Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
11
,
No.
1
,
Febr
uar
y
2021
, pp.
844
~
850
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v11
i
1
.
pp
844
-
850
844
Journ
al h
o
me
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
A deep l
earning
framewo
rk to det
ec
t
Co
vid
-
19
dise
ase via c
hest
X
-
r
ay and
CT scan im
ages
Moham
med
Y
. Kamil
Coll
ege of
Scie
n
ce
s,
Mus
ta
nsi
ri
yah
Univer
si
t
y
,
Ir
aq
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
22, 2
020
Re
vised
Ju
l
29
,
2020
Accepte
d
A
ug
10
, 201
9
COV
ID
-
19
disea
se
has
r
api
dl
y
s
pre
ad
al
l
over
th
e
world
a
t
th
e
b
egi
nning
of
thi
s
y
e
ar.
Th
e
h
ospita
ls
'
rep
orts
have
tol
d
that
low
sensiti
vity
of
RT
-
PC
R
te
sts
in
the
inf
ec
t
ion
ea
rl
y
sta
ge.
At
which
point
,
a
rap
id
a
nd
ac
cur
a
te
dia
gnostic
t
ec
h
niqu
e,
is
nee
d
e
d
to
det
ec
t
th
e
C
ovid
-
19
.
CT
has
bee
n
demons
tra
te
d
to
be
a
suc
ce
ss
ful
tool
in
the
d
i
agnosis
of
diseas
e.
A
deep
le
arn
ing
fra
m
e
work
ca
n
be
de
vel
oped
to
ai
d
in
eva
lu
ating
CT
exa
m
s
to
provide
di
agnosi
s,
thus
saving
tim
e
for
disea
se
c
ontrol
.
In
thi
s
w
ork,
a
d
ee
p
le
arn
ing
m
odel
was
m
odifi
ed
to
C
ovid
-
19
det
e
c
ti
on
via
f
eature
s
ext
ra
ct
ion
from
che
st
X
-
ra
y
and
CT
imag
es.
Ini
ti
a
lly
,
m
an
y
tra
nsfer
-
le
a
rn
ing
m
odel
s
have
appl
i
ed
an
d
compari
son
i
t
,
th
en
a
VG
G
-
19
m
odel
was
tu
ned
to
get
the
b
est
resul
ts
tha
t
c
an
be
ad
opte
d
in
th
e
di
sea
se
di
agnosis.
Diagnostic
per
form
anc
e
wa
s
assess
ed
for
al
l
m
odel
s
used
via
the
d
at
ase
t
th
at
included
1000
images.
T
he
VG
G
-
19
m
o
del
a
chieve
d
th
e
highe
st
accurac
y
of
99%
,
sensiti
vity
of
9
7.
4
%,
and
spe
ci
f
icit
y
of
99
.
4
%
.
Th
e
deep
learni
ng
and
image
proc
essing
demons
tra
te
d
high
per
form
anc
e
in
ea
rl
y
C
ov
id
-
19
det
ection.
It
show
s
to
be
an
auxi
liar
y
d
et
e
ct
ion
wa
y
fo
r
cl
inical
doc
to
rs
and
thus
cont
ribute
to the
cont
rol
of the
p
a
ndemic.
Ke
yw
or
d
s
:
Ar
ti
fici
al
i
ntell
igence
C
hest X
-
r
ay
C
ov
i
d
-
19
CT
scan
D
eep
lear
ning
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Moh
am
m
ed
Y. Kam
il
,
Coll
ege
of
Sci
ences,
Mustansiriy
ah
Un
i
ver
sit
y,
Ba
ghda
d,
Ir
a
q.
Em
a
il
:
m
80
y98@u
om
us
ta
ns
iriy
ah.
e
du.iq
1.
INTROD
U
CTION
Most
countrie
s
in
the
world
hav
e
bee
n
in
fec
ti
ng
with
Coro
nav
ir
us
disease
(C
ovid
-
19)
wit
h
2.
5
m
illi
on
co
nf
ir
m
ed
cases
[1]
.
The
ou
t
br
ea
k
was
declare
d
as
a
“p
ub
li
c
healt
h
em
erg
ency
of
i
nternat
ion
al
con
ce
r
n”
(
PHEIC)
by
the
"
World
Healt
h
Orga
nizat
ion
"
(
WH
O
).
C
ovid
-
19
has
widely
sp
rea
d
over
al
l
worl
d
since
on
Ja
nuary
30,
20
20
[2]
.
