%
m
AP,
alth
o
u
g
h
v
er
y
f
in
e
f
r
ac
tu
r
es
wer
e
s
till
d
if
f
icu
lt
to
d
etec
t
[
2
2
]
.
C
u
r
r
en
tly
,
t
h
e
m
o
s
t
ad
v
a
n
c
ed
m
eth
o
d
s
f
o
r
r
ea
l
-
tim
e
f
r
ac
tu
r
e
d
etec
tio
n
a
r
e
o
b
ject
d
etec
tio
n
f
r
am
ewo
r
k
s
s
u
ch
as
YOL
O.
A
s
y
s
tem
atic
r
ev
iew
s
h
o
we
d
t
h
at
th
e
YOL
O
-
b
ased
m
o
d
el
a
ch
iev
ed
an
ac
c
u
r
ac
y
o
f
u
p
to
9
9
%
in
d
is
tal
r
ad
iu
s
f
r
ac
tu
r
e
d
etec
tio
n
[
2
3
]
.
I
n
ad
d
itio
n
,
u
s
in
g
YOL
O
in
an
AI
-
ass
is
ted
r
ad
io
lo
g
y
s
y
s
tem
r
ed
u
ce
d
r
ep
o
r
t
g
e
n
er
atio
n
tim
e
b
y
an
a
v
er
ag
e
o
f
2
7
%,
d
em
o
n
s
tr
atin
g
its
ef
f
ec
tiv
en
ess
in
clin
ical
wo
r
k
f
lo
ws
[
2
4
]
.
YOL
Ov
5
p
e
r
f
o
r
m
ed
well
in
ce
r
v
ical
s
p
in
e
f
r
ac
tu
r
e
d
etec
tio
n
o
n
th
e
R
SNA
2
0
2
2
C
T
d
ataset
d
esp
ite
it
s
m
ild
wea
k
n
ess
.
I
t
a
ch
iev
ed
an
o
v
er
all
ac
cu
r
ac
y
o
f
9
4
%
an
d
an
AP
o
f
0
.
9
8
in
n
o
r
m
al
ca
s
es
an
d
an
AP
o
f
0
.
9
6
in
f
r
ac
tu
r
e
ca
s
es,
alth
o
u
g
h
r
ec
all
was
lo
wer
in
s
m
all
f
r
ac
tu
r
es
[
2
5
]
.
YOL
Ov
7
g
av
e
h
ig
h
s
p
ee
d
an
d
g
o
o
d
m
AP
s
co
r
es
in
wh
o
le
-
b
o
d
y
f
r
ac
t
u
r
e
d
etec
tio
n
,
b
u
t
p
er
f
o
r
m
ed
p
o
o
r
ly
in
lo
w
-
co
n
tr
ast
X
-
r
ay
[
2
6
]
.
On
th
e
o
th
er
h
a
n
d
,
YOL
Ov
8
p
r
o
v
id
ed
h
ig
h
p
r
ec
is
io
n
an
d
r
ec
all
in
r
ea
l
-
tim
e
d
etec
tio
n
o
n
m
u
ltimo
d
al
im
ag
es
.
Ho
wev
er
,
it
r
e
q
u
ir
es
m
o
r
e
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
an
d
a
s
lig
h
t
d
ec
lin
e
in
m
AP@
0
.
9
5
s
co
r
es
was
o
b
s
er
v
ed
[
2
7
]
.
I
n
ad
d
itio
n
to
r
a
d
io
g
r
a
p
h
s
,
s
o
m
e
n
ew
n
o
n
-
i
n
v
asiv
e
m
et
h
o
d
s
f
o
r
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
h
av
e
also
b
ee
n
p
r
o
p
o
s
ed
.
Fo
r
e
x
am
p
le,
a
m
icr
o
wav
e
im
ag
in
g
s
y
s
tem
is
ca
p
ab
le
o
f
d
etec
tin
g
f
in
e
f
r
ac
tu
r
es
d
o
wn
to
1
m
m
[
2
8
]
.
Similar
ly
,
in
cr
ea
s
ed
g
ain
an
d
ef
f
icien
cy
h
av
e
b
ee
n
ac
h
iev
ed
u
s
in
g
m
etam
ate
r
ial
-
b
ased
an
ten
n
as
f
o
r
ea
r
ly
f
r
ac
tu
r
e
d
etec
tio
n
[
2
9
]
.
Als
o
,
h
ig
h
-
r
eso
lu
tio
n
m
i
cr
o
wav
e
tr
a
n
s
ce
iv
er
s
h
av
e
b
ee
n
ab
le
t
o
d
etec
t
f
r
ac
tu
r
es
with
s
u
b
-
m
illi
m
eter
ac
cu
r
ac
y
[
3
0
]
.
T
h
ese
m
eth
o
d
s
s
h
o
w
th
at
f
r
ac
tu
r
e
d
etec
tio
n
tech
n
o
l
o
g
y
is
n
o
t
l
im
ited
to
co
n
v
en
tio
n
al
im
ag
in
g
.
T
h
e
ef
f
ec
tiv
en
ess
o
f
AI
in
r
a
d
io
lo
g
y
h
as
b
ee
n
h
ig
h
lig
h
ted
in
s
ev
er
al
r
ev
iews.
As
h
as
b
ee
n
s
h
o
wn
,
AI
o
f
ten
p
er
f
o
r
m
s
b
etter
th
a
n
p
h
y
s
ician
s
’
r
esu
lts
in
d
etec
tin
g
h
ip
f
r
ac
tu
r
es
[
3
1
]
.
Acc
o
r
d
in
g
to
a
m
eta
-
an
al
y
s
is
o
f
4
2
s
tu
d
ies,
th
e
d
iag
n
o
s
tic
s
en
s
itiv
ity
o
f
AI
is
c
o
m
p
ar
ab
le
to
th
at
o
f
r
ad
io
lo
g
is
ts
—
9
2
%
in
in
ter
n
al
v
alid
ity
an
d
9
1
%
in
ex
ter
n
a
l
v
ali
d
ity
[
3
2
]
.
I
n
a
d
d
itio
n
,
C
NN
-
b
ased
m
o
d
els
h
av
e
s
h
o
wn
g
r
ea
ter
th
an
9
0
%
ac
cu
r
ac
y
in
s
k
eleta
l
f
r
ac
tu
r
e
d
etec
tio
n
an
d
r
e
d
u
ce
d
in
ter
o
b
s
er
v
e
r
v
ar
iab
ilit
y
[
3
3
]
.
So
m
e
s
tu
d
ies
h
av
e
s
h
o
wn
th
at
in
ter
p
r
etab
le
AI
to
o
ls
s
u
ch
as
Gr
ad
-
C
AM
ar
e
h
elp
f
u
l
in
in
cr
ea
s
in
g
clin
ician
co
n
f
i
d
en
c
e
[
3
4
]
.
Ho
we
v
er
,
i
n
co
n
s
is
ten
t
d
atasets
,
lack
o
f
s
tan
d
ar
d
ize
d
ass
ess
m
en
ts
,
an
d
lim
itatio
n
s
in
clin
ical
v
alid
it
y
s
till
r
em
ain
p
r
o
b
lem
s
.
T
h
i
s
h
as
em
p
h
asized
t
h
e
n
ee
d
f
o
r
tr
an
s
p
a
r
en
t
a
n
d
in
ter
p
r
etab
le
s
y
s
tem
s
to
m
a
k
e
AI
ac
ce
p
tab
le
in
r
ad
io
lo
g
y
.
B
en
ch
m
ar
k
an
al
y
s
is
s
h
o
wed
th
at
YOL
Ov
1
1
an
d
its
p
r
ed
ec
ess
o
r
s
d
em
o
n
s
tr
ated
s
t
ab
ilit
y
in
f
r
ac
tu
r
e
d
etec
tio
n
[
3
5
]
.
R
ec
en
t
r
esear
ch
h
as
al
s
o
s
h
o
wn
th
at
DL
,
esp
ec
ially
C
NN,
is
ca
p
ab
le
o
f
d
em
o
n
s
tr
atin
g
h
u
m
an
-
e
q
u
iv
alen
t
p
er
f
o
r
m
an
ce
i
n
b
o
n
e
f
r
a
ctu
r
e
class
if
icatio
n
an
d
C
AD
[
3
6
]
.
AI
-
b
ased
m
eth
o
d
s
,
esp
ec
ially
YOL
O
an
d
E
f
f
icien
tNet
-
b
as
ed
C
NN,
h
av
e
m
ad
e
f
r
ac
t
u
r
e
d
etec
tio
n
f
aster
an
d
m
o
r
e
ac
cu
r
ate.
T
h
ese
m
o
d
els
ca
n
h
elp
r
ed
u
ce
t
h
e
w
o
r
k
lo
ad
o
f
r
ad
i
o
lo
g
is
ts
an
d
im
p
r
o
v
e
p
atien
t
ca
r
e.
Nev
er
th
eless
,
f
u
tu
r
e
clin
ical
u
s
e
r
eq
u
ir
es
f
o
cu
s
o
n
lar
g
e
-
s
ca
le
clin
ical
tr
ials
,
im
p
r
o
v
in
g
th
e
in
ter
p
r
etab
ilit
y
o
f
m
o
d
els,
a
n
d
g
en
e
r
atin
g
s
tan
d
ar
d
ized
d
atasets
.
T
h
e
f
u
tu
r
e
o
f
au
t
o
m
ated
b
o
n
e
f
r
ac
tu
r
e
d
etec
ti
o
n
will d
ep
en
d
o
n
b
u
ild
in
g
AI
s
y
s
tem
s
th
at
ar
e
b
o
t
h
s
ca
lab
le
an
d
ea
s
y
to
u
n
d
e
r
s
tan
d
.
3.
RE
S
E
ARCH
M
E
T
H
O
D
I
n
th
is
s
ec
tio
n
,
th
e
d
etail
ed
r
esear
ch
m
eth
o
d
is
p
r
esen
t
ed
.
T
h
er
e
a
r
e
5
s
u
b
-
s
ec
tio
n
s
in
th
is
m
eth
o
d
o
l
o
g
y
.
