I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
16
,
No
.
5
,
Octo
b
er
20
26
,
p
p
.
2
8
0
6
~
2
8
1
8
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
16
i
5
.
pp
2
8
0
6
-
2
8
1
8
2806
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Dete
c
tion a
nd
cla
ss
ificatio
n of
thyr
o
id disea
ses
using
ult
ra
so
und
ima
g
es throug
h d
eep learning
tech
niques
K
a
m
a
l Sub
edi
1
,
Ak
a
s
h K
um
a
r
Ya
da
v
2
,
L
o
cha
n P
a
ud
el
3
1
S
c
h
o
o
l
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
A
p
p
l
i
e
d
S
c
i
e
n
c
e
,
Th
e
G
e
o
r
g
e
W
a
sh
i
n
g
t
o
n
U
n
i
v
e
r
si
t
y
,
W
a
s
h
i
n
g
t
o
n
,
D
C
,
U
n
i
t
e
d
S
t
a
t
e
s
2
G
r
a
d
u
a
t
e
S
c
h
o
o
l
o
f
A
r
t
s
a
n
d
S
c
i
e
n
c
e
s,
F
o
r
d
h
a
m U
n
i
v
e
r
s
i
t
y
,
N
e
w
Y
o
r
k
C
i
t
y
,
U
n
i
t
e
d
S
t
a
t
e
s
3
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
M
a
c
h
i
n
e
L
e
a
r
n
i
n
g
D
e
p
a
r
t
m
e
n
t
,
S
y
m
b
i
o
si
s
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
P
u
n
e
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
2
1
,
2
0
2
6
R
ev
is
ed
J
u
l 2
9
,
2
0
2
6
Acc
ep
ted
Au
g
1
9
,
2
0
2
6
Th
y
r
o
id
d
iso
r
d
e
rs
c
o
n
ti
n
u
e
t
o
b
e
a
m
a
jo
r
g
lo
b
a
l
p
u
b
li
c
h
e
a
lt
h
c
o
n
c
e
rn
,
wit
h
u
lt
ra
so
u
n
d
sc
re
e
n
in
g
s
in
d
ica
ti
n
g
th
a
t
t
h
y
r
o
id
n
o
d
u
les
a
ffe
c
t
ro
u
g
h
ly
3
0
%
t
o
5
0
%
o
f
a
d
u
lt
s.
Id
e
n
ti
fy
i
n
g
t
h
e
sp
e
c
ifi
c
n
a
tu
re
o
f
t
h
e
se
n
o
d
u
les
is
c
ru
c
ial
fo
r
e
ffe
c
ti
v
e
trea
tme
n
t,
a
s
c
a
t
c
h
in
g
m
a
li
g
n
a
n
t
n
o
d
u
les
e
a
rly
lea
d
s
to
b
e
tt
e
r
o
u
tco
m
e
s.
A
k
e
y
c
h
a
ll
e
n
g
e
is
th
a
t
b
e
n
ig
n
a
n
d
m
a
li
g
n
a
n
t
n
o
d
u
les
o
ften
lo
o
k
v
isu
a
ll
y
sim
il
a
r
i
n
u
lt
ra
so
u
n
d
sc
a
n
s.
Co
n
se
q
u
e
n
tl
y
,
t
h
is p
a
p
e
r
p
re
s
e
n
ts a t
wo
-
p
h
a
se
d
e
e
p
lea
rn
in
g
a
p
p
ro
a
c
h
f
o
r
th
e
a
u
t
o
m
a
ted
d
e
tec
ti
o
n
a
n
d
c
las
sifica
ti
o
n
o
f
th
y
ro
i
d
n
o
d
u
les
in
u
lt
ra
so
u
n
d
sc
a
n
s.
In
it
ially
,
t
h
e
fra
m
e
wo
rk
u
se
s
Atten
ti
o
n
U
-
Ne
t
se
g
m
e
n
tatio
n
t
o
lo
c
a
te
th
e
re
g
i
o
n
o
f
i
n
tere
st
(R
OI)
fo
r
t
h
e
n
o
d
u
les
.
E
x
p
e
rt
-
d
ra
wn
p
i
x
e
l
-
lev
e
l
a
n
n
o
tatio
n
s
o
f
t
h
y
r
o
id
n
o
d
u
les
we
re
n
o
t
a
v
a
il
a
b
le;
h
e
n
c
e
,
p
se
u
d
o
-
lab
e
ls w
e
re
u
se
d
in
train
in
g
.
Th
e
y
we
re
o
b
tain
e
d
b
y
a
p
p
ly
in
g
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
(CLA
HE)
,
Otsu
th
re
sh
o
l
d
i
n
g
,
a
n
d
m
o
rp
h
o
lo
g
ica
l
o
p
e
ra
ti
o
n
s.
S
u
b
se
q
u
e
n
tl
y
,
a
tailo
re
d
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(
C
NN
)
wa
s
u
ti
li
z
e
d
to
c
a
teg
o
rize
th
e
e
x
trac
ted
th
y
r
o
id
a
re
a
s
a
s
e
it
h
e
r
b
e
n
ig
n
o
r
m
a
li
g
n
a
n
t.
Da
ta
a
u
g
m
e
n
tatio
n
u
sin
g
t
h
e
M
ix
U
p
a
n
d
C
u
tM
i
x
tec
h
n
i
q
u
e
s
h
e
lp
e
d
re
d
u
c
e
t
h
e
risk
o
f
o
v
e
rfit
ti
n
g
,
e
n
h
a
n
c
in
g
t
h
e
m
o
d
e
l'
s
g
e
n
e
ra
li
z
a
b
il
it
y
.
T
h
is
e
x
p
e
rime
n
t
wa
s
tes
ted
o
n
4
8
0
u
lt
ra
so
u
n
d
ima
g
e
s
g
a
th
e
re
d
fro
m
two
tea
c
h
in
g
h
o
s
p
it
a
ls
in
N
e
p
a
l.
Th
e
p
ro
p
o
se
d
fra
m
e
wo
rk
sh
o
ws
th
a
t
se
g
m
e
n
tatio
n
wit
h
we
a
k
su
p
e
rv
isio
n
a
lo
n
g
with
c
las
sifica
ti
o
n
u
sin
g
d
e
e
p
l
e
a
rn
in
g
c
a
n
p
ro
v
i
d
e
e
fficie
n
t
d
i
a
g
n
o
sis
o
f
th
y
r
o
id
n
o
d
u
les
e
v
e
n
with
m
i
n
ima
l
m
e
d
ica
l
d
a
ta
a
n
n
o
tati
o
n
.
Va
li
d
a
ti
o
n
u
sin
g
a
3
-
f
o
ld
stra
ti
f
ied
c
ro
ss
-
v
a
li
d
a
ti
o
n
m
e
th
o
d
d
e
m
o
n
stra
ted
th
a
t
t
h
e
Atten
ti
o
n
U
-
Ne
t
se
g
m
e
n
tatio
n
y
i
e
ld
e
d
a
Dic
e
sc
o
re
o
f
0
.
8
7
7
2
a
n
d
a
n
Io
U
o
f
0
.
7
8
7
3
.
M
e
a
n
wh
il
e
,
th
e
c
las
sifica
ti
o
n
c
o
m
p
o
n
e
n
t
re
a
c
h
e
d
a
n
o
v
e
ra
ll
a
c
c
u
ra
c
y
o
f
9
6
.
7
3
%
,
wit
h
a
se
n
sit
iv
it
y
o
f
5
8
.
3
3
%
,
sp
e
c
ifi
c
it
y
o
f
9
7
.
8
9
%
,
a
n
d
a
m
e
a
n
ROC
-
AU
C
o
f
9
4
.
6
0
%
.
K
ey
w
o
r
d
s
:
Atten
tio
n
U
-
Net
No
d
u
le
s
eg
m
en
tatio
n
Ots
u
th
r
esh
o
ld
in
g
T
h
y
r
o
i
d
d
is
ea
s
e
d
etec
tio
n
Ultr
aso
u
n
d
im
ag
e
a
n
aly
s
is
W
ea
k
s
eg
m
en
tatio
n
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
:
Kam
al
Su
b
ed
i
Sch
o
o
l o
f
E
n
g
in
ee
r
in
g
a
n
d
Ap
p
lied
Scien
ce
,
T
h
e
Geo
r
g
e
W
a
s
h
in
g
to
n
Un
iv
e
r
s
ity
W
ash
in
g
to
n
,
DC
,
Un
ited
State
E
m
ail: K
am
al.
s
u
b
ed
i@
g
wu
.
ed
u
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
th
y
r
o
id
g
lan
d
is
a
s
m
all,
b
u
tter
f
ly
-
s
h
ap
e
d
g
lan
d
th
at
is
lo
ca
ted
in
th
e
an
ter
i
o
r
r
e
g
io
n
o
f
th
e
n
ec
k
.
T
h
e
th
y
r
o
id
g
lan
d
p
r
o
d
u
ce
s
h
o
r
m
o
n
es
wh
o
s
e
f
u
n
ctio
n
is
to
r
eg
u
late
th
e
d
if
f
er
en
t
p
r
o
ce
s
s
es
in
th
e
b
o
d
y
s
u
ch
as
m
etab
o
lis
m
,
h
ea
r
t
r
ate,
an
d
b
o
d
y
tem
p
e
r
atu
r
e.
W
h
en
t
h
er
e
ar
e
s
o
m
e
a
b
n
o
r
m
alities
with
r
eg
ar
d
s
to
th
e
th
y
r
o
id
g
lan
d
,
th
e
n
d
if
f
er
e
n
t
d
is
ea
s
es
will
b
e
ex
p
e
r
ien
ce
d
,
in
clu
d
in
g
h
y
p
er
th
y
r
o
id
is
m
,
h
y
p
o
th
y
r
o
id
is
m
,
G
o
iter
,
an
d
th
y
r
o
id
ca
n
ce
r
[
1
]
.
As o
b
s
er
v
ed
f
r
o
m
th
e
latest s
tu
d
ies,
th
y
r
o
i
d
n
o
d
u
les ar
e
f
o
u
n
d
in
b
etwe
en
3
0
%
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Dete
ctio
n
a
n
d
cla
s
s
ifica
tio
n
o
f th
yro
id
d
is
ea
s
es u
s
in
g
u
ltr
a
s
o
u
n
d
…
(
K
a
ma
l
S
u
b
e
d
i
)
2807
an
d
5
0
%
o
f
th
e
t
o
tal
n
u
m
b
er
o
f
p
e
o
p
le
th
r
o
u
g
h
u
ltra
s
o
u
n
d
s
cr
ee
n
in
g
test
s
.
C
an
ce
r
is
p
r
es
en
t
in
7
%
to
1
5
%
o
f
th
e
th
y
r
o
i
d
n
o
d
u
les
[
2
]
.
Ultr
aso
u
n
d
is
th
e
m
o
s
t
f
r
eq
u
e
n
t
an
d
p
r
im
ar
y
m
o
d
ality
f
o
r
th
e
d
iag
n
o
s
is
o
f
th
y
r
o
id
n
o
d
u
les
d
u
e
to
its
s
af
ety
,
co
s
t
-
ef
f
ec
tiv
en
ess
,
an
d
lack
o
f
io
n
izin
g
r
ad
iatio
n
[
3
]
.
Ultr
aso
u
n
d
im
a
g
es
g
en
er
ally
co
n
tain
p
o
o
r
co
n
tr
ast,
s
p
ec
k
le
n
o
is
e,
an
d
u
n
clea
r
b
o
u
n
d
ar
ies,
m
ak
in
g
it
d
i
f
f
icu
lt
f
o
r
p
h
y
s
ician
s
to
m
ak
e
a
p
r
o
p
e
r
d
iag
n
o
s
is
.
T
h
e
Am
er
ican
T
h
y
r
o
id
Ass
o
ciatio
n
(
AT
A)
h
as
p
r
o
v
id
e
d
d
iag
n
o
s
tic
cr
iter
ia
f
o
r
th
y
r
o
id
n
o
d
u
les
u
s
in
g
u
ltra
s
o
u
n
d
ch
a
r
ac
ter
is
tics
s
u
c
h
as
s
h
ap
e,
ec
h
o
g
en
icity
,
an
d
m
ar
g
in
s
[
4
]
.