It
is
a
hi
gh
ly
co
ntagi
ous
pe
rs
on
-
to
-
pe
rson
tra
ns
m
is
sible
an
d
pne
um
on
ia
cause
d
[3]
.
B
ased
on
the
WHO'
s
repor
t
s,
the
m
or
ta
li
t
y
rate
has
2
-
3%
of
pe
ople
because
of
th
e
virus
.
In
t
he
ab
sence
of
a
preve
ntiv
e
vaccine
for
C
ov
i
d
-
19
dise
ase.
It
is
esse
nt
ia
l
to
diagnost
ic
te
sti
ng
at
an
early
sta
ge
based
on
crit
eria
as
cl
inica
l
sym
pto
m
s,
"R
eve
rse
-
tra
ns
cri
ption
poly
m
erase
chai
n
r
eact
ion
"
(RT
-
P
CR
),
so
as
to
isolat
e
the
infected
people
i
m
m
edi
at
el
y
[4]
.
Ho
w
ever,
there
are
rep
ort
s
sho
wing
the
RT
-
PC
R
te
st
m
igh
t
no
t
be
enou
gh
se
ns
it
ive
f
or
ea
rly
de
te
ct
ion
[
5,
6]
.
So,
com
pu
te
d
tom
og
ra
phy
(CT)
a
pp
ea
re
d
as
a
noni
nv
a
sive
i
m
aging
a
ppr
oach
that
ca
n
detect
spe
c
ific
le
sion
s
in
t
he
lu
ng
ass
oc
ia
te
d
with
C
ovid
-
19
disease
[
7]
.
Chest
CT
is
a
di
agnostic
too
l
for
pne
um
on
ia
and
C
ov
i
d
-
19
,
is
easy
to
do,
an
d
can
outp
ut
an
accurate
diag
nosis.
It
sho
ws
perfect
ra
diogra
phic
featu
r
es
in
al
l
C
ov
i
d
-
19
im
ages,
as
m
ult
ifocal
patc
hy
consolidat
io
n,
gro
und
-
glass
opaci
ti
es,
and
m
ulti
fo
cal
patch
y
con
s
olidati
on
[
8]
.
It
has
be
en
note
d
that
s
ever
a
l
patie
nts
ha
d
a
neg
at
ive
RT
-
P
CR
te
st
wh
il
e
in
the
Chest
C
T
scan
hav
i
ng
po
sit
ive
[9]
.
A
rtific
ia
l
intel
li
g
ence
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A d
ee
p
le
ar
ni
ng fr
am
ew
or
k t
o detec
t
C
ovi
d
-
19
disease
…
(
Mo
hamm
e
d Y.
Kam
il
)
845
(AI)
in
vo
l
ving
m
achine
le
arn
ing
(M
L
)
and
deep
le
ar
ning
(D
L
)
has
gra
nd
evide
nce
su
c
cess
in
the
m
e
dical
i
m
age
under
st
and
i
ng
sco
pe
du
e
t
o
it
s
high
stre
ng
t
h
of
cl
assifi
cat
ion
and
featu
re
extracti
on
[
10
,
11]
.
Conv
olu
ti
onal
neural
net
wor
k
(C
NN)
has
widely
ap
plied
to
detect
a
nd
viral
pne
um
on
ia
an
d
diff
e
r
entia
te
bacteria
l
in
ch
est
rad
i
ogra
phs.
CN
N
has
powe
rful
in
fea
ture
e
xtracti
on
,
in
vo
l
ves
sp
at
ia
l
filt
ers
that
colle
ct
inf
or
m
at
ion
on
the
str
uctu
re
[
12
]
.
D
octo
rs
usual
ly
us
e
X
-
ra
ys
to
dia
gnos
e
lu
ng
in
flam
m
a
ti
on
a
nd
pne
um
on
ia
.
A
ll
hosp
it
al
s
ha
ve
X
-
ra
y
im
a
ging,
it
c
ou
l
d
be
possible
t
o
use
X
-
rays
to
an
al
yz
e
the
lungs
of
C
ovid
-
19
pa
ti
ent.
But X
-
ray a
nal
ysi
s r
eq
uires
takes si
gn
i
ficant
tim
e and
a
rad
i
ology ex
pe
rt
[
13]
.
Li
et
al.