Su
ch
as d
ataset,
r
esizin
g
an
d
la
b
ellin
g
,
au
g
m
e
n
tatio
n
,
d
ata
-
p
r
o
ce
s
s
in
g
,
a
n
d
m
o
d
el
s
el
ec
tio
n
.
3
.
1
.
Da
t
a
s
et
co
llect
i
o
n a
nd
a
nn
o
t
a
t
io
n
T
h
is
s
tu
d
y
em
p
lo
y
s
th
e
h
u
m
an
b
o
n
e
f
r
ac
tu
r
e
C
1
7
d
ataset
f
r
o
m
th
e
M
en
d
ele
y
d
ata
r
ep
o
s
ito
r
y
in
Fig
u
r
e
1
an
d
T
a
b
le
1
[
1
4
]
.
T
h
e
d
ataset
co
n
s
is
ts
o
f
X
-
r
a
y
im
ag
es
with
1
7
t
y
p
es
o
f
f
r
ac
tu
r
e
s
:
av
u
ls
io
n
,
cl
o
s
ed
,
co
m
m
in
u
ted
,
co
m
p
r
ess
io
n
,
d
i
s
lo
ca
ted
,
g
r
ee
n
s
tick
,
h
air
lin
e,
im
p
ac
ted
,
i
n
tr
a
-
ar
ticu
lar
,
lo
n
g
itu
d
in
al,
o
b
li
q
u
e,
o
p
en
,
p
ath
o
l
o
g
ical,
s
eg
m
e
n
tal,
s
p
ir
al,
s
tr
ess
,
an
d
tr
a
n
s
v
er
s
e
f
r
ac
tu
r
es.
T
h
e
r
e
ar
e
2
,
1
9
2
d
ata
p
o
in
ts
in
th
e
co
llectio
n
,
co
n
s
id
er
i
n
g
all
ty
p
es
o
f
f
r
ac
tu
r
es.
T
h
e
im
ag
es
ar
e
in
J
PG,
PNG,
an
d
W
E
B
P
f
o
r
m
ats.
E
ac
h
f
r
ac
tu
r
e
ty
p
e
is
in
a
s
ep
ar
ate
f
o
ld
er
.
T
h
e
d
ataset
co
m
b
in
es
s
am
p
le
s
f
r
o
m
Ka
g
g
le,
R
ad
io
p
ae
d
ia,
an
d
Sh
u
tter
s
to
ck
;
h
en
ce
,
it is
q
u
ite
d
iv
e
r
s
e
in
its
co
n
d
itio
n
s
o
f
im
ag
in
g
.
Fo
r
b
et
ter
p
er
f
o
r
m
an
ce
a
n
d
p
r
ev
en
tio
n
o
f
o
v
er
f
itti
n
g
,
all
im
ag
es
wer
e
r
esized
an
d
n
o
r
m
alize
d
,
an
d
d
if
f
er
e
n
t
au
g
m
e
n
tatio
n
p
r
o
ce
d
u
r
es
wer
e
p
er
f
o
r
m
ed
.
T
h
en
,
th
e
d
ata
was
d
iv
id
ed
in
to
tr
ain
in
g
,
v
alid
atio
n
,
an
d
test
s
ets.
Du
e
to
it
s
v
ar
ied
n
atu
r
e
an
d
well
-
o
r
g
a
n
ized
lab
elin
g
,
it
is
id
ea
l f
o
r
th
e
ass
ess
m
en
t o
f
YOL
O
v
ar
ian
ts
an
d
E
f
f
icien
tNet.
W
e
lab
eled
th
e
d
ata
u
s
in
g
th
e
R
o
b
o
Flo
w
p
latf
o
r
m
,
with
p
r
ec
is
e
b
o
u
n
d
in
g
b
o
x
es
ass
ig
n
ed
to
ea
ch
f
r
ac
tu
r
e
l
o
ca
tio
n
.
I
f
th
er
e
a
r
e
m
u
ltip
le
f
r
ac
tu
r
es
in
an
im
a
g
e,
a
s
ep
a
r
ate
b
o
u
n
d
in
g
b
o
x
is
cr
ea
ted
f
o
r
ea
ch
.
L
ab
eler
s
f
o
llo
wed
q
u
ality
g
u
i
d
elin
es,
an
d
th
e
jo
b
was
d
o
n
e
co
r
r
ec
tly
.
T
h
e
d
ataset
is
th
en
f
o
r
m
atted
t
o
b
e
u
s
ab
le
with
YOL
O,
wh
er
e
th
e
class
lab
el
an
d
b
o
u
n
d
in
g
b
o
x
co
o
r
d
in
ates
f
o
r
ea
ch
im
ag
e
ar
e
ap
p
en
d
ed
to
a
tex
t
f
ile.
T
h
e
d
ataset
is
s
p
lit
t
o
en
s
u
r
e
s
u
f
f
icien
t
d
ata
f
o
r
tr
ain
in
g
a
n
d
r
o
b
u
s
t
v
alid
atio
n
:
7
0
%
tr
ain
in
g
,
1
5
%
v
alid
atio
n
,
an
d
1
5
% testi
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
5
,
No
.
1
,
Ma
r
c
h
2
0
2
6
:
155
-
1
6
7
158
3
.
2
.
Re
s
izing
a
nd
la
bellin
g
As
s
h
o
wn
in
Fig
u
r
e
2
,
in
p
u
t
d
ata
is
n
o
r
m
alize
d
b
y
r
esizin
g
all
im
ag
es
to
6
4
0
×
6
4
0
p
i
x
els
to
b
e
co
m
p
atib
le
with
DL
m
o
d
els.
T
h
e
d
ataset
was
lab
eled
u
s
in
g
th
e
R
o
b
o
Flo
w
an
n
o
tatio
n
t
o
o
l,
wh
ich
m
a
d
e
it
ea
s
y
to
cr
ea
te
b
o
u
n
d
in
g
b
o
x
es
ac
co
r
d
in
g
to
class
I
Ds.
Fig
u
r
e
2
s
h
o
ws
h
o
w
t
h
e
d
ata
is
o
r
g
a
n
ized
u
s
in
g
r
esized
im
ag
es,
an
n
o
tated
b
o
u
n
d
in
g
b
o
x
es,
an
d
d
if
f
er
en
t
c
o
lo
r
c
o
d
es.
T
h
is
s
tr
u
ctu
r
ed
a
n
n
o
tati
o
n
p
r
o
ce
s
s
en
ab
les
ac
cu
r
ate
f
r
ac
tu
r
e
d
etec
tio
n
i
n
r
ea
l
-
life
ap
p
licatio
n
s
an
d
m
ak
e
s
m
o
d
el
tr
ain
in
g
s
im
p
le
an
d
e
f
f
ec
tiv
e.
Fig
u
r
e
1
.
Sam
p
le
d
ataset
T
ab
le
1
.
Data
s
et
s
tatis
tic
s
Ty
p
e
s
o
f
f
r
a
c
t
u
r
e
Q
u
a
n
t
i
t
y
w
i
t
h
i
n
e
a
c
h
t
y
p
e
Ty
p
e
s
o
f
f
r
a
c
t
u
r
e
Q
u
a
n
t
i
t
y
w
i
t
h
i
n
e
a
c
h
t
y
p
e
A
v
u
l
si
o
n
f
r
a
c
t
u
r
e
1
4
1
I
n
t
r
a
-
a
r
t
i
c
u
l
a
r
f
r
a
c
t
u
r
e
1
04
C
l
o
se
d
(
si
mp
l
e
)
f
r
a
c
t
u
r
e
1
1
2
Lo
n
g
i
t
u
d
i
n
a
l
f
r
a
c
t
u
r
e
1
29
C
o
mm
i
n
u
t
e
d
f
r
a
c
t
u
r
e
2
1
9
O
b
l
i
q
u
e
f
r
a
c
t
u
r
e
1
24
C
o
m
p
r
e
ss
i
o
n
-
c
r
u
s
h
f
r
a
c
t
u
r
e
1
50
O
p
e
n
(
c
o
mp
o
u
n
d
)
f
r
a
c
t
u
r
e
8
6
F
r
a
c
t
u
r
e
d
i
s
l
o
c
a
t
i
o
n
1
59
P
a
t
h
o
l
o
g
i
c
a
l
f
r
a
c
t
u
r
e
1
29
G
r
e
e
n
s
t
i
c
k
f
r
a
c
t
u
r
e
1
36
S
e
g
m
e
n
t
a
l
f
r
a
c
t
u
r
e
7
4
H
a
i
r
l
i
n
e
f
r
a
c
t
u
r
e
1
39
S
p
i
r
a
l
f
r
a
c
t
u
r
e
1
34
I
mp
a
c
t
e
d
f
r
a
c
t
u
r
e
1
61
S
t
r
e
ss fr
a
c
t
u
r
e
1
48
Tr
a
n
s
v
e
r
s
e
f
r
a
c
t
u
r
e
4
7
Fig
u
r
e
2
.
Sam
p
le
la
b
eled
d
ata
3
.
3
.
Aug
m
ent
a
t
io
n
T
ab
le
2
illu
s
tr
ates
th
o
s
e
v
ar
i
o
u
s
tech
n
iq
u
es
o
f
d
ata
a
u
g
m
en
tatio
n
h
av
e
b
ee
n
u
s
ed
to
i
n
cr
ea
s
e
th
e
g
en
er
aliza
b
ilit
y
o
f
th
e
m
o
d
el
an
d
o
v
er
co
m
e
th
e
lim
itatio
n
s
o
f
s
m
all
d
ataset
s
.
T
h
ese
m
eth
o
d
s
s
im
u
late
a
v
ar
iety
o
f
im
ag
in
g
s
ettin
g
s
,
allo
win
g
th
e
m
o
d
el
to
p
er
f
o
r
m
s
tab
ly
u
n
d
er
r
ea
l
-
wo
r
ld
co
n
d
iti
o
n
s
.