Desp
ite
th
e
p
r
esen
ce
o
f
g
u
id
elin
es,
th
er
e
ar
e
m
ajo
r
d
is
cr
ep
an
cie
s
in
th
e
in
ter
p
r
etatio
n
s
o
f
u
l
tr
aso
u
n
d
im
ag
es
b
y
d
if
f
er
en
t
p
h
y
s
ician
s
,
an
d
a
d
iag
n
o
s
is
is
g
en
er
ally
m
a
d
e
o
n
ly
af
ter
a
f
in
e
-
n
ee
d
le
asp
ir
atio
n
b
io
p
s
y
,
wh
ich
is
an
i
n
v
asiv
e
p
r
o
ce
s
s
.
Dee
p
lear
n
in
g
h
as b
ec
o
m
e
an
ef
f
ec
tiv
e
m
eth
o
d
f
o
r
an
aly
zin
g
m
ed
ical
im
a
g
es
in
t
h
e
p
ast
d
ec
ad
e.
T
h
e
d
ee
p
lear
n
in
g
f
r
a
m
ewo
r
k
,
esp
ec
ially
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
etwo
r
k
(
C
NN)
ar
ch
itectu
r
e,
is
ab
le
to
lear
n
u
s
ef
u
l
f
ea
tu
r
es
d
ir
ec
tly
f
r
o
m
im
ag
es
r
ath
er
th
an
u
s
in
g
m
an
u
al
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
es
[
2
]
.
T
h
er
e
h
a
v
e
b
ee
n
s
ev
er
al
in
s
tan
ce
s
wh
er
e
d
ee
p
lear
n
in
g
alg
o
r
ith
m
s
h
av
e
b
ee
n
em
p
lo
y
ed
f
o
r
th
y
r
o
id
n
o
d
u
l
e
class
if
icatio
n
an
d
ac
h
iev
ed
s
atis
f
ac
to
r
y
p
er
f
o
r
m
an
ce
.
Fo
r
i
n
s
tan
ce
,
Kim
et
a
l.
[
5
]
d
e
v
elo
p
e
d
m
o
d
els
u
s
in
g
th
e
VGG1
6
,
VGG1
9
,
an
d
R
esNet
ar
ch
itectu
r
es
o
n
a
d
ataset
o
f
m
o
r
e
th
an
1
5
,
0
0
0
t
h
y
r
o
i
d
u
ltra
s
o
u
n
d
im
a
g
es
an
d
o
b
tain
ed
co
m
p
ar
ab
le
ac
c
u
r
ac
y
lev
els
to
ex
p
er
ien
ce
d
r
ad
io
lo
g
is
ts
f
o
r
class
if
icatio
n
task
o
f
b
en
ig
n
an
d
m
alig
n
a
n
t
n
o
d
u
les.
L
ik
ewise,
Das
et
a
l.
[
2
]
co
n
d
u
cted
an
ex
h
a
u
s
tiv
e
r
ev
iew
s
h
o
win
g
th
e
ef
f
icac
y
o
f
d
ee
p
lear
n
in
g
-
b
ased
alg
o
r
ith
m
s
in
t
h
y
r
o
id
n
o
d
u
le
an
aly
s
is
.
T
h
e
ea
r
lier
m
eth
o
d
s
o
f
class
if
icatio
n
o
f
th
e
d
is
ea
s
e
wer
e
b
ased
o
n
m
an
u
ally
d
esig
n
e
d
tex
tu
r
e
f
ea
tu
r
es a
s
well
as c
las
s
ical
m
ac
h
in
e
lear
n
in
g
class
if
ier
s
.
T
h
e
ap
p
r
o
ac
h
o
f
Nu
g
r
o
h
o
et
a
l.
[
1
]
was b
ased
o
n
t
h
e
u
s
e
o
f
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
(
SVM
)
with
tex
tu
r
e
f
ea
tu
r
es
in
o
r
d
er
to
ac
h
iev
e
th
e
ac
cu
r
ac
y
o
f
8
9
%
in
class
if
icatio
n
o
f
th
y
r
o
id
u
ltra
s
o
u
n
d
i
m
ag
es.
Ma
h
m
o
o
d
ian
et
a
l.
[
3
]
p
r
o
p
o
s
ed
an
ap
p
r
o
ac
h
b
ased
o
n
th
e
ex
tr
ac
tio
n
o
f
f
ea
tu
r
es
u
s
in
g
b
is
p
ec
tr
u
m
with
th
e
r
esu
ltin
g
ac
cu
r
ac
y
o
f
9
3
.
2
7
%
in
class
i
f
icatio
n
o
f
th
y
r
o
id
tex
tu
r
es.
Similar
ly
,
Kale
et
a
l.
[
6
]
em
p
lo
y
ed
g
r
ay
-
lev
el
c
o
-
o
cc
u
r
r
en
ce
m
atr
i
x
(
GL
C
M)
f
ea
tu
r
es
an
d
SVM
in
o
r
d
er
to
ac
h
iev
e
8
2
.
2
8
%
ac
cu
r
ac
y
.
T
h
e
p
o
ten
tial
o
f
th
e
ap
p
l
icatio
n
o
f
tex
tu
r
e
an
aly
s
is
u
s
i
n
g
u
ltra
s
o
u
n
d
s
h
ea
r
wav
e
elasto
g
r
ap
h
y
to
p
r
e
d
ict
th
e
o
cc
u
r
r
e
n
ce
o
f
m
alig
n
an
t
th
y
r
o
id
n
o
d
u
les
v
ia
th
e
u
s
e
o
f
SVM,
k
-
n
ea
r
est
n
eig
h
b
o
r
s
(
KNN)
,
an
d
B
ay
esian
class
if
ier
s
was
s
tu
d
ied
b
y
Sin
g
h
et
a
l.
[
7
]
.
B
ib
icu
et
a
l.
[
8
]
u
s
ed
s
tatis
tical
f
ir
s
t
o
r
d
er
f
ea
t
u
r
es
in
o
r
d
er
t
o
d
etec
t
9
0
%
o
f
th
e
s
p
ec
if
icit
y
,
wh
er
ea
s
Ab
b
asian
et
a
l.
[
9
]
u
s
ed
2
7
0
f
ea
tu
r
es
f
r
o
m
u
ltra
s
o
u
n
d
im
a
g
e
tex
tu
r
es
in
o
r
d
er
to
ac
h
iev
e
9
7
.
1
4
%
ac
cu
r
ac
y
o
n
7
0
im
a
g
es
in
to
tal.
Fu
r
th
er
m
o
r
e,
T
y
ag
i
et
a
l.
[
1
0
]
d
ev
elo
p
ed
a
n
in
ter
ac
tiv
e
a
p
p
licatio
n
f
o
r
t
h
e
p
r
e
d
ictio
n
o
f
th
y
r
o
id
d
is
e
ase
u
s
in
g
m
ac
h
in
e
lear
n
in
g
tech
n
iq
u
es.
Dee
p
lear
n
in
g
h
as
b
ee
n
v
e
r
y
ef
f
ec
tiv
e
in
ad
v
an
cin
g
th
y
r
o
id
d
is
ea
s
e
clas
s
if
icatio
n
.
Kim
et
a
l.
[
5
]
b
u
ilt
m
o
d
els
u
s
in
g
VGG1
6
,
VGG1
9
,
an
d
R
esNet
m
o
d
els
o
n
a
lar
g
e
s
ca
le
d
ataset,
a
ch
iev
in
g
a
s
im
ilar
ac
cu
r
ac
y
as
t
h
at
o
f
p
r
o
f
ess
io
n
al
r
ad
io
l
o
g
is
ts
.
C
h
i
et
a
l.
[
1
1
]
f
in
e
-
tu
n
ed
a
p
r
e
-
t
r
ain
ed
Go
o
g
L
eNe
t
m
o
d
el,
wh
ile
Av
er
s
an
o
et
a
l.
[
1
2
]
f
u
r
th
er
tr
ain
ed
th
e
R
esNet
-
1
8
n
eu
r
al
n
etwo
r
k
f
o
r
th
y
r
o
id
u
ltra
s
o
u
n
d
class
if
icatio
n
.
L
iu
et
a
l.
[
1
3
]
p
er
f
o
r
m
e
d
a
tr
an
s
f
er
lear
n
in
g
an
d
h
y
b
r
id
f
ea
tu
r
e
s
tu
d
y
o
n
th
y
r
o
id
u
ltra
s
o
u
n
d
im
a
g
es
an
d
Ab
d
o
lali
et
a
l.
[
1
4
]
p
r
esen
t
ed
an
a
u
to
m
ated
th
y
r
o
id
d
e
tectio
n
an
d
class
if
icatio
n
s
y
s
tem
u
s
in
g
C
NNs.
Gu
an
et
a
l.
[
1
5
]
p
er
f
o
r
m
ed
a
lar
g
e
-
s
ca
le
p
ilo
t
s
tu
d
y
o
n
th
y
r
o
id
n
o
d
u
le
class
if
icatio
n
th
r
o
u
g
h
d
ee
p
lear
n
i
n
g
.
R
ec
en
tly
,
Kim
et
a
l.
[
1
6
]
u
s
ed
R
esNet5
0
,
Den
s
eNe
t2
0
1
,
an
d
E
f
f
icien
tNetV2
-
S d
ee
p
lear
n
i
n
g
ar
ch
itectu
r
es f
o
r
Mu
ltiv
iew
th
y
r
o
id
u
ltra
s
o
u
n
d
class
if
icatio
n
,
o
b
tain
in
g
class
if
icatio
n
ac
cu
r
ac
ies
o
f
8
1
%
-
8
3
%.
T
h
ese
s
tu
d
ies
h
ig
h
lig
h
t
th
e
ad
v
an
tag
es
o
f
l
ea
r
n
ed
f
ea
tu
r
es
o
v
er
m
an
u
ally
d
esig
n
ed
o
n
es,
h
o
wev
er
,
t
h
ey
r
eq
u
i
r
e
a
lar
g
e
am
o
u
n
t
o
f
lab
elled
d
atasets
an
d
d
ea
l w
ith
class
if
icatio
n
s
ep
ar
ately
f
r
o
m
s
eg
m
en
tatio
n
.
Fo
r
th
e
p
r
o
b
lem
o
f
s
eg
m
en
t
atio
n
,
th
e
U
-
Net
m
o
d
el
[
1
7
]
h
as
b
ec
o
m
e
h
ig
h
ly
p
o
p
u
lar
in
m
ed
ical
im
ag
in
g
d
u
e
t
o
its
ab
ilit
y
to
g
en
er
ate
ac
cu
r
ate
s
eg
m
en
tat
io
n
m
ap
s
e
v
en
with
r
elativ
e
ly
s
m
all
d
atasets
.
Atten
tio
n
U
-
Net
[
1
8
]
f
u
r
t
h
er
i
m
p
r
o
v
e
d
s
eg
m
e
n
tatio
n
p
er
f
o
r
m
an
ce
b
y
in
te
g
r
atin
g
atten
tio
n
g
ates
th
at
f
o
cu
s
o
n
r
elev
an
t
im
ag
e
r
eg
io
n
s
.
C
h
in
a
et
a
l.
[
1
9
]
p
r
esen
ted
a
n
ew
v
o
lu
m
etr
ic
s
eg
m
en
tatio
n
ap
p
r
o
ac
h
u
s
in
g
iter
ativ
e
r
an
d
o
m
walk
s
f
o
r
a
n
aly
zin
g
u
ltra
s
o
u
n
d
im
ag
es.
Yan
g
et
a
l.
[
2
0
]
i
n
tr
o
d
u
ce
d
a
m
u
ltit
ask
lear
n
in
g
f
r
am
ewo
r
k
f
o
r
th
y
r
o
id
n
o
d
u
le
s
eg
m
en
tatio
n
b
y
in
co
r
p
o
r
atin
g
th
y
r
o
id
r
eg
io
n
p
r
io
r
s
,
wh
ile
Do
n
g
et
a
l.
[
2
1
]
p
r
o
p
o
s
ed
a
d
u
al
-
p
ath
atten
tio
n
m
ec
h
an
is
m
with
UNe
t++
an
d
ac
h
iev
ed
a
Dice
s
co
r
e
o
f
0
.
8
3
1
0
.
Ho
w
ev
er
,
m
o
s
t
ex
is
tin
g
s
eg
m
en
tatio
n
ap
p
r
o
ac
h
es
s
till
r
ely
h
ea
v
ily
o
n
m
a
n
u
ally
a
n
n
o
tated
g
r
o
u
n
d
tr
u
th
im
a
g
es,
wh
ich
m
o
tiv
ates
o
u
r
p
s
eu
d
o
-
m
ask
a
p
p
r
o
ac
h
f
o
r
we
ak
ly
s
u
p
er
v
is
ed
lea
r
n
in
g
.
Ho
wev
er
,
th
er
e
ar
e
s
till
s
o
m
e
is
s
u
es
th
at
n
ee
d
to
b
e
ad
d
r
ess
ed
.
Firstl
y
,
m
o
s
t
d
ee
p
lear
n
in
g
m
o
d
el
s
cu
r
r
en
tly
u
s
ed
f
o
r
an
aly
zin
g
t
h
y
r
o
id
n
o
d
u
les
d
em
an
d
lar
g
e
am
o
u
n
ts
o
f
d
ata
with
p
ix
el
-
wi
s
e
an
n
o
tatio
n
s
m
ad
e
b
y
m
e
d
ical
p
r
o
f
ess
io
n
als.
T
h
i
s
p
r
o
ce
s
s
is
ex
tr
em
ely
lab
o
r
io
u
s
an
d
c
o
s
tly
;
th
er
ef
o
r
e,
it
ca
n
n
o
t
b
e
u
s
ed
i
n
th
e
m
ajo
r
ity
o
f
d
e
v
elo
p
in
g
c
o
u
n
tr
ies
wh
er
e
th
er
e
is
an
a
b
u
n
d
an
ce
o
f
p
atien
ts
with
th
y
r
o
id
d
is
ea
s
es
an
d
a
lack
o
f
q
u
alif
ied
p
r
o
f
ess
io
n
als.
Sec
o
n
d
ly
,
m
o
s
t
r
esear
c
h
er
s
r
e
g
ar
d
s
eg
m
e
n
tatio
n
a
n
d
class
if
icatio
n
as
two
in
d
ep
en
d
en
t
p
r
o
ce
s
s
es
with
o
u
t
u
tili
zin
g
th
e
c
o
n
n
ec
tio
n
b
et
wee
n
th
em
.
T
h
ir
d
ly
,
m
o
s
t
cu
r
r
en
t
d
ata
b
ases
ar
e
co
llected
in
h
ig
h
ly
e
q
u
ip
p
ed
h
o
s
p
itals
in
d
e
v
elo
p
ed
co
u
n
tr
ie
s
;
h
en
ce
,
it
is
d
if
f
icu
lt
to
d
r
aw
co
n
cl
u
s
io
n
s
ab
o
u
t
th
e
p
er
f
o
r
m
an
ce
o
f
th
ese
m
eth
o
d
s
o
n
d
ata
co
llected
in
s
m
all
h
o
s
p
itals
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
16
,
No
.
5
,
Octo
b
e
r
20
26
:
2
8
0
6
-
2
8
1
8
2808
T
h
is
p
ap
er
p
r
o
p
o
s
es
a
m
o
d
if
ied
ar
ch
itectu
r
e
to
im
p
lem
en
t
d
ee
p
lear
n
in
g
u
s
in
g
a
two
-
p
h
ase
tech
n
iq
u
e
t
h
at
ca
n
b
e
u
s
ed
to
o
v
er
co
m
e
th
e
p
r
o
b
lem
s
m
en
ti
o
n
ed
ea
r
lier
.
T
h
is
ar
ch
itectu
r
e
is
co
m
p
o
s
ed
o
f
two
m
o
d
u
les.
First,
th
e
a
tten
tio
n
U
-
Net
[
1
8
]
alg
o
r
ith
m
is
u
tili
ze
d
to
id
en
tify
th
e
ar
ea
o
f
in
ter
est
f
r
o
m
th
e
u
ltra
s
o
n
ic
im
ag
es.
B
ec
au
s
e
it
is
q
u
ite
ch
allen
g
in
g
to
g
ath
e
r
an
n
o
tate
d
d
ata
b
y
ex
p
er
ts
,
p
s
eu
d
o
-
m
ask
in
g
is
p
er
f
o
r
m
ed
u
s
in
g
s
ev
e
r
al
o
p
er
atio
n
s
,
in
clu
d
in
g
c
o
n
tr
ast
lim
ited
ad
ap
tiv
e
h
is
to
g
r
a
m
eq
u
a
lizatio
n
(
C
L
AHE
)
with
Ots
u
th
r
esh
o
ld
in
g
,
a
n
d
m
o
r
p
h
o
lo
g
ical
o
p
er
atio
n
s
.
Nex
t,
a
s
p
ec
ialized
C
NN
m
o
d
el
is
d
ep
lo
y
e
d
to
class
if
y
n
o
d
u
les
as
b
en
ig
n
o
r
m
alig
n
an
t.
Fu
r
th
er
m
o
r
e,
t
h
e
Mix
U
p
[
2
2
]
an
d
C
u
tMix
[
2
3
]
a
p
p
r
o
ac
h
es
ar
e
im
p
lem
en
ted
d
u
r
i
n
g
t
h
e
tr
ain
in
g
p
h
ase
to
r
ed
u
ce
o
v
er
f
itti
n
g
b
y
a
u
g
m
e
n
tin
g
s
am
p
les,
wh
ich
is
ex
tr
em
el
y
im
p
o
r
tan
t
d
u
e
to
th
e
s
ca
r
city
o
f
th
e
d
ataset.
T
h
is
ap
p
r
o
ac
h
was
d
ev
elo
p
ed
an
d
ev
alu
ated
o
n
4
8
0
u
ltra
s
o
u
n
d
im
ag
es c
o
llected
f
r
o
m
two
ed
u
ca
tio
n
al
h
o
s
p
itals
in
Nep
al.
Desp
ite
th
e
p
r
o
g
r
ess
in
th
y
r
o
i
d
n
o
d
u
le
an
aly
s
is
,
a
cr
itical
r
esear
ch
g
ap
r
em
ain
s
:
th
er
e
is
a
lack
o
f
m
eth
o
d
s
th
at
co
m
b
in
e
wea
k
ly
s
u
p
er
v
is
ed
s
eg
m
en
tatio
n
with
class
if
icatio
n
in
a
u
n
if
ied
p
ip
elin
e,
p
ar
ticu
lar
ly
f
o
r
s
ettin
g
s
wh
er
e
p
ix
el
-
lev
el
an
n
o
tatio
n
s
ar
e
u
n
a
v
ailab
le.
E
x
is
tin
g
m
eth
o
d
s
eith
er
r
e
q
u
i
r
e
ex
p
e
n
s
iv
e
ex
p
e
r
t
an
n
o
tatio
n
s
o
r
tr
ea
t
s
eg
m
en
tat
io
n
an
d
class
if
icatio
n
as
s
ep
ar
ate,
d
is
co
n
n
ec
ted
task
s
.
T
h
is
wo
r
k
ad
d
r
ess
es
th
i
s
g
ap
th
r
o
u
g
h
th
r
ee
m
ain
co
n
tr
ib
u
tio
n
s
:
i
)
a
u
n
if
ied
two
-
s
ta
g
e
f
r
am
ewo
r
k
th
at
i
n
teg
r
ates
wea
k
ly
s
u
p
er
v
is
ed
Atten
tio
n
U
-
Net
s
eg
m
en
tati
o
n
with
C
NN
-
b
ased
clas
s
if
i
ca
tio
n
,
en
ab
lin
g
th
y
r
o
id
n
o
d
u
le
d
etec
tio
n
an
d
m
alig
n
an
cy
class
if
icatio
n
with
o
u
t
p
ix
el
-
lev
el
an
n
o
tatio
n
s
;
ii
)
a
p
s
eu
d
o
-
m
ask
g
en
er
atio
n
m
ec
h
an
is
m
u
s
in
g
C
L
AHE
-
en
h
an
ce
d
p
r
ep
r
o
ce
s
s
in
g
co
m
b
in
ed
with
Ots
u
th
r
esh
o
ld
in
g
an
d
m
o
r
p
h
o
lo
g
ical
o
p
er
atio
n
s
,
p
r
o
v
id
in
g
ef
f
ec
tiv
e
s
p
atial
g
u
id
an
ce
f
o
r
th
e
s
eg
m
en
tatio
n
m
o
d
el;
an
d
iii
)
a
co
m
p
r
eh
en
s
iv
e
ev
alu
ati
o
n
o
n
4
8
0
th
y
r
o
id
u
ltra
s
o
u
n
d
im
ag
es
with
th
r
ee
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
.
T
h
e
s
ig
n
if
ican
ce
o
f
th
is
wo
r
k
lies
in
i
ts
p
o
ten
tial
to
m
a
k
e
AI
-
ass
is
ted
th
y
r
o
id
d
iag
n
o
s
i
s
ac
ce
s
s
ib
le
in
lo
w
-
r
eso
u
r
ce
s
ettin
g
s
wh
er
e
ex
p
er
t
a
n
n
o
tatio
n
s
an
d
lar
g
e
an
n
o
tated
d
atasets
ar
e
u
n
av
ail
ab
le.
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
d
escr
ib
es
t
h
e
p
r
o
p
o
s
ed
two
-
s
tag
e
f
r
am
ewo
r
k
in
d
etail.
T
h
e
e
n
d
-
to
-
en
d
p
ip
elin
e
is
illu
s
tr
ated
in
Fig
u
r
e
1.
Fig
u
r
e
1
.
B
lo
ck
d
iag
r
am
o
f
th
e
p
r
o
p
o
s
ed
th
y
r
o
id
n
o
d
u
le
d
et
ec
tio
n
an
d
class
if
icatio
n
p
ip
eli
n
e
2
.
1
.
Da
t
a
s
et
d
escript
io
n
T
h
is
s
t
u
d
y
u
t
ili
ze
s
a
c
o
l
lec
ti
o
n
o
f
4
8
0
g
r
a
y
s
c
al
e
th
y
r
o
id
u
lt
r
aso
u
n
d
i
m
a
g
es
o
b
ta
in
e
d
f
r
o
m
Nep
al
g
u
n
j
Me
d
ic
al
C
o
lle
g
e
a
n
d
T
e
ac
h
i
n
g
H
o
s
p
it
al
a
n
d
L
u
m
b
in
i
Me
d
i
ca
l
C
o
ll
eg
e
a
n
d
T
ea
ch
in
g
H
o
s
p
it
al
i
n
Ne
p
a
l.
I
m
a
g
e
ac
q
u
is
iti
o
n
o
c
cu
r
r
e
d
d
u
r
i
n
g
p
atie
n
t
u
ltr
as
o
u
n
d
e
x
am
i
n
ati
o
n
s
at
b
o
th
i
n
s
tit
u
t
io
n
s
,
a
n
d
a
ll
n
e
ce
s
s
ar
y
et
h
ic
al
ap
p
r
o
v
als
we
r
e
o
b
ta
in
ed
f
r
o
m
t
h
e
I
n
s
tit
u
ti
o
n
a
l
R
ev
iew
B
o
a
r
d
s
o
f
t
h
e
h
o
s
p
i
tals
.
All
p
a
tie
n
t
i
d
e
n
ti
f
i
ca
ti
o
n
d
a
ta
wer
e
a
n
o
n
y
m
i
ze
d
b
e
f
o
r
e
u
s
i
n
g
th
e
u
lt
r
as
o
u
n
d
s
c
an
s
f
o
r
a
n
al
y
s
is
.
D
u
r
i
n
g
t
h
e
la
b
e
li
n
g
p
r
o
c
es
s
,
t
h
e
im
a
g
es
we
r
e
class
i
f
i
ed
b
y
a
n
e
x
p
e
r
t
r
a
d
i
o
lo
g
is
t
.
T
h
e
b
e
n
i
g
n
c
ate
g
o
r
y
o
f
t
h
e
d
atas
et
is
m
u
ch
m
o
r
e
a
b
u
n
d
an
t
th
a
n
t
h
e
m
al
ig
n
a
n
t
c
at
eg
o
r
y
,
t
h
e
r
e
b
y
g
i
v
i
n
g
r
is
e
t
o
a
cl
ass
im
b
a
la
n
c
e
i
s
s
u
e
c
o
m
m
o
n
am
o
n
g
m
e
d
ic
al
i
m
a
g
e
d
at
asets
.
2
.
2
.
I
ma
g
e
prepro
ce
s
s
ing
a
nd
da
t
a
a
ug
m
ent
a
t
io
n
All
im
ag
es
wer
e
r
escaled
s
u
ch
th
at
th
e
r
eso
lu
tio
n
o
f
th
e
im
ag
es
b
ec
am
e
2
2
4
×
2
2
4
p
ix
els,
wh
ich
was
th
e
s
ize
o
f
th
e
in
p
u
t
r
e
q
u
ir
e
d
b
y
th
e
n
e
u
r
al
n
etwo
r
k
.