Au
g
m
e
n
tatio
n
in
clu
d
es
r
an
d
o
m
r
o
tatio
n
f
o
r
s
lig
h
t
ch
an
g
es
in
im
ag
e
an
g
le
(
±
1
5
°),
s
ca
lin
g
f
o
r
v
ar
iat
io
n
in
f
r
ac
tu
r
e
s
ize
(
±
1
0
%),
h
o
r
izo
n
tal
a
n
d
v
er
tic
al
f
lip
p
in
g
to
s
im
u
late
o
r
ien
ta
tio
n
ch
an
g
es,
b
r
ig
h
tn
ess
ad
ju
s
tm
en
t
(
±
2
0
%)
th
at
r
ef
lects
ch
an
g
es
in
X
-
r
ay
ex
p
o
s
u
r
e
an
d
in
ten
s
ity
,
an
d
r
a
n
d
o
m
b
lu
r
(
u
p
to
2
p
i
x
els)
th
at
s
im
u
lates
m
o
tio
n
o
r
f
o
cu
s
b
lu
r
in
r
ea
l
im
ag
es.
B
y
u
s
in
g
th
ese
a
u
g
m
en
tatio
n
s
r
an
d
o
m
ly
d
u
r
i
n
g
t
r
ain
in
g
,
th
e
d
at
aset
b
ec
o
m
es
m
o
r
e
d
iv
er
s
e,
wh
ich
s
ig
n
if
ica
n
tly
i
m
p
r
o
v
es th
e
p
er
f
o
r
m
an
ce
an
d
s
tab
ilit
y
o
f
th
e
m
o
d
el.
T
ab
le
2
.
Au
g
m
en
tatio
n
tec
h
n
i
q
u
es a
n
d
p
ar
am
eter
s
Te
c
h
n
i
q
u
e
s
P
a
r
a
me
t
e
r
s
Te
c
h
n
i
q
u
e
s
P
a
r
a
me
t
e
r
s
R
a
n
d
o
m
r
o
t
a
t
i
o
n
±
1
5
°
V
e
r
t
i
c
a
l
f
l
i
p
p
0
.
2
S
c
a
l
i
n
g
±
1
0
%
B
r
i
g
h
t
n
e
ss
a
d
j
u
s
t
me
n
t
±
2
0
%
H
o
r
i
z
o
n
t
a
l
f
lip
p
=
0
.
5
R
a
n
d
o
m
b
l
u
r
U
p
t
o
2
p
i
x
e
l
s
S
t
o
c
h
a
st
i
c
a
p
p
l
i
c
a
t
i
o
n
R
a
n
d
o
m
d
u
r
i
n
g
t
r
a
i
n
i
n
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
C
o
mp
a
r
a
tive
a
n
a
lysi
s
o
f YOL
O
va
r
ia
n
ts
a
n
d
E
fficien
tN
et
fo
r
d
etec
tin
g
b
o
n
e
fr
a
ctu
r
es
…
(
S
h
a
ta
b
d
i S
a
r
ke
r
)
159
3
.
4
.
Da
t
a
prepro
ce
s
s
ing
As s
h
o
w
n
i
n
Fi
g
u
r
e
3
,
X
-
r
a
y
i
m
a
g
es
ar
e
p
r
e
p
r
o
c
ess
e
d
i
n
m
u
l
tip
le
s
t
ep
s
t
o
m
ak
e
t
h
e
m
co
m
p
ati
b
le
wit
h
th
e
s
el
ec
t
e
d
m
o
d
e
l
,
wit
h
o
u
t
c
o
m
p
r
o
m
is
i
n
g
t
h
e
d
i
ag
n
o
s
ti
c
v
a
l
u
e
.
T
h
e
i
m
a
g
es
a
r
e
r
esiz
ed
t
o
6
4
0
×6
4
0
p
i
x
e
ls
f
o
r
th
e
YOL
O
m
o
d
el
an
d
2
2
4
×2
2
4
p
ix
els
f
o
r
E
f
f
icien
tNetB
0
,
t
o
m
atch
th
e
s
p
atial
d
im
en
s
io
n
s
o
f
th
e
in
p
u
t
to
th
e
m
o
d
el.
T
h
is
in
c
r
ea
s
es
p
er
f
o
r
m
an
ce
in
b
o
th
tr
ain
i
n
g
a
n
d
e
s
tim
a
tio
n
.
I
m
ag
e
co
n
tr
ast
is
en
h
an
ce
d
to
clar
if
y
b
o
n
e
s
tr
u
ctu
r
e
an
d
f
r
ac
tu
r
e
lin
es.
L
o
ca
l
co
n
tr
ast
is
en
h
a
n
ce
d
u
s
in
g
c
o
n
tr
ast
-
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
(
C
L
AHE
)
,
wh
ich
h
ig
h
lig
h
ts
f
in
e
f
r
ac
t
u
r
e
lin
e
s
with
in
th
e
b
o
n
e,
wh
ich
ar
e
n
o
t
ea
s
ily
s
ee
n
o
n
unc
o
n
tr
asted
X
-
r
ay
im
a
g
es.
A
G
a
u
s
s
ia
n
b
l
u
r
f
i
lt
e
r
is
u
s
ed
f
o
r
n
o
i
s
e
r
e
d
u
c
t
i
o
n
,
w
h
i
c
h
r
e
d
u
c
e
s
b
a
c
k
g
r
o
u
n
d
a
r
t
i
f
a
ct
s
a
n
d
h
i
g
h
-
f
r
e
q
u
e
n
c
y
n
o
i
s
e
.
T
h
i
s
s
t
a
b
i
li
z
es
f
e
at
u
r
e
e
x
t
r
ac
t
i
o
n
a
n
d
h
e
l
p
s
t
h
e
m
o
d
e
l
f
o
c
u
s
m
o
r
e
d
e
e
p
l
y
o
n
r
e
l
e
v
a
n
t
f
r
ac
t
u
r
e
p
a
t
t
e
r
n
s
.
M
o
r
p
h
o
l
o
g
i
c
a
l
o
p
e
r
ati
o
n
s
a
n
d
O
t
s
u
t
h
r
es
h
o
l
d
i
n
g
w
er
e
u
s
e
d
t
o
s
e
p
a
r
a
te
b
o
n
e
r
e
g
i
o
n
s
f
r
o
m
s
o
f
t
t
is
s
u
e
.
T
h
i
s
m
a
d
e
t
h
e
d
ia
g
n
o
s
t
i
c
a
n
a
l
y
s
is
m
o
r
e
a
c
c
u
r
a
t
e
b
y
r
e
d
u
c
i
n
g
th
e
i
n
f
l
u
e
n
c
e
o
f
i
r
r
e
l
e
v
a
n
t
a
n
at
o
m
i
c
a
l
f
e
a
t
u
r
es
.
Pix
el
in
ten
s
ity
v
alu
es
d
u
r
in
g
t
r
ain
in
g
ar
e
n
o
r
m
alize
d
t
o
th
e
[
0
,
1
]
r
an
g
e
b
y
d
iv
i
d
in
g
b
y
2
5
5
,
wh
ich
en
s
u
r
es
n
u
m
er
ical
s
tab
ilit
y
an
d
co
n
s
is
ten
t
g
r
ad
ien
t
u
p
d
ates.
I
m
ag
es
ar
e
k
ep
t
in
g
r
ay
s
ca
le
to
av
o
id
o
v
er
co
m
p
licatio
n
an
d
k
ee
p
k
ey
s
tr
u
ctu
r
al
in
f
o
r
m
atio
n
in
tact.
B
ased
o
n
v
ar
io
u
s
v
is
u
al
in
s
p
ec
tio
n
s
an
d
h
is
to
g
r
am
a
n
aly
s
is
,
th
is
p
r
ep
r
o
ce
s
s
in
g
p
ip
elin
e
in
cr
ea
s
ed
t
h
e
p
er
f
o
r
m
a
n
ce
a
n
d
s
tab
ilit
y
o
f
th
e
m
o
d
el
wh
ile
r
etain
in
g
im
p
o
r
tan
t d
ia
g
n
o
s
tic
in
f
o
r
m
atio
n
.
Fig
u
r
e
3
.
Data
p
r
ep
r
o
ce
s
s
in
g
3
.
5
.
M
o
del selec
t
io
n a
nd
a
rc
hite
ct
ure
T
h
is
s
tu
d
y
ev
alu
ated
s
ev
en
ad
v
an
ce
d
DL
m
o
d
els
–
YOL
Ov
8
,
YOL
Ov
9
,
YOL
Ov
1
0
,
YOL
Ov
1
1
,
E
f
f
icien
tNetB
0
,
Den
s
eNe
t1
6
9
,
an
d
R
esNet5
0
–
with
th
e
o
b
je
ctiv
e
o
f
d
eter
m
in
in
g
th
e
b
est
m
o
d
el
f
o
r
ac
c
u
r
ate
an
d
r
eliab
le
f
r
ac
tu
r
e
d
etec
tio
n
in
m
e
d
ical
X
-
r
ay
im
a
g
es.
T
o
en
s
u
r
e
f
air
an
d
co
n
s
is
ten
t
co
m
p
ar
is
o
n
s
,
all
m
o
d
els
ar
e
tr
ain
ed
a
n
d
ev
alu
ated
af
ter
p
r
ep
r
o
ce
s
s
in
g
with
th
e
s
am
e
ca
r
e
an
d
ap
p
ly
in
g
d
ata
au
g
m
en
tatio
n
.
T
h
e
ev
al
u
atio
n
r
esu
lts
s
h
o
wed
th
at
YOL
Ov
1
1
ac
h
iev
ed
t
h
e
h
ig
h
est
ac
cu
r
ac
y
a
n
d
is
c
o
n
s
id
er
ed
th
e
m
o
s
t
s
u
itab
le
m
o
d
el
f
o
r
th
is
a
p
p
lica
tio
n
.
I
n
th
e
f
ield
o
f
m
e
d
ical
d
i
ag
n
o
s
tics
,
wh
er
e
ev
en
a
s
m
all
m
is
tak
e
ca
n
h
av
e
a
b
ig
im
p
ac
t,
h
ig
h
ac
cu
r
ac
y
is
v
er
y
im
p
o
r
tan
t.
Fo
r
th
is
r
ea
s
o
n
,
th
e
s
elec
tio
n
o
f
YOL
Ov
1
1
ca
n
b
e
ca
lled
lo
g
ical
an
d
ju
s
tifia
b
le.