Alth
o
u
g
h
th
e
im
ag
es
wer
e
i
n
g
r
ay
s
c
ale
f
o
r
m
at
to
b
eg
i
n
with
,
it
was
n
ec
es
s
ar
y
to
co
n
v
er
t
th
e
im
ag
es
in
to
th
r
ee
-
c
h
a
n
n
el
im
ag
es
b
ec
au
s
e
we
h
ad
to
f
ee
d
th
e
im
ag
es
in
to
a
r
eg
u
lar
C
NN
with
th
r
e
e
ch
an
n
els.
T
h
e
n
e
x
t
s
tep
in
v
o
lv
ed
n
o
r
m
alizin
g
th
e
p
ix
el
v
alu
es
b
y
d
iv
id
i
n
g
th
em
b
y
2
5
5
an
d
e
n
s
u
r
in
g
ev
e
r
y
p
ix
el
is
in
[
0
,
1
]
r
an
g
e.
I
n
o
r
d
e
r
to
s
o
lv
e
th
e
is
s
u
e
o
f
class
im
b
alan
ce
wh
er
e
m
alig
n
an
t sam
p
les we
r
e
s
ig
n
if
ican
tly
f
ewe
r
th
an
b
en
ig
n
s
am
p
les,
o
v
er
s
am
p
lin
g
an
d
d
ata
au
g
m
en
tatio
n
wer
e
p
er
f
o
r
m
ed
.
T
h
is
was
ac
h
iev
ed
b
y
g
en
er
atin
g
n
ew
im
ag
es
th
r
o
u
g
h
h
o
r
izo
n
tal
f
li
p
s
,
ap
p
l
y
in
g
s
m
all
r
o
tatio
n
a
n
g
les
o
f
±
1
5
°,
a
n
d
ad
ju
s
tin
g
th
e
b
r
ig
h
tn
ess
b
y
s
ca
lin
g
it
b
etwe
en
v
alu
es
o
f
0
.
8
an
d
1
.
2
[
2
4
]
.
T
h
r
o
u
g
h
a
u
g
m
en
tatio
n
,
t
h
e
to
tal
n
u
m
b
er
o
f
i
m
ag
es
in
cr
ea
s
ed
to
4
9
0
.
Mo
r
eo
v
er
,
Mix
Up
[
2
2
]
an
d
C
u
tMix
[
2
3
]
s
tr
ateg
ies
wer
e
u
tili
ze
d
d
u
r
in
g
th
e
tr
ai
n
in
g
o
f
t
h
e
m
o
d
el.
Mix
Up
g
en
e
r
ates
n
ew
s
am
p
le
s
b
y
c
o
m
b
in
in
g
o
r
m
ix
in
g
two
im
ag
es
to
g
eth
er
alo
n
g
with
th
eir
co
r
r
esp
o
n
d
in
g
lab
els,
wh
ile
C
u
tMix
in
v
o
l
v
es
cu
ttin
g
o
u
t
a
p
o
r
tio
n
o
f
o
n
e
im
ag
e
an
d
p
asti
n
g
it
o
n
to
an
o
th
er
im
ag
e.
Ad
d
itio
n
ally
,
a
class
-
weig
h
te
d
b
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
l
o
s
s
f
u
n
ctio
n
was
em
p
lo
y
ed
d
u
r
in
g
C
NN
tr
ain
in
g
,
with
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Dete
ctio
n
a
n
d
cla
s
s
ifica
tio
n
o
f th
yro
id
d
is
ea
s
es u
s
in
g
u
ltr
a
s
o
u
n
d
…
(
K
a
ma
l
S
u
b
e
d
i
)
2809
th
e
m
alig
n
a
n
t
class
r
ec
eiv
in
g
a
weig
h
t
p
r
o
p
o
r
tio
n
al
to
th
e
in
v
er
s
e
o
f
its
class
f
r
e
q
u
en
c
y
(
weig
h
t
r
atio
o
f
ap
p
r
o
x
im
ately
3
4
:1
f
o
r
m
alig
n
an
t
v
s
.
b
e
n
ig
n
)
.
T
h
is
weig
h
ti
n
g
s
tr
ateg
y
e
n
s
u
r
es
th
at
m
is
class
if
icatio
n
s
o
f
th
e
r
ar
e
m
alig
n
a
n
t
class
co
n
tr
ib
u
t
e
s
ig
n
if
ican
tly
m
o
r
e
to
th
e
g
r
ad
ien
t
u
p
d
ates,
th
er
eb
y
co
u
n
t
er
ac
tin
g
th
e
n
atu
r
al
b
ias o
f
th
e
o
p
tim
izer
to
war
d
th
e
m
ajo
r
ity
b
en
ig
n
class
.
2
.
3
.
P
s
eudo
-
ma
s
k
g
ener
a
t
io
n
T
h
e
f
ac
t
th
at
th
e
i
n
itial
d
ata
lack
ed
p
ix
el
-
wis
e
an
n
o
tatio
n
s
p
r
o
v
i
d
ed
b
y
th
e
m
ed
ical
p
r
o
f
ess
io
n
al
s
m
ad
e
it
n
ec
ess
ar
y
to
cr
ea
te
a
n
au
to
m
ated
p
r
o
ce
d
u
r
e
to
g
e
n
er
ate
p
s
eu
d
o
-
m
ask
s
u
s
ed
in
th
e
tr
ain
in
g
o
f
th
e
s
e
g
m
e
n
t
a
ti
o
n
n
e
t
w
o
r
k
.
T
h
e
p
r
o
c
e
s
s
i
n
c
l
u
d
e
s
t
h
r
ee
s
t
a
g
es
.
A
t
f
i
r
s
t
,
C
L
A
HE
i
s
p
e
r
f
o
r
m
e
d
o
n
t
h
e
i
n
p
u
t
u
l
t
r
a
s
o
u
n
d
i
m
a
g
e
.
T
h
e
m
ai
n
i
d
e
a
b
e
h
i
n
d
C
L
A
H
E
i
s
t
o
p
a
r
ti
t
i
o
n
t
h
e
im
a
g
e
i
n
t
o
s
e
v
e
r
a
l
s
m
al
l
p
a
r
t
s
a
n
d
e
q
u
a
l
i
z
e
t
h
e
h
i
s
t
o
g
r
a
m
w
it
h
i
n
e
a
c
h
p
a
r
t
i
n
d
e
p
e
n
d
e
n
t
l
y
.
T
h
u
s
,
t
h
e
b
o
u
n
d
a
r
ie
s
o
f
t
h
e
t
h
y
r
o
i
d
n
o
d
u
l
e
b
e
c
o
m
e
m
o
r
e
n
o
t
i
c
e
a
b
l
e
.
Seco
n
d
ly
,
Ots
u
th
r
esh
o
ld
in
g
alo
n
g
with
a
b
in
ar
y
in
v
er
s
e
co
n
f
ig
u
r
atio
n
was
u
s
ed
in
o
r
d
er
t
o
d
if
f
er
en
tiate
th
e
d
ar
k
a
r
ea
o
f
th
e
n
o
d
u
le
f
r
o
m
o
th
er
b
ac
k
g
r
o
u
n
d
ar
ea
s
o
f
th
e
im
ag
e.
T
h
e
Ots
u
tech
n
iq
u
e
au
to
m
atica
lly
f
in
d
s
th
e
o
p
ti
m
al
th
r
esh
o
ld
v
alu
e
b
y
m
ax
im
izin
g
th
e
v
ar
ian
ce
b
etwe
en
th
e
two
class
es.
Alth
o
u
g
h
Ots
u
th
r
esh
o
ld
in
g
a
s
s
u
m
es
a
b
im
o
d
al
in
ten
s
ity
d
is
tr
ib
u
tio
n
,
wh
ich
ca
n
b
e
co
m
p
r
o
m
is
ed
b
y
s
p
ec
k
le
n
o
is
e
in
u
ltra
s
o
u
n
d
im
ag
es,
t
h
e
p
r
ec
ed
in
g
C
L
AHE
en
h
an
ce
m
en
t
s
tep
m
itig
ates
th
is
li
m
itatio
n
b
y
lo
ca
lly
n
o
r
m
alizin
g
c
o
n
tr
ast
an
d
im
p
r
o
v
in
g
th
e
s
ep
ar
a
b
ilit
y
o
f
n
o
d
u
le
a
n
d
b
a
ck
g
r
o
u
n
d
r
eg
i
o
n
s
.
W
e
ev
alu
ated
alter
n
ativ
e
lo
ca
l
ad
a
p
tiv
e
th
r
esh
o
ld
in
g
m
et
h
o
d
s
(
e.
g
.
,
Sau
v
o
la
an
d
B
er
n
s
en
)
b
u
t
f
o
u
n
d
th
at
g
lo
b
al
Ots
u
th
r
esh
o
ld
in
g
o
n
C
L
AHE
-
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
es
p
r
o
d
u
ce
d
m
o
r
e
co
n
s
is
ten
t
an
d
co
h
er
en
t
b
i
n
ar
y
m
ask
s
f
o
r
th
e
p
s
eu
d
o
-
lab
el
g
e
n
er
atio
n
,
as
lo
ca
l
m
eth
o
d
s
ten
d
ed
to
o
v
er
-
s
eg
m
en
t
in
r
eg
io
n
s
with
g
r
ad
u
al
in
ten
s
ity
tr
an
s
itio
n
s
ty
p
ical
o
f
u
ltra
s
o
u
n
d
im
ag
e
r
y
.
T
h
ir
d
ly
,
m
o
r
p
h
o
lo
g
ical
o
p
e
n
in
g
a
n
d
cl
o
s
in
g
w
er
e
p
er
f
o
r
m
ed
u
s
in
g
an
ellip
tical
s
tr
u
ctu
r
e
elem
en
t
o
f
s
ize
7
×
7
to
elim
in
ate
n
o
is
e
an
d
im
p
r
o
v
e
ed
g
e
d
etec
tio
n
o
f
th
e
b
i
n
ar
y
m
ask
.
L
astl
y
,
a
co
n
to
u
r
d
etec
tio
n
tec
h
n
iq
u
e
was
u
s
ed
t
o
d
etec
t
all
co
n
n
ec
ted
c
o
m
p
o
n
en
ts
with
in
th
e
im
ag
e,
a
n
d
th
e
lar
g
est co
m
p
o
n
en
t w
as c
o
n
s
id
er
ed
th
e
p
s
eu
d
o
-
m
ask
o
f
th
e
th
y
r
o
id
n
o
d
u
le.
2
.
4
.
At
t
ent
io
n U
-
Net
f
o
r
no
du
le
s
eg
m
ent
a
t
io
n
T
h
e
s
eg
m
en
tatio
n
ar
c
h
itectu
r
e
f
o
llo
ws
th
e
atten
tio
n
m
ec
h
a
n
is
m
o
f
U
-
Net
[
1
8
]
,
an
ex
ten
s
io
n
to
t
h
e
co
n
v
en
tio
n
al
U
-
Net
[
1
7
]
ar
c
h
itectu
r
e
with
th
e
u
s
e
o
f
atten
tio
n
g
ates
at
ea
ch
s
k
ip
co
n
n
ec
ti
o
n
.
T
h
is
p
ar
ticu
lar
ar
ch
itectu
r
e
was
cr
ea
ted
s
p
ec
i
f
ically
f
o
r
b
i
o
m
ed
ical
im
ag
in
g
ap
p
licatio
n
s
,
esp
ec
ially
th
o
s
e
in
v
o
lv
in
g
im
a
g
e
s
eg
m
en
tatio
n
.
I
t
co
n
s
is
ts
o
f
t
wo
p
ath
s
ca
lle
d
en
c
o
d
er
s
an
d
d
ec
o
d
er
s
,
r
esp
ec
tiv
ely
.
T
h
e
en
co
d
er
s
d
ec
r
ea
s
e
th
e
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
e
m
a
p
s
wh
ile
in
cr
ea
s
in
g
th
e
n
u
m
b
er
o
f
f
ilter
s
,
wh
e
r
ea
s
th
e
r
ev
e
r
s
e
h
ap
p
en
s
with
th
e
d
ec
o
d
er
s
,
wh
o
s
e
f
u
n
ctio
n
is
to
r
esto
r
e
th
e
s
p
atial
d
im
en
s
io
n
s
.
Fo
u
r
co
n
v
o
lu
tio
n
al
b
lo
ck
s
ar
e
in
clu
d
ed
in
th
e
en
c
o
d
er
p
ath
o
f
o
u
r
Atten
tio
n
U
-
Net
ar
ch
itec
tu
r
e
with
f
ilter
s
o
f
3
2
,
6
4
,
1
2
8
,
a
n
d
2
5
6
as
s
h
o
wn
in
Fig
u
r
e
2
.
E
ac
h
b
lo
ck
co
m
p
r
is
es
two
co
n
s
ec
u
tiv
e
3
×
3
co
n
v
o
l
u
tio
n
al
lay
er
s
a
n
d
th
en
b
atch
n
o
r
m
aliza
tio
n
is
ap
p
lied
an
d
th
e
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
an
d
th
en
2
×
2
m
ax
-
p
o
o
lin
g
o
p
er
atio
n
f
o
r
d
o
wn
s
am
p
lin
g
.
T
h
r
ee
u
p
s
am
p
l
in
g
b
lo
c
k
s
ar
e
i
n
co
r
p
o
r
ated
i
n
th
e
d
ec
o
d
e
r
p
at
h
with
f
ilter
s
o
f
1
2
8
,
6
4
,
a
n
d
3
2
,
u
s
in
g
2
×
2
t
r
an
s
p
o
s
ed
c
o
n
v
o
l
u
tio
n
s
.
An
atte
n
tio
n
g
ate
m
ec
h
an
is
m
is
em
p
lo
y
ed
at
ea
ch
s
k
ip
co
n
n
ec
tio
n
to
e
m
p
h
asize
th
e
r
elev
an
t
a
r
ea
s
in
th
e
im
ag
e.
T
h
e
o
u
tp
u
t
f
r
o
m
th
e
atten
tio
n
g
at
e
g
en
er
ates
a
weig
h
t
m
ap
,
wh
ic
h
f
o
c
u
s
es
o
n
th
e
n
o
d
u
le
r
e
g
io
n
wh
ile
s
u
p
p
r
ess
in
g
th
e
b
ac
k
g
r
o
u
n
d
r
e
g
io
n
.
T
h
is
is
s
ig
n
if
ican
t
wh
en
d
ea
lin
g
with
u
ltra
s
o
u
n
d
im
ag
es,
wh
er
e
th
e
b
o
u
n
d
a
r
y
o
f
th
e
n
o
d
u
le
is
d
if
f
icu
lt
to
d
is
tin
g
u
is
h
b
ec
au
s
e
o
f
lo
w
co
n
tr
ast an
d
s
p
ec
k
le
n
o
is
e.
T
h
e
s
eg
m
en
tatio
n
o
u
tp
u
t o
f
th
e
m
o
d
el
is
a
b
in
ar
y
m
ask
o
f
d
im
en
s
io
n
s
2
2
4
×
2
2
4
b
y
em
p
lo
y
in
g
th
e
1
×
1
co
n
v
o
lu
tio
n
l
ay
er
an
d
s
ig
m
o
id
ac
tiv
atio
n
.
Fig
u
r
e
2
.
Atten
tio
n
U
-
Net
ar
c
h
itectu
r
e
u
s
ed
f
o
r
th
y
r
o
id
n
o
d
u
le
s
eg
m
en
tatio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
16
,
No
.
5
,
Octo
b
e
r
20
26
:
2
8
0
6
-
2
8
1
8
2810
2
.
5
.
Reg
i
o
n o
f
inte
re
s
t
ex
t
ra
ct
io
n
On
ce
th
e
Atten
tio
n
U
-
Net
h
as
g
en
er
ated
th
e
b
in
ar
y
s
eg
m
en
t
atio
n
m
ask
,
b
o
u
n
d
in
g
b
o
x
e
x
tr
ac
tio
n
is
p
er
f
o
r
m
ed
to
c
r
o
p
o
u
t
th
e
R
eg
io
n
o
f
I
n
ter
est
f
r
o
m
th
e
r
aw
u
l
tr
aso
u
n
d
im
a
g
e
as
s
h
o
wn
in
F
ig
u
r
e
3
.
B
o
u
n
d
i
n
g
b
o
x
g
e
n
er
atio
n
in
v
o
lv
es
d
eter
m
in
in
g
th
e
r
o
w
an
d
co
lu
m
n
v
alu
es
f
o
r
th
e
m
in
im
u
m
a
n
d
m
ax
im
u
m
r
o
ws
an
d
co
lu
m
n
s
wh
er
e
t
h
e
n
o
d
u
le
a
p
p
ea
r
s
in
th
e
s
eg
m
e
n
tatio
n
m
as
k
.
T
h
is
r
eg
i
o
n
is
th
en
e
x
tr
ac
te
d
f
r
o
m
th
e
o
r
ig
in
al
im
ag
e
an
d
f
ed
in
to
th
e
class
if
icatio
n
m
o
d
el.
T
h
is
p
r
o
ce
s
s
h
e
lp
s
en
s
u
r
e
th
at
th
e
n
eu
r
al
n
etwo
r
k
p
ay
s
atten
tio
n
to
th
e
n
o
d
u
le
o
n
ly
an
d
n
o
t
an
y
o
th
er
r
eg
io
n
s
with
in
t
h
e
r
aw
im
ag
e.
Af
ter
g
e
n
er
a
tin
g
th
e
p
r
ed
icted
s
eg
m
en
tatio
n
m
ask
,
m
o
r
p
h
o
l
o
g
ical
o
p
er
atio
n
s
wer
e
p
e
r
f
o
r
m
ed
to
im
p
r
o
v
e
th
e
o
u
tp
u
t
m
ask
b
o
u
n
d
a
r
ies.
I
n
p
ar
ticu
lar
,
th
e
m
o
r
p
h
o
lo
g
ical
clo
s
in
g
with
a
3
×3
cir
cu
lar
k
e
r
n
el
was
p
er
f
o
r
m
e
d
f
o
r
clo
s
in
g
s
m
all
h
o
les
with
in
th
e
p
r
ed
icted
n
o
d
u
le
ar
ea
,
f
o
llo
wed
b
y
th
e
m
o
r
p
h
o
lo
g
ical
o
p
en
in
g
f
o
r
r
em
o
v
in
g
s
m
all
f
al
s
e
-
p
o
s
itiv
e
clu
s
ter
s
less
th
an
5
0
p
ix
els in
s
ize.
Fig
u
r
e
3
.
E
x
am
p
le
o
u
tp
u
ts
o
f
t
h
e
s
eg
m
en
tatio
n
p
ip
elin
e
s
h
o
win
g
o
r
ig
in
al
im
a
g
es,
p
s
eu
d
o
-
m
ask
s
,
an
d
p
r
ed
icted
m
ask
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Dete
ctio
n
a
n
d
cla
s
s
ifica
tio
n
o
f th
yro
id
d
is
ea
s
es u
s
in
g
u
ltr
a
s
o
u
n
d
…
(
K
a
ma
l
S
u
b
e
d
i
)
2811
2
.
6
.
Cus
t
o
m
CNN
f
o
r
c
la
s
s
if
ica
t
io
n
T
h
e
ex
tr
ac
ted
p
atc
h
es
f
r
o
m
t
h
e
th
y
r
o
id
n
o
d
u
les
ar
e
lab
eled
b
ased
o
n
th
e
d
ev
elo
p
e
d
C
NN
m
o
d
el
th
at
u
s
es
th
e
Ker
as
f
u
n
ctio
n
al
API
.
T
h
is
m
o
d
el
co
m
p
r
is
es
th
r
ee
co
n
v
o
lu
tio
n
b
lo
c
k
s
with
an
in
cr
ea
s
e
in
th
e
n
u
m
b
er
o
f
f
ilter
s
in
cr
em
en
tall
y
f
r
o
m
3
2
to
6
4
to
f
in
ally
1
2
8
f
ilter
s
.
T
h
ese
th
r
ee
co
n
v
o
lu
tio
n
b
lo
ck
s
h
av
e
a
p
air
o
f
3
×3
co
n
v
o
lu
tio
n
al
lay
er
s
in
co
r
p
o
r
atin
g
b
atch
n
o
r
m
a
lizatio
n
alo
n
g
with
R
eL
U
ac
tiv
atio
n
f
u
n
ctio
n
s
t
o
b
o
o
s
t
th
e
lear
n
i
n
g
ca
p
a
b
ilit
ies.
Af
ter
th
is
,
th
e
d
o
wn
s
am
p
lin
g
p
r
o
ce
s
s
em
p
lo
y
s
m
a
x
p
o
o
lin
g
with
a
2
×2
k
er
n
el
f
o
r
e
x
tr
ac
tio
n
o
f
r
elev
an
t
f
e
atu
r
es.
Par
am
eter
s
ar
e
r
ed
u
c
ed
,
an
d
to
p
r
ev
e
n
t
o
v
er
f
itti
n
g
,
a
g
lo
b
al
a
v
er
ag
e
p
o
o
lin
g
p
r
o
ce
s
s
is
u
s
ed
.
Mo
r
e
o
v
er
,
d
r
o
p
o
u
t
r
eg
u
la
r
izatio
n
with
a
0
.
5
r
ate
is
in
c
o
r
p
o
r
ated
in
th
e
C
NN
m
o
d
el.
T
h
e
class
if
icatio
n
p
r
o
ce
s
s
tak
es
p
lace
with
th
e
h
el
p
o
f
s
ig
m
o
id
ac
tiv
atio
n
with
a
th
r
esh
o
ld
v
alu
e
o
f
0
.
5
to
d
is
tin
g
u
is
h
b
etwe
en
b
e
n
ig
n
a
n
d
m
alig
n
an
t
n
o
d
u
les
as sh
o
wn
in
Fig
u
r
e
4.
Fig
u
r
e
4
.
Ar
c
h
itectu
r
e
f
o
r
cu
s
t
o
m
C
NN
u
s
ed
f
o
r
b
en
ig
n
-
m
ali
g
n
an
t c
lass
if
icatio
n
2
.
7
.
T
ra
ini
ng
p
ro
t
o
co
l
T
r
ain
in
g
f
o
r
b
o
th
th
e
Atten
tio
n
U
-
Net
an
d
C
NN
m
o
d
els
was
ca
r
r
ied
o
u
t
with
th
e
Ad
a
m
o
p
tim
izatio
n
alg
o
r
ith
m
.
W
e
ch
o
s
e
a
s
tar
tin
g
lear
n
in
g
r
ate
o
f
10
−
5
to
m
ai
n
tain
s
tab
le
co
n
v
er
g
en
ce
w
h
ile
th
e
n
etwo
r
k
was
b
ein
g
tr
ain
e
d
.
As
th
e
o
b
jectiv
e
f
u
n
ctio
n
,
b
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
was
ap
p
lied
to
b
o
th
th
e
s
eg
m
en
tatio
n
an
d
th
e
class
if
icatio
n
task
s
.
A
r
ed
u
ctio
n
in
th
e
lear
n
in
g
r
at
e
d
u
r
i
n
g
t
r
ain
in
g
was
attain
ed
th
r
o
u
g
h
t
h
e
R
ed
u
ce
L
R
On
Plateau
ca
llb
ac
k
.
E
ar
ly
s
to
p
p
i
n
g
was
ad
o
p
te
d
to
av
o
id
o
v
er
f
itti
n
g
,
with
th
e
p
atien
ce
s
et
to
1
5
ep
o
ch
s
.
T
h
e
b
atc
h
s
ize
an
d
m
ax
im
u
m
ep
o
c
h
s
wer
e
8
an
d
1
0
0
,
r
esp
ec
tiv
ely
.
W
e
c
o
n
d
u
cted
3
-
f
o
ld
s
tr
atif
ie
d
cr
o
s
s
-
v
alid
atio
n
to
ev
alu
ate
o
u
r
m
o
d
el
b
ec
au
s
e
it
en
s
u
r
es
th
e
r
eliab
ilit
y
o
f
o
u
r
r
esu
lts
r
eg
ar
d
less
o
f
th
e
d
ata
s
p
lit.
2
.
8
.
L
o
s
s
f
un
ct
io
ns
a
nd
ev
a
lua
t
io
n m
et
rics
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
o
b
j
ec
tiv
e
f
u
n
ctio
n
s
ap
p
lied
t
o
tr
ain
th
e
two
n
etwo
r
k
s
,
to
g
eth
e
r
with
t
h
e
m
etr
ics
ad
o
p
ted
f
o
r
m
ea
s
u
r
i
n
g
h
o
w
well
th
e
s
eg
m
e
n
tatio
n
an
d
class
if
icatio
n
s
tag
es
p
er
f
o
r
m
with
in
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
.
2
.
8
.
1
.
Cro
s
s
ent
ro
py
lo
s
s
Fo
r
th
e
b
in
ar
y
class
if
icatio
n
t
ask
(
b
en
i
g
n
o
r
m
alig
n
a
n
t)
,
th
e
cu
s
to
m
ized
C
NN
is
tr
ain
ed
u
s
in
g
th
e
b
in
ar
y
c
r
o
s
s
-
en
tr
o
p
y
o
b
jectiv
e
,
wh
o
s
e
m
ath
em
atica
l f
o
r
m
is
g
iv
en
b
y
:
=
−
1
∑
[
.