YOL
Ov
1
1
is
th
e
lates
t
v
er
s
io
n
o
f
th
e
YOL
O
s
er
ies,
wh
ich
b
r
in
g
s
m
an
y
ar
c
h
itectu
r
al
im
p
r
o
v
em
en
ts
.
I
ts
ad
v
an
ce
d
b
ac
k
b
o
n
e
n
etw
o
r
k
is
ca
p
ab
le
o
f
ca
p
tu
r
in
g
h
ig
h
-
r
eso
lu
tio
n
f
i
n
e
f
ea
tu
r
es,
wh
ich
is
h
elp
f
u
l
in
d
etec
tin
g
o
f
ten
-
m
is
s
ed
f
r
ac
tu
r
e
lin
es.
Mu
lti
-
s
ca
le
f
ea
tu
r
e
in
teg
r
atio
n
is
ac
co
m
p
lis
h
ed
th
r
o
u
g
h
th
e
n
ec
k
ar
ch
itectu
r
e,
wh
ich
co
m
b
in
es
th
e
p
ath
ag
g
r
e
g
atio
n
n
etwo
r
k
an
d
th
e
ad
v
an
ce
d
f
ea
tu
r
e
p
y
r
a
m
id
n
etwo
r
k
.
I
t
is
v
er
y
im
p
o
r
tan
t to
id
en
tify
f
r
ac
tu
r
es o
f
d
if
f
er
en
t sh
ap
es,
s
izes,
an
d
p
o
s
itio
n
s
.
Mo
r
eo
v
er
,
YOL
Ov
1
1
h
as
an
an
ch
o
r
-
f
r
ee
d
etec
tio
n
h
ea
d
,
wh
ic
h
r
ed
u
ce
s
d
e
p
en
d
e
n
ce
o
n
p
r
ed
ef
in
e
d
an
ch
o
r
s
.
T
h
is
r
esu
lts
in
in
cr
e
ased
b
o
u
n
d
in
g
b
o
x
lo
ca
tio
n
a
cc
u
r
ac
y
an
d
co
n
f
i
d
en
ce
s
co
r
e
s
,
wh
ich
im
p
r
o
v
es
m
o
d
el
f
lex
ib
ilit
y
an
d
g
e
n
er
ali
za
b
ilit
y
to
d
if
f
er
en
t
d
atasets
.
W
h
ile
h
ig
h
r
ec
all
is
im
p
o
r
tan
t
in
m
ed
ical
u
s
e,
it
i
s
eq
u
ally
im
p
o
r
tan
t
to
r
e
d
u
ce
f
alse
p
o
s
itiv
es
in
o
r
d
er
to
av
o
id
u
n
n
ec
ess
ar
y
tr
ea
tm
e
n
t
o
r
m
is
d
iag
n
o
s
is
[
3
5
]
.
Fin
ally
,
YOL
Ov
1
1
d
em
o
n
s
tr
a
ted
in
cr
ea
s
ed
ac
cu
r
ac
y
,
r
o
b
u
s
tn
ess
,
an
d
ad
ap
tab
ilit
y
co
m
p
ar
ed
to
o
th
er
m
o
d
els
–
YOL
Ov
8
,
YOL
Ov
9
,
YOL
Ov
1
0
,
E
f
f
icien
tNetB
0
,
Den
s
eNe
t1
6
9
,
a
n
d
R
esNet5
0
.
I
t
s
tan
d
s
o
u
t
as
a
v
er
y
s
tr
o
n
g
ca
n
d
id
ate
f
o
r
u
s
e
in
a
u
to
m
ated
f
r
ac
tu
r
e
d
etec
tio
n
an
d
clin
ical
d
ec
is
io
n
-
m
ak
in
g
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
tr
ain
in
g
s
etu
p
an
d
h
y
p
er
p
ar
am
eter
s
f
o
r
YOL
Ov
1
1
-
b
ased
b
o
n
e
f
r
ac
tu
r
e
d
etec
t
io
n
ar
e
s
h
o
wn
in
T
ab
le
3
.
T
h
e
m
o
d
el
u
s
es
th
e
Ad
am
W
o
p
tim
izer
,
wh
ich
h
as
an
in
itial
lear
n
in
g
r
ate
o
f
0
.
0
0
0
4
7
6
an
d
a
m
o
m
en
tu
m
o
f
0
.
9
.
T
h
is
s
etu
p
h
elp
s
to
u
p
d
ate
weig
h
ts
q
u
ic
k
ly
an
d
e
f
f
icien
tly
o
v
er
2
0
0
ep
o
ch
s
.
Pre
-
tr
ain
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
5
,
No
.
1
,
Ma
r
c
h
2
0
2
6
:
155
-
1
6
7
160
weig
h
ts
(
p
r
e
-
tr
ain
ed
=
tr
u
e
)
f
r
o
m
th
e
C
OC
O
d
ataset
ar
e
u
s
ed
to
s
p
ee
d
u
p
tr
ain
in
g
.
Als
o
,
if
v
alid
atio
n
f
o
r
1
0
0
ep
o
ch
s
d
o
es n
o
t c
h
a
n
g
e
th
e
p
er
f
o
r
m
an
ce
,
t
r
ain
in
g
is
s
to
p
p
ed
b
y
ap
p
l
y
in
g
ea
r
l
y
s
to
p
p
in
g
(
p
atien
ce
=
1
0
0
)
.
Au
to
m
atic
m
ix
ed
p
r
ec
is
io
n
(
AM
P)
is
u
s
ed
to
in
cr
ea
s
e
co
m
p
u
tatio
n
al
ef
f
icien
c
y
.
I
t
co
m
b
in
es
1
6
-
b
it
an
d
3
2
-
b
it
f
lo
atin
g
-
p
o
in
t,
r
esu
ltin
g
in
less
m
em
o
r
y
a
n
d
f
aster
tr
ain
in
g
.
T
o
m
ain
tain
lear
n
e
d
f
ea
tu
r
es,
a
s
in
g
le
co
n
v
o
l
u
tio
n
al
lay
er
(
m
o
d
el.
2
3
.
d
f
l.c
o
n
v
.
weig
h
t)
is
f
r
o
ze
n
.
Fo
r
r
eg
u
lar
izatio
n
,
in
p
u
t
im
a
g
es
ar
e
r
esized
to
6
4
0
×6
4
0
p
ix
els
wit
h
a
b
atc
h
s
ize
o
f
1
6
an
d
a
weig
h
t
d
ec
a
y
o
f
0
.
0
0
0
5
.
T
h
e
r
o
b
u
s
tn
ess
o
f
d
atasets
is
im
p
r
o
v
ed
b
y
d
ata
au
g
m
en
tatio
n
m
eth
o
d
s
s
u
ch
as
er
asin
g
(
er
asin
g
=0
.
4
)
,
h
o
r
iz
o
n
tal
f
lip
p
in
g
(
f
lip
lr
=0
.
5
)
,
a
n
d
r
an
d
au
g
m
en
t.
Ou
tp
u
ts
ar
e
s
av
ed
in
T
o
r
c
h
Scr
ip
t
f
o
r
m
at
f
o
r
d
ep
lo
y
m
e
n
t,
with
a
m
a
x
im
u
m
d
etec
tio
n
lim
it
o
f
3
0
0
item
s
p
er
im
ag
e
an
d
o
v
er
lap
p
in
g
m
ask
s
m
an
ag
ed
b
y
a
m
ask
r
atio
o
f
f
o
u
r
.
Ob
ject
tr
ac
k
in
g
is
co
n
f
ig
u
r
e
d
with
B
o
ts
o
r
t.y
am
l to
en
s
u
r
e
p
r
ec
is
e
d
etec
tio
n
.
T
ab
le
3
.
YOL
Ov
1
1
-
b
ased
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
m
o
d
el
h
y
p
er
p
ar
am
et
er
s
P
a
r
a
me
t
e
r
s
V
a
l
u
e
P
a
r
a
me
t
e
r
s
V
a
l
u
e
B
a
t
c
h
si
z
e
16
A
M
P
T
r
u
e
N
u
mb
e
r
o
f
e
p
o
c
h
s
2
0
0
M
a
x
_
d
e
t
3
0
0
O
p
t
i
mi
z
e
r
A
u
t
o
(
A
d
a
mW
s
e
l
e
c
t
e
d
)
F
o
r
mat
To
r
c
h
s
c
r
i
p
t
Pre
-
t
r
a
i
n
e
d
w
e
i
g
h
t
s
T
r
u
e
Tr
a
c
k
e
r
B
o
t
so
r
t
.
y
a
m
l
Le
a
r
n
i
n
g
r
a
t
e
(
l
r
0
)
0
.
0
0
0
4
7
6
A
u
t
o
a
u
g
me
n
t
R
a
n
d
a
u
g
m
e
n
t
M
o
me
n
t
u
m
0
.
9
O
v
e
r
l
a
p
mas
k
En
a
b
l
e
d
(
mas
k
r
a
t
i
o
=
4
)
P
a
t
i
e
n
c
e
1
00
Er
a
si
n
g
0
.
4
I
mag
e
si
z
e
6
4
0
×
6
4
0
p
i
x
e
l
s
F
l
i
p
p
i
n
g
(
f
l
i
p
l
r
)
0
.
5
W
e
i
g
h
t
d
e
c
a
y
0
.
0
0
0
5
Fre
e
z
i
n
g
l
a
y
e
r
‘
mo
d
e
l
.
2
3
.
d
f
l
.
c
o
n
v
.
w
e
i
g
h
t
’
4
.
1
.
M
o
del
ev
a
lua
t
i
o
n
T
ab
le
4
s
u
m
m
ar
izes
a
n
a
d
v
a
n
ce
d
YOL
Ov
1
1
-
b
ased
DL
m
o
d
el
wh
ich
is
d
e
v
elo
p
e
d
to
ac
cu
r
ately
d
etec
t
an
d
class
if
y
cr
ac
k
s
o
r
f
r
ac
tu
r
es
in
b
o
n
e
X
-
r
a
y
im
ag
es.
T
h
is
m
o
d
el
h
as
a
to
tal
o
f
3
1
9
lay
er
s
.
I
t
co
n
tain
s
a
to
tal
o
f
9
,
4
3
4
,
3
7
1
p
a
r
am
eter
s
,
o
f
wh
ich
9
,
4
3
4
,
3
5
5
a
r
e
a
ctu
ally
tr
ain
ab
le
–
th
at
is
,
th
e
y
ca
n
b
e
ch
a
n
g
ed
d
u
r
i
n
g
tr
ain
in
g
to
m
a
k
e
th
e
m
o
d
el
m
o
r
e
ac
cu
r
ate.
T
h
ese
p
ar
am
eter
s
ar
e
f
in
e
-
tu
n
ed
d
u
r
i
n
g
tr
ain
in
g
s
o
th
at
th
e
m
o
d
el
ca
n
u
n
d
er
s
tan
d
s
u
b
tle
a
n
d
c
o
m
p
lex
f
r
ac
t
u
r
e
p
atter
n
s
with
in
X
-
r
ay
im
ag
es
a
n
d
ca
n
tell
v
e
r
y
ac
cu
r
ately
wh
er
e
th
e
f
r
ac
tu
r
e
is
.
T
h
e
co
m
p
u
tatio
n
al
co
s
t
o
f
th
e
m
o
d
el
is
o
n
ly
2
1
.
6
g
i
g
a
f
lo
ati
n
g
p
o
in
t
o
p
er
atio
n
s
(
GFLO
Ps
)
,
m
ea
n
in
g
it
ca
n
wo
r
k
q
u
ic
k
ly
with
o
u
t
m
u
ch
p
r
o
ce
s
s
in
g
p
o
wer
.
I
t
ca
n
th
er
e
f
o
r
e
b
e
u
s
ed
in
r
ea
l
-
ti
m
e
s
y
s
tem
s
,
s
u
ch
as
d
etec
tin
g
f
r
ac
tu
r
es
i
m
m
ed
iately
af
ter
an
X
-
r
ay
is
tak
en
in
a
h
o
s
p
ital.
T
h
e
s
tr
u
ctu
r
e
o
f
t
h
e
m
o
d
el
is
d
ef
in
e
d
in
th
e
‘
YOL
O1
1
s
.
y
am
l’
f
ile.
He
r
e'
s
h
o
w
ea
ch
p
ar
t
o
f
th
e
n
et
wo
r
k
is
laid
o
u
t,
wh
at
lay
er
s
th
er
e
ar
e,
h
o
w
d
ata
f
lo
ws
–
all
th
e
in
f
o
r
m
atio
n
.
On
th
e
o
th
er
h
an
d
,
th
e
'
d
ata.
y
am
l'
f
ile
co
n
tain
s
im
p
o
r
tan
t d
etails o
f
th
e
tr
ain
in
g
an
d
v
alid
atio
n
d
atasets
–
s
u
ch
as
w
h
ich
class
es
ar
e
th
er
e
(
eg
,
f
r
a
ctu
r
e,
n
o
f
r
ac
tu
r
e)
,
wh
e
r
e
th
e
im
ag
es
ar
e
s
to
r
ed
,
an
d
wh
er
e
th
e
lab
el
f
iles
ar
e.
T
h
ese
two
YAM
L
f
iles
–
o
n
e
f
o
r
th
e
ar
ch
itectu
r
e
an
d
t
h
e
o
t
h
er
f
o
r
th
e
d
ataset
co
n
f
ig
u
r
atio
n
–
t
o
g
eth
er
cr
ea
te
a
well
-
o
r
g
an
ize
d
an
d
s
y
s
tem
atic
s
y
s
tem
f
o
r
tr
ain
i
n
g
a
n
d
e
v
alu
atin
g
th
e
e
n
tire
m
o
d
el.
T
h
at
is
,
th
e
en
tire
p
r
o
ce
s
s
i
s
d
o
n
e
in
a
clea
n
f
r
am
ewo
r
k
,
wh
ich
h
elp
s
im
p
r
o
v
e
m
o
d
el
ac
cu
r
ac
y
an
d
p
er
f
o
r
m
an
ce
.
T
ab
le
4
.
Mo
d
e
l p
ar
am
ete
r
s
P
a
r
a
me
t
e
r
s
V
a
l
u
e
P
a
r
a
me
t
e
r
s
V
a
l
u
e
M
o
d
e
l
l
a
y
e
r
s
3
1
9
G
r
a
d
i
e
n
t
s
9
,
4
3
4
,
3
5
5
M
o
d
e
l
p
a
r
a
me
t
e
r
s
9
,
4
3
4
,
3
7
1
G
F
LO
P
s
2
1
.
6
D
a
t
a
c
o
n
f
i
g
u
r
a
t
i
o
n
D
a
t
a
.
y
a
ml
M
o
d
e
l
c
o
n
f
i
g
u
r
a
t
i
o
n
Y
O
LO
1
1
s
.
y
a
ml
4
.
2
.
H
a
rdwa
re
des
cr
iptio
n
T
ab
le
5
illu
s
tr
ates
th
at
all
test
s
wer
e
p
er
f
o
r
m
ed
o
n
a
c
o
m
p
u
ter
with
a
1
0
th
Ge
n
I
n
tel
C
o
r
e
i7
-
1
0
7
0
0
p
r
o
ce
s
s
o
r
,
NVI
DI
A
R
T
X
3
0
5
0
g
r
ap
h
ics
ca
r
d
with
1
6
GB
DDR4
R
AM
,
an
d
8
GB
m
em
o
r
y
.
T
h
e
s
y
s
tem
was
r
u
n
n
in
g
o
n
th
e
U
b
u
n
tu
2
4
.
0
4
o
p
e
r
atin
g
s
y
s
tem
.
T
h
is
h
ar
d
war
e
co
n
f
ig
u
r
atio
n
g
a
v
e
u
s
en
o
u
g
h
p
o
wer
to
co
m
f
o
r
ta
b
ly
tr
ain
,
test
,
an
d
ev
alu
ate
th
e
DL
m
o
d
els we
u
s
ed
.
T
ab
le
5
.
Sy
s
tem
s
p
ec
if
icati
on
C
o
m
p
o
n
e
n
t
S
p
e
c
i
f
i
c
a
t
i
o
n
C
o
m
p
o
n
e
n
t
S
p
e
c
i
f
i
c
a
t
i
o
n
C
P
U
1
0
t
h
G
e
n
I
n
t
e
l
C
o
r
e
i
7
1
0
7
0
0
R
A
M
1
6
G
B
D
D
R
4
OS
U
b
u
n
t
u
2
4
.
0
4
G
P
U
N
V
I
D
I
A
R
TX
3
0
5
0
8
G
B
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
C
o
mp
a
r
a
tive
a
n
a
lysi
s
o
f YOL
O
va
r
ia
n
ts
a
n
d
E
fficien
tN
et
fo
r
d
etec
tin
g
b
o
n
e
fr
a
ctu
r
es
…
(
S
h
a
ta
b
d
i S
a
r
ke
r
)
161
4
.
3
.
Ana
ly
s
is
o
f
re
s
ults
T
o
u
n
d
er
s
tan
d
th
e
im
p
r
o
v
e
m
en
t
o
f
YOL
Ov
1
1
in
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
,
it
is
co
m
p
ar
ed
with
YOL
Ov
1
0
,
YOL
Ov
8
,
YOL
Ov
9
,
E
f
f
icien
tNetB
0
,
Den
s
eNe
t1
6
9
,
an
d
R
esNet5
0
m
o
d
els.
T
h
e
m
o
s
t
im
p
o
r
tan
t
co
m
p
ar
is
o
n
was w
ith
YOL
Ov
1
0
,
as b
o
th
ar
e
n
ewe
r
v
er
s
io
n
s
o
f
th
e
s
am
e
ar
ch
i
tectu
r
e.
Fr
o
m
th
e
r
esu
lts
,
wh
ich
wer
e
g
iv
en
in
Fig
u
r
e
4
,
YOL
Ov
1
1
(
Fig
u
r
e
4
(
a)
)
s
h
o
wed
b
etter
b
eh
av
io
r
th
a
n
YOL
Ov
1
0
(
Fig
u
r
e
4
(
b
)
)
in
all
ty
p
es o
f
s
ig
n
if
ican
t lo
s
s
o
r
lo
s
s
m
ea
n
s
d
u
r
in
g
tr
ain
in
g
.
Fo
r
e
x
am
p
le,
b
o
x
r
eg
r
ess
io
n
lo
s
s
–
wh
ich
tells
h
o
w
wel
l
th
e
m
o
d
el
lear
n
s
f
r
ac
tu
r
e
l
o
ca
tio
n
s
–
d
r
o
p
p
ed
f
r
o
m
0
.
6
5
t
o
0
.
3
0
f
o
r
YOL
Ov
1
0
,
wh
ile
YOL
O
v
1
1
d
r
o
p
p
e
d
e
v
en
b
etter
f
r
o
m
0
.
7
5
to
0
.
2
2
.
T
h
is
s
u
g
g
ests
th
at
o
v
er
tim
e,
YOL
Ov
1
1
h
as
lear
n
ed
th
e
lo
ca
tio
n
o
f
b
o
n
e
r
u
p
tu
r
es
m
o
r
e
ac
c
u
r
ately
.
W
h
ile
class
if
icatio
n
lo
s
s
-
wh
ic
h
t
ells
h
o
w
well
th
e
m
o
d
el
ca
n
tell
if
a
n
i
m
ag
e
h
as f
r
ac
tu
r
es
–
d
r
o
p
p
ed
r
elativ
ely
q
u
ick
ly
f
o
r
YOL
Ov
1
0
(
f
r
o
m
6
.
0
t
o
0
.
7
)
,
YOL
Ov
1
1
d
r
o
p
p
ed
r
elativ
el
y
s
lo
wly
b
u
t
m
o
r
e
s
tead
ily
f
r
o
m
4
.
0
t
o
1
.
0
.
T
h
i
s
m
ea
n
s
YOL
Ov
1
1
lear
n
s
m
o
r
e
s
tab
ly
an
d
r
eliab
ly
t
o
r
e
co
g
n
ize
s
u
b
tle
an
d
co
m
p
lex
f
r
ac
t
u
r
es.