(
̂
)
+
(
1
−
)
.
l
og
(
1
−
̂
]
Her
e,
N
d
en
o
tes
th
e
to
tal
n
u
m
b
er
o
f
s
am
p
les,
−
r
ep
r
esen
ts
th
e
g
r
o
u
n
d
-
tr
u
th
lab
el,
an
d
̂
co
r
r
e
s
p
o
n
d
s
to
th
e
p
r
ed
icted
p
r
o
b
a
b
ilit
y
f
o
r
th
e
m
alig
n
an
t c
ateg
o
r
y
.
2
.
8
.
2
.
Dice
l
o
s
s
I
n
th
e
ca
s
e
o
f
th
e
Atten
tio
n
U
-
Net
s
eg
m
en
tatio
n
p
h
ase,
d
ice
lo
s
s
i
s
u
s
ed
in
o
r
d
er
to
m
ak
e
th
e
p
r
ed
ictio
n
m
ask
o
v
er
lap
with
th
e
p
s
eu
d
o
-
m
ask
.
T
h
e
ex
p
r
ess
io
n
f
o
r
d
ice
s
im
ilar
ity
co
ef
f
ici
en
t
(
DSC
)
an
d
ca
n
b
e
wr
itten
as:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
16
,
No
.
5
,
Octo
b
e
r
20
26
:
2
8
0
6
-
2
8
1
8
2812
=
2
|
⋂
|
|
|
+
|
|
wh
er
e
r
ep
r
esen
ts
th
e
p
r
ed
icted
s
eg
m
en
tatio
n
,
an
d
co
r
r
esp
o
n
d
s
to
th
e
g
r
o
u
n
d
-
tr
u
t
h
m
ask
.
T
h
e
d
ice
lo
s
s
,
wh
ich
is
m
in
im
ized
d
u
r
in
g
tr
a
in
in
g
,
is
th
en
:
=
1
−
2
.
8
.
3
.
I
nte
rsect
io
n o
v
er
un
io
n
(
I
o
U)
An
o
th
er
im
p
o
r
tan
t
m
ea
s
u
r
e
o
f
th
e
q
u
ality
o
f
th
e
s
eg
m
e
n
tatio
n
m
o
d
el
is
in
ter
s
ec
tio
n
o
v
er
u
n
io
n
(
I
o
U)
.
I
o
U
is
co
m
p
u
ted
b
ase
d
o
n
h
o
w
s
im
ilar
th
e
r
e
g
io
n
s
o
f
th
e
p
r
ed
icted
m
ask
ar
e
to
th
e
r
eg
io
n
s
o
f
th
e
tar
g
et
m
ask
:
=
|
⋂
|
|
⋃
|
2
.
8
.
4
.
M
ix
U
p
a
ug
m
ent
a
t
io
n s
t
r
a
t
eg
y
Mix
Up
is
a
d
ata
au
g
m
en
tatio
n
m
eth
o
d
u
s
ed
in
tr
ain
in
g
to
im
p
r
o
v
e
g
en
er
aliza
tio
n
an
d
av
o
i
d
o
v
er
f
itti
n
g
.
T
h
is
s
tr
ateg
y
cr
ea
t
es
n
ew
tr
ain
in
g
ex
am
p
les
b
y
tak
in
g
th
e
weig
h
ted
s
u
m
o
f
two
d
ata
s
am
p
les
an
d
th
eir
co
r
r
esp
o
n
d
in
g
lab
els.
T
h
e
m
ix
ed
s
am
p
les ar
e
g
e
n
er
ated
as:
̃
=
+
(
1
−
)
̃
=
+
(
1
−
)
wh
er
e
(
,
)
an
d
(
,
)
ar
e
r
an
d
o
m
l
y
s
am
p
led
tr
ain
in
g
p
air
s
,
a
n
d
f
o
llo
ws
a
B
eta
d
is
tr
ib
u
tio
n
,
co
n
tr
o
llin
g
th
e
p
r
o
p
o
r
tio
n
o
f
m
ix
in
g
,
wh
er
e
α
is
th
e
c
o
n
ce
n
tr
atio
n
p
a
r
am
eter
o
f
t
h
e
B
eta
d
is
tr
ib
u
tio
n
(
s
et
to
0
.
2
in
o
u
r
ex
p
er
im
en
ts
)
,
c
o
n
tr
o
llin
g
th
e
s
tr
en
g
th
o
f
in
ter
p
o
latio
n
b
etwe
en
th
e
two
s
am
p
les.
2
.
8
.
5
.
Cla
s
s
if
ica
t
io
n
m
et
rics
Fo
r
th
e
co
m
p
r
eh
en
s
iv
e
ass
ess
m
en
t
o
f
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
cu
s
to
m
ized
C
NN
class
if
ier
,
th
e
f
o
llo
win
g
m
etr
ics
o
f
b
in
ar
y
class
if
icatio
n
h
av
e
b
ee
n
co
m
p
u
ted
b
ased
o
n
t
h
e
tr
u
e
p
o
s
itiv
es
(
T
P),
tr
u
e
n
eg
at
iv
es
(
T
N)
,
f
alse
p
o
s
itiv
es
(
FP
)
,
an
d
f
alse
n
eg
ativ
es
(
FN)
v
alu
es
f
r
o
m
th
e
co
n
f
u
s
io
n
m
atr
ix
.
T
h
e
im
p
o
r
tan
t m
etr
ics in
clu
d
e:
a.
Sen
s
itiv
ity
(
R
ec
all)
: M
ea
s
u
r
es
th
e
p
er
ce
n
ta
g
e
o
f
tr
u
e
m
alig
n
an
t c
ases
th
at
ar
e
co
r
r
ec
tly
d
et
ec
ted
:
=
+
b.
Sp
ec
if
icity
:
Me
asu
r
es
th
e
p
er
ce
n
tag
e
o
f
tr
u
e
b
en
ig
n
ca
s
es
th
at
ar
e
co
r
r
ec
tly
d
etec
ted
.
T
h
is
m
ea
s
u
r
e
is
im
p
o
r
tan
t
b
ec
au
s
e
o
f
th
e
cla
s
s
im
b
alan
ce
is
s
u
e
to
s
ee
w
h
eth
er
th
e
alg
o
r
ith
m
ca
n
av
o
id
u
n
n
ec
ess
ar
y
b
io
p
s
ies:
=
+
c.
Pre
cisi
o
n
: M
ea
s
u
r
es th
e
p
er
ce
n
tag
e
o
f
t
r
u
e
p
o
s
itiv
es (
m
alig
n
an
t)
am
o
n
g
all
m
alig
n
a
n
t p
r
e
d
ictio
n
s
:
=
+
d.
F1
-
Sco
r
e:
Har
m
o
n
ic
m
ea
n
o
f
p
r
ec
is
io
n
an
d
r
ec
all,
wh
er
e
p
r
ec
is
io
n
is
an
d
r
ec
all
is
:
1
=
2
+
wh
er
e
d
en
o
tes
p
r
ec
is
io
n
an
d
d
en
o
tes
s
en
s
itiv
ity
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Dete
ctio
n
a
n
d
cla
s
s
ifica
tio
n
o
f th
yro
id
d
is
ea
s
es u
s
in
g
u
ltr
a
s
o
u
n
d
…
(
K
a
ma
l
S
u
b
e
d
i
)
2813
2
.
9
.
I
m
ple
m
ent
a
t
io
n
d
et
a
ils
T
h
e
ex
p
er
im
en
ts
wer
e
ca
r
r
i
ed
o
u
t
in
ac
co
r
d
an
ce
with
t
h
e
f
o
llo
win
g
s
etu
p
f
o
r
r
e
p
r
o
d
u
cib
ilit
y
p
u
r
p
o
s
es.
T
h
e
em
p
lo
y
ed
d
ata
s
et
co
m
p
r
is
ed
4
9
0
im
ag
es
o
f
th
y
r
o
id
u
ltra
s
o
u
n
d
s
with
a
r
o
u
g
h
ly
3
4
:1
r
atio
o
f
b
en
ig
n
to
m
alig
n
an
t
ca
s
es
(
4
7
6
b
en
i
g
n
a
n
d
1
4
m
alig
n
an
t
)
.
P
r
ep
r
o
ce
s
s
in
g
in
v
o
lv
e
d
r
esizin
g
th
e
im
a
g
es
to
t
h
e
d
im
en
s
io
n
s
o
f
2
2
4
×
2
2
4
,
c
o
lo
r
co
n
v
e
r
s
io
n
f
r
o
m
g
r
a
y
s
ca
le
to
R
GB
,
an
d
n
o
r
m
aliza
tio
n
t
o
th
e
r
an
g
e
[
0
,
1
]
.
Fo
r
th
e
p
s
eu
d
o
‑
m
ask
in
g
p
r
o
c
ess
,
C
L
AHE
with
a
c
lip
p
in
g
th
r
esh
o
ld
o
f
2
.
0
an
d
a
tile
g
r
id
h
av
in
g
d
im
en
s
io
n
s
(
8
×8
)
was
ap
p
lied
,
al
o
n
g
with
Ots
u
's
th
r
esh
o
ld
in
g
an
d
m
o
r
p
h
o
lo
g
ical
o
p
er
atio
n
s
with
a
7
×7
ellip
tical
k
er
n
el.
All
tr
ials
wer
e
ex
ec
u
ted
o
n
a
s
in
g
le
A1
0
0
ac
ce
ler
ato
r
f
r
o
m
NVI
DI
A
u
s
in
g
th
e
Ker
as
API
b
u
ilt
o
n
T
en
s
o
r
Flo
w,
an
d
n
etwo
r
k
tr
ai
n
in
g
e
m
p
lo
y
ed
t
h
e
Ad
a
m
o
p
ti
m
izatio
n
alg
o
r
ith
m
h
av
in
g
a
l
ea
r
n
in
g
r
ate
o
f
1
e‑
5
an
d
a
m
in
i‑
b
atch
s
ize
eq
u
al
to
8
,
u
p
to
1
0
0
e
p
o
ch
s
,
p
atien
ce
f
o
r
ea
r
ly
s
to
p
p
in
g
o
f
1
5
an
d
d
r
o
p
o
u
t
o
f
0
.
5
.
T
h
e
in
tr
o
d
u
ce
d
s
eg
m
e
n
tatio
n
ar
c
h
i
tectu
r
e,
d
e
r
iv
ed
f
r
o
m
th
e
Atten
tio
n
U‑
s
h
ap
e
d
n
etwo
r
k
with
g
atin
g
,
h
as
ar
o
u
n
d
2
.
0
1
m
illi
o
n
tr
ain
a
b
le
p
a
r
am
e
ter
s
wh
ile
th
e
cu
s
to
m
C
NN
class
if
ier
h
as
ab
o
u
t
0
.
8
m
illi
o
n
p
a
r
am
eter
s
.
T
h
e
to
tal
n
u
m
b
er
o
f
p
ar
am
eter
s
a
m
o
u
n
ts
to
ju
s
t
a
f
ew
GFLO
Ps
an
d
th
e
i
n
f
er
e
n
ce
r
e
q
u
ir
es
lo
w
m
em
o
r
y
f
o
r
b
o
th
s
tag
es,
th
u
s
v
alid
atin
g
th
at
t
h
e
ar
ch
itectu
r
e
is
lig
h
tweig
h
t
an
d
d
ep
lo
y
a
b
le
o
n
r
eg
u
lar
clin
ical
h
ar
d
war
e
.
I
n
f
er
en
ce
laten
cy
,
d
eter
m
in
e
d
u
s
in
g
a
GPU
an
d
a
c
o
n
s
u
m
er
C
PU,
p
r
o
v
es
th
at
th
e
f
r
am
ewo
r
k
p
r
o
ce
s
s
es
u
ltra
s
o
u
n
d
im
a
g
es in
clin
ically
ac
ce
p
tab
le
tim
ef
r
a
m
es f
o
r
d
i
ag
n
o
s
tic
ap
p
licatio
n
s
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
Seg
m
ent
a
t
i
o
n
p
er
f
o
rma
nce
Qu
an
titativ
e
ev
alu
atio
n
o
f
att
en
tio
n
U‑
Net
s
eg
m
en
tatio
n
p
a
r
t
is
ca
r
r
ied
o
u
t
u
s
in
g
t
h
e
f
o
ll
o
win
g
two
m
etr
ics:
Dice
s
co
r
e
an
d
I
o
U
(
also
k
n
o
wn
as
J
ac
ca
r
d
in
d
ex
)
.