Similar
r
esu
lts
wer
e
s
ee
n
in
d
is
tr
ib
u
tio
n
f
o
ca
l
lo
s
s
(
DFL)
.
YOL
Ov
1
1
m
an
ag
ed
to
r
e
d
u
ce
th
is
lo
s
s
f
r
o
m
1
.
3
8
to
0
.
9
5
,
w
h
ich
is
s
im
ilar
to
YOL
Ov
1
0
'
s
r
ed
u
ctio
n
f
r
o
m
1
.
3
3
t
o
0
.
9
8
-
b
u
t
s
lig
h
tly
b
etter
.
Ov
er
all,
th
e
d
ata
s
h
o
w
th
at
YOL
Ov
1
1
is
n
o
t
o
n
l
y
b
ett
er
th
an
p
r
e
v
io
u
s
m
o
d
els
b
u
t
also
lear
n
s
m
o
r
e
ac
cu
r
ately
,
co
n
s
is
ten
tly
,
an
d
r
eliab
ly
in
d
etec
tin
g
b
o
n
e
f
r
ac
t
u
r
es.
T
h
e
r
esu
lts
o
f
th
e
v
alid
atio
n
p
h
ase
p
r
o
v
ed
th
e
p
o
wer
o
f
YOL
Ov
1
1
m
o
r
e
clea
r
ly
.
YOL
Ov
1
1
s
h
o
wed
th
at
its
v
alid
atio
n
b
o
x
lo
s
s
an
d
D
FL
lo
s
s
d
ec
r
ea
s
ed
s
lo
wly
an
d
s
m
o
o
th
ly
–
m
ea
n
in
g
th
e
m
o
d
el
lear
n
ed
s
tead
ily
.
B
u
t
f
o
r
YOL
Ov
1
0
,
th
ese
g
r
ap
h
s
f
lu
ctu
ated
,
s
h
o
win
g
th
at
Y
OL
Ov
1
0
d
id
n
o
t
lear
n
as
s
tab
ly
an
d
h
a
d
tr
o
u
b
le
ad
ap
tin
g
well
to
n
ew
d
ata.
T
h
e
s
am
e
th
in
g
ca
n
b
e
s
ee
n
f
o
r
c
lass
if
icatio
n
lo
s
s
–
Y
OL
Ov
1
1
co
n
s
is
ten
tly
r
ed
u
ce
d
lo
s
s
,
b
u
t
YOL
Ov
1
0
'
s
g
r
ap
h
was
m
u
ch
m
o
r
e
v
o
latile.
T
h
is
im
p
lies
th
at
YOL
Ov
1
1
ca
n
m
o
r
e
co
n
s
is
ten
tly
an
d
r
eliab
ly
u
n
d
e
r
s
tan
d
wh
eth
er
t
h
er
e
ar
e
f
r
ac
tu
r
es
in
th
e
im
ag
e.
YOL
Ov
1
1
was
ah
ea
d
in
d
e
tectio
n
p
er
f
o
r
m
an
ce
as
well.
T
h
e
m
o
d
el
s
h
o
wed
a
n
im
p
r
o
v
em
e
n
t
in
p
r
ec
is
io
n
f
r
o
m
ab
o
u
t
0
.
1
0
to
0
.
4
8
,
in
d
ica
tin
g
th
at
it
r
ed
u
ce
d
th
e
ten
d
en
c
y
to
f
alsely
d
etec
t
“f
r
ac
tu
r
es”
(
f
alse
p
o
s
itiv
es).
W
h
ile
YOL
Ov
1
0
ev
en
t
u
ally
r
ea
ch
ed
t
h
e
s
am
e
p
lace
in
ac
c
u
r
ac
y
,
YOL
Ov
1
1
m
ain
tain
ed
a
g
o
o
d
b
ala
n
c
e
b
etwe
en
c
o
r
r
ec
t
d
etec
tio
n
an
d
f
alse
alar
m
s
th
r
o
u
g
h
o
u
t
t
r
ain
in
g
.
YOL
Ov
1
1
was
also
m
o
r
e
s
tab
le
in
ter
m
s
o
f
r
ec
all
-
h
o
w
m
a
n
y
tr
u
e
f
r
ac
tu
r
es
th
e
m
o
d
el
was
ab
le
to
d
etec
t.
At
f
ir
s
t,
Y
OL
Ov
1
0
s
h
o
wed
a
r
tific
ially
h
ig
h
r
ec
all
(
b
ec
au
s
e
it
was
s
a
y
in
g
“
f
r
ac
tu
r
e”
to
o
m
an
y
p
lace
s
)
,
b
u
t
later
it
d
ec
r
ea
s
ed
.
I
n
co
n
tr
ast,
YOL
Ov
1
1
m
ain
tain
s
a
co
n
s
tan
t
r
ec
all
o
f
ar
o
u
n
d
0
.
4
5
th
r
o
u
g
h
o
u
t,
wh
ich
is
p
r
ac
ticall
y
m
o
r
e
m
ea
n
in
g
f
u
l a
n
d
r
eliab
l
e.
T
h
e
tr
u
e
p
o
wer
o
f
YOL
Ov
1
1
is
m
o
s
t
clea
r
ly
s
ee
n
in
th
e
m
AP
r
esu
lts
.
Her
e,
YOL
Ov
1
1
'
s
m
AP@
5
0
v
alu
e
in
c
r
ea
s
ed
f
r
o
m
0
.
1
5
t
o
0
.
5
2
,
wh
ile
YOL
Ov
1
0
o
n
ly
m
an
ag
ed
to
r
is
e
f
r
o
m
0
.
1
5
to
0
.
4
5
.
T
h
at
is
,
u
n
d
er
th
e
s
am
e
co
n
d
itio
n
s
,
YOL
O
v
1
1
lea
r
n
s
to
d
etec
t
f
r
ac
tu
r
e
s
m
u
ch
m
o
r
e
ac
cu
r
ately
.
Similar
r
esu
lts
wer
e
o
b
tain
ed
b
y
lo
o
k
i
n
g
at
th
e
m
o
r
e
s
tr
in
g
en
t
ev
alu
ati
o
n
m
e
tr
ic
m
AP@
5
0
-
9
5
.
W
h
ile
b
o
th
m
o
d
els
im
p
r
o
v
ed
,
YOL
Ov
1
1
'
s
v
alu
e
in
cr
ea
s
ed
f
r
o
m
0
.
0
8
to
0
.
3
8
–
m
a
r
g
in
ally
s
u
r
p
ass
in
g
YOL
Ov
1
0
'
s
in
cr
ea
s
e
f
r
o
m
0
.
0
7
to
0
.
3
8
.
T
h
is
m
ea
n
s
th
at
at
d
if
f
er
en
t
in
ter
s
ec
tio
n
o
v
er
u
n
io
n
(
I
o
U
)
th
r
esh
o
ld
s
(
i.e
.
,
d
if
f
er
e
n
t
s
tiff
n
ess
cr
iter
ia)
,
YOL
Ov
1
1
is
ab
le
to
d
etec
t
f
r
ac
tu
r
es
s
lig
h
tly
m
o
r
e
ac
cu
r
ate
ly
.
Ov
er
all,
th
e
r
esu
lts
s
h
o
w
t
h
at
YOL
Ov
1
1
ca
n
ad
ap
t
to
n
ew
d
ata
b
etter
th
a
n
p
r
ev
io
u
s
v
e
r
s
io
n
s
(
b
etter
g
en
e
r
aliza
tio
n
)
.
I
ts
b
o
u
n
d
in
g
b
o
x
e
s
tim
atio
n
(
lo
ca
tio
n
o
f
f
r
ac
tu
r
e)
is
m
o
r
e
ac
cu
r
ate
,
an
d
its
o
v
er
all
d
etec
t
io
n
ca
p
ab
ilit
y
is
also
m
o
r
e
p
o
wer
f
u
l.
T
h
er
ef
o
r
e,
YOL
Ov
1
1
is
a
m
o
r
e
r
eliab
le
an
d
ef
f
ec
tiv
e
ch
o
ice
f
o
r
au
t
o
m
atic
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
in
r
ea
l
h
o
s
p
itals
o
r
m
ed
ical
im
ag
in
g
s
y
s
tem
s
.
4
.
4
.
E
v
a
lua
t
i
o
n m
et
rics
Var
io
u
s
im
p
o
r
ta
n
t
m
etr
ics,
in
clu
d
in
g
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
an
d
p
r
ec
is
io
n
-
r
ec
all
(
PR
)
cu
r
v
e,
ar
e
u
s
ed
to
u
n
d
er
s
tan
d
h
o
w
well
th
e
YOL
Ov
1
1
m
o
d
el
p
er
f
o
r
m
s
in
b
o
n
e
f
r
ac
tu
r
e
d
etec
tio
n
.
T
h
e
m
o
d
el
s
h
o
we
d
a
p
r
ec
is
io
n
o
f
0
.
9
9
at
th
e
1
.
0
c
o
n
f
id
en
ce
le
v
el,
wh
ich
m
ea
n
s
th
at
with
v
er
y
h
ig
h
co
n
f
id
e
n
ce
,
th
e
m
o
d
el
ca
n
ac
cu
r
ately
d
etec
t
f
r
ac
tu
r
es
w
ith
alm
o
s
t
n
o
er
r
o
r
.
Ho
wev
e
r
,
lo
wer
in
g
th
e
co
n
f
id
en
ce
le
v
el
in
cr
ea
s
es
f
alse
p
o
s
itiv
es,
th
er
eb
y
d
ec
r
ea
s
in
g
p
r
ec
is
io
n
.
On
th
e
o
th
e
r
h
an
d
,
th
e
r
ec
all,
i.e
.
,
h
o
w
m
a
n
y
r
ea
l
f
r
ac
tu
r
es
th
e
m
o
d
e
l
ca
n
d
etec
t,
r
ea
ch
es a
m
a
x
im
u
m
o
f
0
.
7
0
wh
en
th
e
co
n
f
i
d
en
c
e
lev
el
is
0
.
0
,
an
d
th
e
r
ec
all
g
r
ad
u
ally
d
ec
r
ea
s
es
as
th
e
co
n
f
i
d
en
ce
i
n
cr
ea
s
es.
T
h
e
F1
-
s
co
r
e
g
r
ap
h
s
h
o
ws
th
at
t
h
e
b
est
b
alan
ce
b
etwe
en
p
r
ec
is
io
n
an
d
r
ec
all
is
o
b
tain
ed
at
a
co
n
f
i
d
en
ce
lev
el
o
f
0
.
6
9
0
,
wh
ile
th
e
h
ig
h
est
F1
-
s
co
r
e
was
0
.
4
6
.
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el
is
n
o
t
eq
u
al
f
o
r
d
if
f
e
r
en
t
t
y
p
es
o
f
er
o
s
io
n
.
Fo
r
ex
a
m
p
le,
th
e
F1
-
s
co
r
e
in
th
e
“c
lo
s
ed
s
im
p
le
f
r
ac
tu
r
e”
class
was
v
er
y
g
o
o
d
,
b
u
t in
s
o
m
e
class
es
th
e
r
esu
lts
wer
e
p
o
o
r
,
wh
ich
s
u
g
g
ests
th
at
th
er
e
wer
e
to
o
m
an
y
f
ea
tu
r
es o
r
th
at
th
er
e
was
r
elativ
ely
litt
le
d
at
a.
Av
er
ag
e
ac
cu
r
ac
y
(
AP)
wa
s
m
ea
s
u
r
ed
b
y
PR
cu
r
v
e,
wh
er
e
“c
lo
s
ed
-
s
im
p
le
-
f
r
ac
tu
r
e”
a
n
d
“c
o
m
p
r
ess
io
n
-
c
r
u
s
h
-
f
r
ac
tu
r
e”
class
es
o
b
tain
ed
0
.
8
5
5
a
n
d
0
.
8
1
9
AP,
r
esp
ec
ti
v
ely
,
s
h
o
win
g
v
e
r
y
g
o
o
d
d
etec
tio
n
ab
ilit
y
.
I
n
co
n
tr
ast,
th
e
AP
i
n
th
e
“lo
n
g
itu
d
in
al
-
f
r
ac
t
u
r
e”
a
n
d
“tr
an
s
v
e
r
s
e
-
f
r
ac
tu
r
e”
class
es
wer
e
o
n
ly
0
.
0
9
5
an
d
0
.
1
0
0
,
in
d
icatin
g
th
at
s
u
ch
f
r
ac
tu
r
es
wer
e
q
u
ite
d
if
f
ic
u
lt
to
d
etec
t.
Fin
ally
,
th
e
o
v
e
r
all
m
AP
o
f
th
e
m
o
d
el
at
th
e
0
.
5
I
o
U
th
r
esh
o
ld
was
0
.
4
6
2
.
T
h
is
s
h
o
ws
th
at
p
er
f
o
r
m
an
ce
i
s
f
ai
r
ly
g
o
o
d
,
b
u
t
th
er
e
is
s
till
r
o
o
m
f
o
r
im
p
r
o
v
em
e
n
t,
esp
ec
ially
f
o
r
d
if
f
icu
lt a
n
d
u
n
d
er
r
ep
r
esen
te
d
f
r
ac
tu
r
e
class
es
.
T
h
e
co
m
p
ar
ativ
e
r
esu
lts
in
T
a
b
les
6
a
n
d
7
s
h
o
w
th
at
th
er
e
is
s
o
m
e
tr
ad
e
-
o
f
f
b
etwe
en
p
r
e
cisi
o
n
an
d
r
ec
all
am
o
n
g
th
e
d
i
f
f
er
en
t
test
ed
m
o
d
e
ls
.
T
h
e
p
r
o
p
o
s
ed
YOL
Ov
1
1
m
o
d
el
ac
h
iev
es
a
v
er
y
h
ig
h
ac
cu
r
ac
y
,
i.e
.
,
0
.
9
9
,
wh
ile
E
f
f
icien
tNetB
0
h
as
an
ac
cu
r
ac
y
o
f
0
.
9
4
,
wh
ic
h
s
h
o
ws
th
at
th
ese
m
o
d
els
ar
e
l
ess
lik
ely
to
f
alsely
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
5
,
No
.
1
,
Ma
r
c
h
2
0
2
6
:
155
-
1
6
7
162
id
en
tify
n
o
n
-
b
r
o
k
e
n
b
o
n
es
as
f
r
ac
tu
r
es.
T
h
is
is
v
er
y
im
p
o
r
tan
t
f
r
o
m
a
clin
ical
p
o
i
n
t
o
f
v
iew,
as
ad
d
itio
n
al
m
is
d
iag
n
o
s
es
ca
n
lead
t
o
u
n
n
ec
ess
ar
y
p
atien
t
tr
ea
tm
e
n
t,
a
d
d
itio
n
al
s
u
f
f
e
r
in
g
,
an
d
i
n
cr
ea
s
ed
d
iag
n
o
s
tic
co
s
ts
.
Hen
ce
,
th
e
h
ig
h
ac
cu
r
ac
y
o
f
Y
OL
Ov
1
1
en
s
u
r
es m
o
r
e
r
eliab
l
e
an
d
ef
f
ec
tiv
e
d
etec
tio
n
i
n
m
ed
ical
s
y
s
tem
s
.
I
n
f
r
ac
tu
r
e
d
et
ec
tio
n
,
'
r
ec
all'
r
ef
er
s
to
h
o
w
ef
f
icien
tly
t
h
e
m
o
d
el
ca
n
d
etec
t
ac
tu
al
f
r
ac
t
u
r
es.
T
h
e
r
ec
all
o
f
R
esNet5
0
(
0
.
4
3
)
an
d
YOL
Ov
1
0
(
0
.
5
8
)
is
r
elativ
el
y
lo
w,
wh
ich
m
ea
n
s
th
at
th
ese
m
o
d
els
ca
n
o
f
ten
m
is
s
f
r
ac
tu
r
es.
Fals
e
n
eg
ativ
e
r
esu
lts
ar
e
p
ar
ticu
lar
ly
d
an
g
er
o
u
s
b
ec
au
s
e
if
a
f
r
ac
tu
r
e
is
p
r
esen
t,
if
it
g
o
es
u
n
d
etec
ted
,
th
e
p
atien
t'
s
tr
ea
t
m
en
t
m
ay
b
e
d
elay
ed
,
th
e
b
o
n
e
m
ay
n
o
t
r
o
tate
p
r
o
p
er
ly
,
o
r
l
o
n
g
-
ter
m
f
u
n
ctio
n
al
lo
s
s
m
ay
o
cc
u
r
.
(
a)
(
b
)
Fig
u
r
e
4
.
T
r
ain
in
g
g
r
ap
h
with
2
0
0
ep
o
ch
s
b
ased
o
n
(
a
)
YOL
O1
1
an
d
(
b
)
YOL
Ov
1
0
T
ab
le
6
.
T
esti
n
g
p
er
f
o
r
m
an
ce
o
f
YOL
Ov
1
1
with
YOL
Ov
1
0
,
YOL
Ov
9
,
YOL
Ov
8
an
d
E
f
f
ic
ien
tNetB
0
M
o
d
e
l
Ep
o
c
h
C
l
a
s
s
Tr
a
i
n
a
b
l
e
p
a
r
a
m
e
t
e
r
s
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
mA
P
@
0
.
5
Y
O
LO
v
1
0
1
0
0
A
l
l
2
.
7
1
M
0
.
8
0
0
.
5
8
0
.
3
7
0
.
3
4
7
2
0
0
0
.
9
2
0
.
7
8
0
.
4
5
0
.
4
4
7
Y
O
LO
v
9
1
0
0
A
l
l
2
5
.
4
5
M
0
.
8
6
0
.
7
0
0
.
3
6
0
.
3
5
2
2
0
0
0
.
9
4
0
.
8
0
0
.
4
4
0
.
4
5
2
Y
O
LO
v
8
1
0
0
A
l
l
2
5
.
8
6
M
0
.
8
5
0
.
6
0
0
.
4
4
0
.
4
7
2
2
0
0
0
.
9
4
0
.
7
0
0
.
5
0
0
.
4
8
8
Ef
f
i
c
i
e
n
t
N
e
t
B
0
1
0
0
A
l
l
4
.
0
2
M
0
.
6
5
0
.
6
3
0
.
6
3
0
.
7
0
0
2
0
0
0
.
9
4
0
.
9
3
0
.
9
3
0
.
9
7
9
P
r
o
p
o
se
d
Y
O
LO
v
11
1
0
0
A
l
l
9
.
4
3
M
0
.
9
0
0
.
6
0
0
.
3
8
0
.
3
5
2
2
0
0
0
.
9
9
0
.
7
0
0
.
4
6
0
.
4
6
2
T
ab
le
7
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
o
f
YOL
Ov
1
1
with
Den
s
eNe
t1
6
9
an
d
R
esNet5
0
M
o
d
e
l
Ep
o
c
h
Tr
a
i
n
a
b
l
e
p
a
r
a
m
e
t
e
r
s
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
mA
P
@
0
.
5
P
r
o
p
o
se
d
(
Y
O
LO
v
11)
D
e
n
seN
e
t
1
6
9
2
0
0
2
0
0
0
.
4
3
M
1
2
.
5
1
M
0
.
9
9
0
.
5
6
0
.
7
0
0
.
5
6
0
.
4
6
0
.
5
5
0
.
4
6
2
0
.
5
7
5
R
e
sN
e
t
5
0
2
0
0
2
3
.
5
7
M
0
.
4
3
0
.
4
3
0
.
4
2
0
.
4
0
2
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
C
o
mp
a
r
a
tive
a
n
a
lysi
s
o
f YOL
O
va
r
ia
n
ts
a
n
d
E
fficien
tN
et
fo
r
d
etec
tin
g
b
o
n
e
fr
a
ctu
r
es
…
(
S
h
a
ta
b
d
i S
a
r
ke
r
)
163
On
th
e
o
th
er
h
an
d
,
th
e
YOL
Ov
1
1
m
o
d
el
s
h
o
ws
h
ig
h
p
r
ec
is
io
n
(
0
.
9
9
)
an
d
m
o
d
er
ate
r
ec
all
(
0
.