T
ab
le
1
s
u
m
m
ar
izes
th
ese
o
u
tco
m
es,
s
h
o
win
g
th
at
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
g
o
t
an
av
er
ag
e
Dice
s
co
r
e
e
q
u
al
to
0
.
8
7
7
2
±
0
.
0
0
5
4
a
n
d
th
e
a
v
er
ag
e
I
o
U
eq
u
al
to
0
.
7
8
7
3
±
0
.
0
0
7
5
.
T
h
is
im
p
lies
th
at
th
e
m
o
d
el
g
ets
s
ig
n
i
f
ican
t
ag
r
ee
m
en
t
b
etwe
en
s
eg
m
en
ted
a
n
d
r
ef
er
e
n
ce
m
ask
s
in
s
p
ite
o
f
a
lack
o
f
e
x
p
er
t
p
ix
el‑
le
v
el
an
n
o
tatio
n
s
.
I
t
is
also
ev
id
en
t
f
r
o
m
th
e
c
o
n
s
is
ten
cy
o
f
r
esu
lts
ac
r
o
s
s
d
if
f
er
en
t
f
o
ld
s
(
s
tan
d
a
r
d
d
ev
iatio
n
b
ein
g
less
th
an
0
.
0
1
f
o
r
b
o
th
m
etr
ics).
As
ca
n
b
e
s
ee
n
f
r
o
m
Fig
u
r
e
3
,
th
e
m
o
d
el
c
o
p
ed
well
with
is
o
latio
n
o
f
n
o
d
u
les
an
d
g
en
e
r
atio
n
o
f
a
p
p
r
o
p
r
iate
b
o
u
n
d
a
r
ies
f
o
r
th
em
.
I
t
is
ex
p
lain
ed
b
y
th
e
atten
tio
n
co
m
p
o
n
en
t
o
f
th
e
m
o
d
el
wh
ich
f
o
r
ce
s
th
e
n
etwo
r
k
to
p
ay
atten
tio
n
to
th
e
n
o
d
u
le
p
ar
t w
h
ile
ig
n
o
r
in
g
th
e
b
ac
k
g
r
o
u
n
d
o
n
e.
T
ab
le
1
.
Seg
m
e
n
tatio
n
p
e
r
f
o
r
m
an
ce
o
f
atten
ti
o
n
U
-
Net
ac
r
o
s
s
th
r
ee
f
o
ld
s
M
e
t
r
i
c
s
F
o
l
d
I
F
o
l
d
I
I
F
o
l
d
I
I
I
M
e
a
n
+
/
-
S
t
d
D
i
c
e
S
c
o
r
e
0
.
8
7
1
6
0
.
8
7
7
9
0
.
8
8
2
3
0
.
8
7
7
2
+
/
-
0
.
0
0
5
4
I
o
U
(
Jac
c
a
r
d
)
0
.
7
7
9
4
0
.
7
8
8
2
0
.
7
9
4
2
0
.
7
8
7
3
+
/
-
0
.
0
0
7
5
3
.
2
.
Cla
s
s
if
ica
t
io
n
p
er
f
o
r
m
a
nce
T
h
e
ex
tr
ac
te
d
r
e
g
io
n
s
o
f
in
te
r
est
(
i.e
.
,
R
OI
s
)
wer
e
s
u
p
p
lie
d
to
t
h
e
class
if
ier
,
wh
ich
m
a
p
p
ed
e
v
er
y
r
eg
io
n
o
n
to
o
n
e
o
f
two
d
ia
g
n
o
s
tic
g
r
o
u
p
s
:
class
0
,
d
en
o
tin
g
b
en
i
g
n
,
o
r
class
1
,
d
en
o
tin
g
m
alig
n
an
t
.
C
las
s
if
icatio
n
o
u
tco
m
es
o
b
tain
ed
v
ia
s
tr
atif
ied
th
r
ee
‑
f
o
ld
c
r
o
s
s
‑
v
alid
atio
n
ar
e
s
u
m
m
ar
iz
ed
in
T
ab
le
2
,
th
e
co
n
f
u
s
io
n
m
atr
ix
is
d
ep
icted
in
Fig
u
r
e
5
,
a
n
d
th
e
r
ec
eiv
er
o
p
er
atin
g
ch
ar
ac
ter
is
tic
(
R
OC
)
p
lo
t
is
illu
s
tr
ated
i
n
Fig
u
r
e
6
.
T
ab
le
2
.
C
lass
if
icatio
n
p
er
f
o
r
m
an
ce
m
etr
ics ac
r
o
s
s
th
r
ee
cr
o
s
s
-
v
alid
atio
n
f
o
ld
s
M
e
t
r
i
c
s
F
o
l
d
I
F
o
l
d
I
I
F
o
l
d
I
I
I
M
e
a
n
+
/
-
S
t
d
A
c
c
u
r
a
c
y
(
%)
9
7
.
5
6
9
8
.
1
6
9
4
.
4
8
9
6
.
7
3
+
/
-
1
.
9
7
S
e
n
s
i
t
i
v
i
t
y
(
%)
4
0
.
0
0
7
5
.
0
0
6
0
.
0
0
5
8
.
3
3
+
/
-
1
7
.
5
6
S
p
e
c
i
f
i
c
i
t
y
(
%)
9
9
.
3
7
9
8
.
7
4
9
5
.
5
7
9
7
.
8
9
+
/
-
2
.
0
4
P
r
e
c
i
s
i
o
n
(
%)
6
6
.
6
7
6
0
.
0
0
3
0
.
0
0
5
2
.
2
2
+
/
-
1
9
.
5
3
F1
-
S
c
o
r
e
(
%)
5
0
.
0
0
6
6
.
6
7
4
0
.
0
0
5
2
.
2
2
+
/
-
1
3
.
4
7
R
O
C
-
A
U
C
(
%)
9
5
.
9
7
9
2
.
7
7
9
5
.
0
6
9
4
.
6
0
+
/
-
1
.
6
5
T
h
e
m
o
d
el
o
b
tain
ed
an
o
v
er
a
ll
ac
cu
r
ac
y
o
f
9
6
.
7
3
%
with
a
s
tan
d
ar
d
d
ev
iatio
n
o
f
1
.
9
7
%
ac
r
o
s
s
th
e
th
r
ee
f
o
l
d
s
.
Ad
d
itio
n
ally
,
th
e
m
o
d
el
o
b
tain
ed
a
s
p
ec
if
icity
o
f
9
7
.
8
9
%,
wh
ic
h
d
em
o
n
s
tr
ates
its
h
ig
h
ca
p
ac
ity
t
o
id
en
tify
b
en
ig
n
n
o
d
u
les
co
r
r
e
ctly
.
C
lin
ically
,
h
ig
h
s
p
ec
if
icity
im
p
lies
f
ewe
r
f
alse
p
o
s
iti
v
e
r
esu
lts
,
th
er
eb
y
r
ed
u
cin
g
th
e
n
ee
d
f
o
r
in
v
asiv
e
b
io
p
s
ies an
d
m
in
im
izin
g
u
n
n
ec
ess
ar
y
p
atien
t a
n
x
iety
in
m
ed
ical
en
v
ir
o
n
m
e
n
ts
.
C
o
n
v
er
s
ely
,
th
e
s
en
s
itiv
ity
o
b
t
ain
ed
b
y
th
e
m
o
d
el
was
co
m
p
ar
ativ
ely
lo
w
at
5
8
.
3
3
%.
T
h
e
p
r
im
ar
y
r
ea
s
o
n
f
o
r
th
is
lo
wer
s
en
s
itiv
ity
is
th
e
in
h
er
en
t c
lass
im
b
alan
ce
with
in
t
h
e
d
ataset.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
16
,
No
.
5
,
Octo
b
e
r
20
26
:
2
8
0
6
-
2
8
1
8
2814
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
m
atr
i
x
s
h
o
win
g
th
e
class
if
icatio
n
r
esu
lts
Fig
u
r
e
6
.
R
OC
cu
r
v
e
s
h
o
win
g
th
e
d
is
cr
im
in
ativ
e
ab
ilit
y
o
f
th
e
class
if
icatio
n
m
o
d
el
3
.
3
.
Crit
ica
l
a
na
ly
s
is
o
f
s
ens
i
t
iv
it
y
I
t
is
im
p
o
r
tan
t
to
d
is
cu
s
s
th
e
m
ea
n
s
en
s
itiv
ity
o
f
5
8
.
3
3
%
±
1
7
.
5
6
%.
I
n
ter
m
s
o
f
d
etec
tin
g
th
y
r
o
i
d
ca
n
ce
r
,
lo
w
s
en
s
itiv
ity
m
ea
n
s
th
at
f
alse
n
eg
ativ
es
(
th
e
d
etec
tio
n
o
f
m
alig
n
a
n
t
n
o
d
u
les)
ar
e
n
o
t
g
o
o
d
.
I
n
o
th
er
wo
r
d
s
,
a
late
d
iag
n
o
s
is
o
f
th
e
d
is
ea
s
e
will
lead
to
th
e
d
ev
elo
p
m
en
t
o
f
th
e
d
is
ea
s
e
an
d
a
s
m
aller
n
u
m
b
er
o
f
way
s
to
tr
ea
t
th
e
d
is
ea
s
e
in
th
e
f
u
tu
r
e.
T
h
e
r
ea
s
o
n
f
o
r
th
e
l
o
w
s
en
s
itiv
ity
is
th
e
h
ig
h
im
b
alan
ce
o
f
class
es
in
th
e
d
ata,
s
in
ce
th
e
r
atio
o
f
b
en
ig
n
an
d
m
alig
n
an
t t
u
m
o
r
s
in
t
h
e
d
ata
is
3
4
:1
.
3
.
4
.
Co
m
pa
riso
n
wit
h e
x
is
t
ing
s
t
ud
ies
T
ab
le
3
o
u
tlin
es
h
o
w
o
u
r
n
e
wly
d
ev
elo
p
ed
f
r
am
ewo
r
k
p
er
f
o
r
m
s
r
elativ
e
to
ea
r
lier
m
o
d
e
ls
u
s
ed
f
o
r
th
y
r
o
id
n
o
d
u
le
class
if
icatio
n
.
T
ab
le
3
o
u
r
p
r
o
p
o
s
ed
two
‑
s
tag
e
f
r
am
ewo
r
k
attain
e
d
th
e
to
p
class
if
icatio
n
p
er
f
o
r
m
an
ce
,
with
a
v
al
u
e
o
f
9
6
.
7
3
%,
th
er
eb
y
s
u
r
p
ass
in
g
all
co
m
p
etin
g
m
o
d
els
ev
alu
at
ed
ac
r
o
s
s
d
if
f
er
en
t
d
ataset
s
izes.
Ag
u
s
tin
et
a
l.
[
2
5
]
p
r
e
v
io
u
s
ly
ass
ess
ed
b
a
s
elin
e
m
o
d
els
o
n
th
e
T
DI
D
d
ataset,
wh
er
e
th
e
s
tan
d
ar
d
U‑
Net
[
1
7
]
ac
h
iev
e
d
an
ac
cu
r
ac
y
o
f
9
1
.
0
0
%,
w
h
ile
th
e
a
tten
tio
n
U‑
Net
[
1
8
]
r
ea
ch
e
d
8
6
.
0
0
%.
Alth
o
u
g
h
th
e
a
tten
tio
n
U‑
Net
is
ef
f
ec
tiv
e
at
s
eg
m
e
n
tin
g
th
y
r
o
id
n
o
d
u
les
b
y
s
u
p
p
r
ess
in
g
b
ac
k
g
r
o
u
n
d
n
o
is
e
th
r
o
u
g
h
its
atten
tio
n
m
ec
h
an
i
s
m
,
it
is
less
ef
f
ec
tiv
e
at
lear
n
in
g
th
e
h
ig
h
er
‑
lev
el
f
ea
tu
r
es
r
eq
u
ir
e
d
to
r
eliab
ly
s
ep
ar
ate
b
en
i
g
n
f
r
o
m
m
alig
n
a
n
t
lesi
o
n
s
wh
en
u
s
ed
as
a
s
tan
d
alo
n
e
class
if
ier
.
T
o
o
v
e
r
co
m
e
th
is
lim
itatio
n
,
o
u
r
f
r
am
ewo
r
k
s
ep
ar
ates
th
e
two
task
s
.
I
n
th
e
f
ir
s
t
s
tag
e,
th
e
Atten
tio
n
U‑
Net
s
er
v
es
s
o
lely
to
p
er
f
o
r
m
p
r
ec
is
e
n
o
d
u
le
s
eg
m
e
n
tatio
n
,
af
ter
wh
ich
th
e
s
eg
m
en
ted
r
e
g
io
n
s
ar
e
p
ass
ed
to
a
d
ed
icate
d
C
NN
in
th
e
s
ec
o
n
d
s
tag
e,
wh
ich
ex
tr
ac
ts
m
o
r
e
d
is
cr
im
in
ativ
e
f
ea
tu
r
es
f
o
r
class
if
icatio
n
.
T
h
is
c
o
m
b
in
ati
o
n
allo
ws
t
h
e
s
y
s
tem
to
e
x
p
lo
it
th
e
ad
v
an
tag
es
o
f
b
o
th
ar
c
h
itectu
r
es,
y
ield
in
g
im
p
r
o
v
em
e
n
ts
o
f
5
.
7
3
%
o
v
e
r
th
e
co
n
v
e
n
tio
n
al
U‑
Net
an
d
1
0
.
7
3
%
o
v
er
th
e
s
tan
d
alo
n
e
at
ten
tio
n
‑
b
ased
U‑
Net
r
e
p
o
r
ted
b
y
Ag
u
s
tin
et
a
l.
[
2
5
]
.
T
a
b
l
e
3
.
C
o
m
p
a
r
a
ti
v
e
a
n
a
l
y
s
i
s
o
f
t
h
e
p
r
o
p
o
s
e
d
f
r
a
m
e
w
o
r
k
w
i
th
e
x
i
s
t
i
n
g
s
t
u
d
i
es
f
o
r
t
h
y
r
o
i
d
n
o
d
u
l
e
c
l
a
s
s
i
f
ic
a
t
i
o
n
.
A
c
c
u
r
a
c
y
(
%
)
d
e
n
o
t
es
t
h
e
o
v
e
r
a
l
l
c
l
ass
i
f
i
c
at
i
o
n
a
c
c
u
r
a
c
y
r
e
p
o
r
t
e
d
b
y
e
a
c
h
s
t
u
d
y
o
n
t
h
e
i
r
r
es
p
e
c
t
i
v
e
d
a
t
as
e
t
S
t
u
d
y
M
e
t
h
o
d
D
a
t
a
s
e
t
S
i
z
e
A
c
c
u
r
a
c
y
(
%)
N
u
g
r
o
h
o
e
t
a
l
.
[
1
]
G
LC
M
+
S
V
M
39
8
9
.
7
4
K
a
l
e
e
t
a
l
.
[
6
]
G
LC
M
+
S
V
M
85
8
2
.
2
8
G
u
a
n
e
t
a
l
.
[
1
5
]
I
n
c
e
p
t
i
o
n
-
v3
2
8
3
6
9
0
.
5
0
K
i
m
e
t
a
l
.
[
1
6
]
R
e
sN
e
t
5
0
1
0
4
8
8
2
.
0
0
K
i
m
e
t
a
l
.
[
1
6
]
D
e
n
seN
e
t
2
0
1
1
0
4
8
8
3
.
0
0
K
i
m
e
t
a
l
.
[
1
6
]
Ef
f
i
c
i
e
n
t
N
e
t
v
2
1
0
4
8
8
1
.
0
0
A
g
u
s
t
i
n
e
t
a
l
.
[
2
5
]
U
-
N
e
t
2
9
8
9
1
.
0
0
A
g
u
s
t
i
n
e
t
a
l
.
[
2
5
]
A
t
t
e
n
t
i
o
n
U
-
N
e
t
2
9
8
8
6
.
0
0
R
a
s
h
e
d
e
t
a
l
.
[
2
6
]
R
e
sN
e
t
1
0
1
(
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
)
4
8
3
9
4
.
7
5
R
a
s
h
e
d
e
t
a
l
.
[
2
6
]
Ef
f
i
c
i
e
n
t
N
e
t
B
0
(
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
)
4
8
3
9
3
.
0
9
D
o
n
g
e
t
a
l
.
[
2
1
]
U
N
e
t
+
+
+
D
u
a
l
-
p
a
t
h
A
t
t
e
n
t
i
o
n
4
8
0
9
4
.
5
0
Zh
a
n
g
e
t
a
l
.
[
2
7
]
Y
O
LO
v
3
5
0
0
8
4
.
0
0
Li
e
t
a
l
.
[
2
8
]
V
i
T
-
b
a
se
d
n
o
m
o
g
r
a
m D
L
m
o
d
e
l
2
2
4
7
8
8
.
6
0
Pr
o
p
o
sed
Fr
a
m
e
w
o
r
k
A
t
t
e
n
t
i
o
n
U
-
N
e
t
+ C
N
N
4
8
0
9
6
.
7
3
3
.
5
.
T
ra
ini
ng
a
nd
co
nv
er
g
en
ce
a
na
ly
s
is
Fig
u
r
es
7
a
n
d
8
s
h
o
w
th
e
l
ea
r
n
in
g
cu
r
v
e
f
o
r
t
r
ain
in
g
a
n
d
v
alid
atio
n
p
er
f
o
r
m
an
ce
i
n
ter
m
s
o
f
ac
cu
r
ac
y
an
d
lo
s
s
f
o
r
s
u
b
s
eq
u
en
t
ep
o
ch
s
.
Fro
m
th
ese
f
ig
u
r
e
s
,
it
ca
n
b
e
o
b
s
er
v
ed
th
at
th
er
e
is
co
n
v
er
g
en
ce
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Dete
ctio
n
a
n
d
cla
s
s
ifica
tio
n
o
f th
yro
id
d
is
ea
s
es u
s
in
g
u
ltr
a
s
o
u
n
d
…
(
K
a
ma
l
S
u
b
e
d
i
)
2815
th
e
n
etwo
r
k
with
o
u
t
o
v
er
f
itti
n
g
.
T
h
e
ap
p
licatio
n
o
f
r
e
g
u
lar
izatio
n
ap
p
r
o
ac
h
es
lik
e
Mix
U
p
[
2
2
]
an
d
C
u
tMix
[
2
3
]
,
b
atch
n
o
r
m
aliza
tio
n
,
an
d
d
r
o
p
o
u
t
allo
wed
u
s
to
h
av
e
a
s
im
ilar
p
er
f
o
r
m
an
ce
b
et
wee
n
tr
ain
in
g
an
d
v
alid
atio
n
d
atasets
.
Mo
r
eo
v
er
,
th
e
ap
p
r
o
ac
h
o
f
R
ed
u
ce
L
R
On
Plateau
was
also
p
o
s
itiv
e
s
in
ce
it
r
ed
u
ce
d
th
e
lear
n
in
g
r
ate
a
f
ter
th
e
v
alid
atio
n
lo
s
s
b
ec
am
e
s
tag
n
an
t.
Fig
u
r
e
7
.
T
r
ain
in
g
ac
c
u
r
ac
y
v
s
.
v
alid
atio
n
ac
c
u
r
ac
y
cu
r
v
e
Fig
u
r
e
8
.
T
r
ain
in
g
l
o
s
s
an
d
v
al
id
atio
n
lo
s
s
cu
r
v
es
3
.
6
.
Q
ua
lit
a
t
iv
e
v
is
ua
l a
na
ly
s
is
T
h
e
r
aw
im
ag
es
o
f
th
e
th
y
r
o
i
d
u
ltra
s
o
u
n
d
ar
e
illu
s
tr
ated
in
Fig
u
r
e
9
.
Fro
m
th
ese
im
ag
es,
o
n
e
ca
n
u
n
d
er
s
tan
d
th
at
t
h
er
e
a
r
e
s
ev
e
r
al
ch
allen
g
es
in
th
y
r
o
id
u
ltra
s
o
u
n
d
s
s
u
ch
as
lo
w
c
o
n
tr
ast,
s
p
ec
k
le
n
o
is
e,
a
n
d
v
is
u
al
am
b
ig
u
ity
b
etwe
en
b
en
i
g
n
an
d
m
alig
n
an
t
T
u
m
o
r
s
.
T
h
e
r
esu
lts
o
f
class
if
icatio
n
ar
e
s
h
o
wn
in
Fig
u
r
e
1
0
.
Her
e,
r
eg
io
n
s
m
ar
k
e
d
with
g
r
e
en
b
o
r
d
e
r
s
r
ep
r
esen
t b
en
ig
n
c
o
n
d
itio
n
s
,
wh
er
ea
s
th
o
s
e
m
ar
k
ed
with
r
ed
b
o
r
d
e
r
s
s
h
o
w
m
alig
n
a
n
t
co
n
d
itio
n
s
.
B
ar
ch
ar
ts
o
f
p
r
o
b
ab
ilit
ies
r
ep
r
e
s
en
t
co
n
f
id
e
n
ce
le
v
el
o
f
t
h
e
p
r
ed
ictio
n
s
m
a
d
e
b
y
th
e
alg
o
r
ith
m
.
As ex
p
ec
ted
,
t
h
e
co
n
f
id
e
n
ce
lev
els f
o
r
b
en
ig
n
ca
s
es a
r
e
h
ig
h
er
th
a
n
m
alig
n
a
n
t c
ases
.
Fig
u
r
e
9
.
E
x
am
p
le
r
aw
t
h
y
r
o
id
u
ltra
s
o
u
n
d
im
ag
es f
r
o
m
th
e
d
ataset
3
.
7
.
Clini
ca
l im
pli
ca
t
io
ns
a
n
d lim
it
a
t
io
ns
Ov
er
all,
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
h
ig
h
l
y
p
r
o
m
is
in
g
in
ter
m
s
o
f
b
ein
g
ap
p
lied
in
clin
ical
s
ettin
g
s
,
esp
ec
ially
in
lo
w
-
r
eso
u
r
ce
a
r
e
as
wh
er
e
th
er
e
is
s
h
o
r
tag
e
o
f
tr
ain
ed
r
a
d
io
lo
g
is
ts
.
Sin
ce
th
e
s
y
s
tem
h
as
a
v
er
y
g
o
o
d
lev
el
o
f
s
p
ec
if
icity
o
f
9
7
.
8
9
%,
o
n
e
ca
n
b
e
s
u
r
e
th
at
t
h
er
e
will
b
e
n
o
ca
s
es
o
f
wr
o
n
g
ly
lab
eled
b
en
i
g
n
n
o
d
u
les.
Hen
ce
,
th
is
ca
n
p
o
te
n
tially
d
ec
r
ea
s
e
th
e
n
ee
d
f
o
r
u
n
n
ec
ess
ar
y
b
io
p
s
y
p
r
o
ce
d
u
r
e
s
.
No
n
eth
eless
,
th
is
s
tu
d
y
h
as
s
ev
er
al
im
p
o
r
tan
t
lim
itatio
n
s
th
at
m
u
s
t
b
e
ac
k
n
o
wled
g
ed
:
i
)
Sm
all
d
ataset
s
ize
:
T
h
e
ev
alu
atio
n
was
co
n
d
u
cte
d
o
n
o
n
l
y
4
8
0
th
y
r
o
id
u
ltra
s
o
u
n
d
im
a
g
es,
wh
ich
is
s
u
b
s
tan
tially
s
m
aller
th
an
d
atasets
u
s
ed
in
co
m
p
ar
ab
le
s
tu
d
ies.
A
lar
g
er
d
ataset
wo
u
ld
p
r
o
v
id
e
m
o
r
e
r
eliab
le
p
e
r
f
o
r
m
an
ce
esti
m
at
es.
ii)
Sev
er
e
class
im
b
alan
ce
:
T
h
e
d
ataset
ex
h
ib
its
an
ex
tr
em
e
b
en
ig
n
-
to
-
m
alig
n
an
t
r
atio
o
f
ap
p
r
o
x
i
m
ately
3
4
:1
,
wh
ic
h
s
ig
n
if
ican
tly
b
iases
th
e
m
o
d
el
to
war
d
th
e
b
en
ig
n
class
an
d
d
i
r
ec
tly
co
n
tr
i
b
u
tes
to
t
h
e
lo
w
s
en
s
itiv
ity
.
iii)
L
ac
k
o
f
ex
ter
n
al
v
alid
atio
n
a
n
d
a
n
n
o
tatio
n
:
All
ex
p
e
r
im
en
ts
wer
e
co
n
d
u
cted
u
s
in
g
i
n
ter
n
al
th
r
ee
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
o
n
a
s
in
g
le
d
ataset.
T
h
e
ab
s
en
ce
o
f
ex
ter
n
al
v
alid
atio
n
o
n
in
d
e
p
en
d
e
n
t
d
atasets
lim
it
s
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.