7
0
)
,
in
d
icatin
g
a
g
o
o
d
b
ala
n
ce
.
T
h
is
m
ea
n
s
th
e
m
o
d
el
is
a
b
le
to
d
etec
t
m
o
s
t
f
r
ac
tu
r
es,
an
d
th
e
r
ate
o
f
f
alse
p
o
s
itiv
es is
lo
w.
Alth
o
u
g
h
E
f
f
icien
tNetB
0
s
h
o
ws
a
r
ec
all
o
f
0
.
9
3
,
its
lo
w
p
r
ec
is
io
n
in
cr
ea
s
es th
e
p
o
s
s
ib
ilit
y
o
f
f
alse
p
o
s
itiv
es,
i.e
.
,
s
o
m
etim
es
th
e
m
o
d
el
m
ay
f
alsely
r
e
p
o
r
t a
f
r
ac
tu
r
e
wh
e
n
th
er
e
is
n
o
f
r
a
ctu
r
e.
I
n
s
u
m
m
ar
y
,
r
ed
u
cin
g
f
alse
n
eg
ativ
es
is
p
ar
am
o
u
n
t
i
n
f
r
ac
tu
r
e
d
etec
tio
n
,
as
th
e
clin
ical
r
is
k
o
f
m
is
s
i
n
g
a
f
r
ac
tu
r
e
,
ev
en
i
f
p
r
esen
t,
is
h
ig
h
.
I
n
th
is
r
esp
ec
t,
YOL
Ov
1
1
m
ay
p
r
o
v
e
to
b
e
a
s
af
e
an
d
r
eliab
le
s
y
s
tem
f
o
r
clin
ical
u
s
e
d
u
e
to
its
h
ig
h
p
r
ec
is
io
n
an
d
g
o
o
d
r
e
ca
ll.
4
.
5
.
Vis
ua
liza
t
io
n
T
h
e
m
u
lti
-
class
f
r
ac
tu
r
e
d
etec
tio
n
p
er
f
o
r
m
an
ce
o
f
th
e
YOL
Ov
1
1
-
b
as
ed
m
o
d
el
is
s
h
o
wn
in
Fig
u
r
e
5
.
Fig
u
r
e
5
(
a)
s
h
o
ws
a
co
n
f
u
s
i
o
n
m
atr
ix
,
s
h
o
win
g
th
e
co
r
r
ec
t
an
d
in
co
r
r
ec
t
p
r
e
d
ictio
n
s
o
f
th
e
m
o
d
el
f
o
r
1
7
f
r
ac
tu
r
e
ty
p
es.
E
n
tr
ies
in
o
r
ig
in
ally
d
ia
g
o
n
al
lin
es
ar
e
c
o
r
r
ec
t
p
r
e
d
ictio
n
s
o
f
th
e
m
o
d
el,
an
d
d
ash
ed
lin
es
in
d
icate
m
is
class
if
icatio
n
s
.
So
m
e
class
es,
s
u
ch
as
'
s
p
ir
al
f
r
ac
tu
r
e'
an
d
'
s
eg
m
en
tal
f
r
ac
tu
r
e'
,
ar
e
m
o
r
e
ch
allen
g
in
g
to
i
d
en
tify
,
b
u
t
th
e
m
o
d
el
s
h
o
wed
g
o
o
d
ac
c
u
r
ac
y
in
class
es
s
u
ch
as
'
co
m
p
r
ess
io
n
-
cr
u
s
h
f
r
ac
tu
r
e'
(
0
.
7
0
)
a
n
d
'
o
p
en
co
m
p
o
u
n
d
f
r
ac
tu
r
e'
(
0
.
6
2
)
.
Fig
u
r
e
5
(
b
)
s
h
o
ws
th
e
f
r
ac
tu
r
e
lo
ca
lizatio
n
o
f
th
e
m
o
d
el
in
th
e
X
-
r
ay
im
ag
e
.
Her
e,
th
e
b
o
u
n
d
in
g
b
o
x
es
an
d
co
n
f
id
e
n
ce
s
co
r
es
in
d
icate
th
e
co
r
r
ec
t
id
en
tif
icatio
n
o
f
d
if
f
e
r
en
t
f
r
ac
tu
r
e
ty
p
es.
T
h
ese
v
is
u
aliza
tio
n
s
s
h
o
w
th
at
th
e
m
o
d
el
is
n
o
t
o
n
ly
ab
le
to
d
etec
t
f
r
ac
tu
r
es
,
b
u
t
also
p
in
p
o
in
t
ex
ac
tly
wh
er
e
th
e
y
ar
e.
O
n
th
e
o
n
e
h
a
n
d
,
it
s
h
o
ws
th
e
l
o
ca
lizatio
n
ca
p
ab
ilit
y
;
o
n
th
e
o
t
h
e
r
h
an
d
,
it
h
ig
h
lig
h
ts
th
e
p
o
wer
o
f
th
e
m
o
d
el
in
class
if
icatio
n
an
d
th
e
o
p
p
o
r
tu
n
ity
to
r
ed
u
ce
m
is
id
en
tific
atio
n
.
(
a)
(
b
)
Fig
u
r
e
5.
YOL
Ov
1
1
of
(
a)
c
o
n
f
u
s
io
n
m
atr
ix
a
n
d
(
b
)
test
in
g
p
er
f
o
r
m
a
n
ce
o
f
r
an
d
o
m
d
ata
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
5
,
No
.
1
,
Ma
r
c
h
2
0
2
6
:
155
-
1
6
7
164
Fig
u
r
e
6
s
h
o
ws
h
o
w
well
th
e
p
r
o
p
o
s
ed
YOL
Ov
1
1
m
o
d
el
p
er
f
o
r
m
e
d
f
o
r
1
7
f
r
ac
t
u
r
e
ty
p
es.
Fig
u
r
e
6
(
a)
s
h
o
ws
th
at
th
e
m
o
d
el
p
r
o
d
u
ce
s
v
er
y
ac
cu
r
ate
r
esu
lts
at
h
ig
h
co
n
f
id
en
ce
,
in
d
icatin
g
lo
w
f
alse
p
o
s
itiv
es.
T
h
e
F1
-
co
n
f
id
e
n
ce
cu
r
v
e
in
Fig
u
r
e
6
(
b
)
s
h
o
ws
an
o
v
er
all
F1
-
s
co
r
e
o
f
ab
o
u
t
0
.
4
6
,
in
d
icatin
g
a
b
alan
ce
d
p
e
r
f
o
r
m
an
ce
o
f
th
e
m
o
d
el.
T
h
e
r
ec
all
-
co
n
f
id
en
ce
c
u
r
v
e
in
Fig
u
r
e
6
(
c
)
s
h
o
ws
th
at
th
e
m
o
d
el'
s
r
ec
all
is
ab
o
u
t
0
.
7
0
,
a
n
d
th
e
r
ec
all
d
ec
r
ea
s
es
with
in
cr
ea
s
in
g
c
o
n
f
id
en
ce
–
m
ea
n
in
g
th
at
s
o
m
e
f
r
ac
tu
r
es
m
a
y
b
e
m
is
s
ed
wh
ile
p
r
ed
ictin
g
m
o
r
e
r
eliab
ly
.
T
h
e
m
AP@
0
.
5
v
alu
e
in
Fig
u
r
e
6
(
d
)
is
0
.
4
6
2
,
wh
ich
s
h
o
ws
a
r
esp
ec
ta
b
le
b
alan
ce
b
etwe
en
p
r
ec
is
io
n
an
d
r
ec
all.
I
n
s
u
m
m
a
r
y
,
th
e
YOL
Ov
1
1
m
o
d
el
is
r
e
liab
le
an
d
s
tab
le
in
d
etec
tin
g
d
if
f
er
en
t
t
y
p
es
o
f
f
r
ac
tu
r
es,
wh
ich
m
ak
es
it
s
u
it
ab
le
f
o
r
au
to
m
atic
f
r
ac
tu
r
e
d
etec
tio
n
in
m
e
d
ical
X
-
r
ay
im
ag
es.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
6
.
YOL
Ov
1
1
-
b
ased
of
(
a)
p
r
ec
is
io
n
c
u
r
v
e
,
(
b
)
F1
-
s
co
r
e
cu
r
v
e
,
(
c)
r
ec
all
cu
r
v
e
,
an
d
(
d
)
m
AP c
u
r
v
e
Fig
u
r
e
7
co
m
p
a
r
es
th
e
p
r
ed
i
ctio
n
p
er
f
o
r
m
a
n
ce
o
f
f
i
v
e
m
o
d
els
–
YOL
Ov
1
1
,
YOL
Ov
1
0
,
YOL
Ov
9
,
YOL
Ov
8
,
an
d
E
f
f
icien
tNetB
0
–
o
n
a
1
7
-
class
f
r
ac
tu
r
e
d
etec
tio
n
d
ataset.
E
ac
h
r
o
w
s
h
o
ws
a
f
r
ac
tu
r
e
class
,
an
d
ea
ch
co
lu
m
n
ex
p
r
ess
es
th
e
p
r
ed
ictio
n
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el.
T
h
e
r
esu
lts
s
h
o
w
th
at
YOL
Ov
1
1
o
u
tp
er
f
o
r
m
s
th
e
o
t
h
er
m
o
d
el
s
in
all
class
e
s
an
d
ac
h
iev
es
th
e
h
ig
h
est
o
v
er
all
ac
cu
r
ac
y
.
T
h
is
p
r
o
v
es
th
at
YOL
Ov
1
1
is
p
ar
ticu
lar
ly
ca
p
a
b
le
o
f
d
etec
tin
g
co
m
p
lex
f
r
ac
tu
r
es,
esp
ec
ially
in
ter
m
s
o
f
p
r
e
cisi
o
n
an
d
r
ec
all.
Fig
u
r
e
7
.
Sam
p
le
d
etec
ted
im
a
g
es f
o
r
YOL
Ov
1
1
,
YOL
Ov
1
0
,
YOL
Ov
9
,
YOL
Ov
8
,
a
n
d
E
f
f
icien
tNetB
0
r
esp
ec
tiv
ely
Evaluation Warning : The document was created with Spire.PDF for Python.