[
14]
de
velo
ped
a
ne
ur
al
netw
ork
m
od
el
is
cal
le
d
(CO
N
V
Net)
to
the
de
te
ct
io
n
of
C
ov
i
d
-
19
via
chest
CT
i
m
ages.
The
data
was
use
d
c
ons
ist
s
of
43
56
i
m
ages
for
33
22
patie
nts
a
ge
around
49±
15
ye
ars.
T
hey
ba
sed
on
the
Re
stNet
50
m
od
el
to
de
vel
op
the
al
gorith
m
cou
ld
us
e
a
rob
us
t
dia
gnosi
s
for
C
ov
i
d
-
19
.
T
he
sensiti
vity
and
s
pecifici
ty
of
t
he
w
ork
wer
e
repo
rted
as
90
%
a
nd
96%
,
res
pe
ct
ively
.
Bhan
dar
y
et
al
.
[15]
pro
pose
d
a
m
od
ifie
d
Al
exN
et
m
od
el
by
us
in
g
a
sup
port
vector
m
achine
a
nd
c
om
par
ed
it
against
Softm
a
x.
T
hey im
plem
ented
on
1018 im
ages o
f
the
ch
est
X
-
r
ay
b
e
longin
g
to
the
LID
C
-
I
DRI
da
ta
base
to
detect
pneu
m
on
ia
and
c
an
cer.
T
he
al
gori
thm
’s
per
f
orm
ance
has
e
valu
at
ed,
a
nd
it
ha
s
97.
27%
acc
uracy
,
98.09%
se
ns
it
iv
it
y,
and
95.
63
%
sp
eci
fici
ty
.
Wang
et
al.
[16]
m
od
ifie
d
th
e
In
cepti
on
m
od
el
via
the
transf
e
r
-
le
arn
in
g
m
et
ho
d
to
prov
i
de
a
cl
in
ic
al
diagno
sis
that
cou
l
d
a
id
C
ov
i
d
-
19
’s
gr
a
phic
al
featu
res
ext
racti
on.
They
hav
e
c
ollec
te
d
1119
im
ages
(CT
scan
)
of
dia
gnos
e
d
with
vi
ral
pn
eum
on
ia
and
C
ov
i
d
-
19
ca
se
s.
The
evaluati
ng
data
set
sh
owed
a
n
accuracy
of
79
.3
%
with
a
sen
sit
ivit
y
of
67
%
and
s
pecifi
ci
ty
of
83%.
Xu
et
al.
[17]
est
ablishe
d
a
scree
ning
s
yst
e
m
fo
r
the
detect
ion
of
C
ov
i
d
-
19
diseas
e
based
on
dee
p
le
arn
i
ng
te
ch
niques
us
in
g
CT
i
m
ages.
This
syst
em
was
bu
il
t
on
Re
sn
et
18
m
od
el
,
al
so
c
ollec
te
d
the
i
m
ages
fr
om
three
ho
sp
it
al
s
in
China,
a
bo
ut
618
im
ages.
The
im
ple
m
e
nts
res
ult
of
fe
r
ed
that
accu
ra
cy
,
sensiti
vity
,
and
s
pecifici
ty
was
86.7%,
93.1
%
,
90%,
re
sp
ect
i
vely
.
D
hur
gh
a
m
et
al.
[18]
de
velo
ped
a
m
achine
le
ar
ning techn
iq
ue
t
hat
ai
ded
in
detect
ing
the
infecti
on
e
f
fici
ently
fo
r
C
ov
i
d
-
19
dis
e
ase
by
us
in
g
chest
CT
i
m
ages.
They
us
ed
the
FFT
-
Ga
bor
m
et
ho
d
an
d
achieve
d
an
acc
ur
acy
of
95.
37%,
sp
eci
fici
ty
94.76%
,
an
d
sensiti
vity
95
.
99%
via
exam
ined
on a
dataset
consist
of 47
0
im
ages f
or
C
ov
i
d
-
19
pa
ti
ents.
The
m
otivati
on
in
t
his
w
ork,
a
n
a
uto
m
ated
dia
gnos
is
syst
e
m
dev
el
opm
ent
is
able
to
analy
z
e
the
le
sion
fro
m
rad
iolo
gy
im
ages
and
ai
de
do
in
g
a
rap
i
d
an
d
accur
at
e
diagnosis.
T
he
rest
of
this
stud
y
consi
sts
of:
S
ect
ion
2
pres
ents
the
m
et
ho
dolo
gy
of
t
he
pro
posed
de
ep
le
ar
ning
fr
am
ewo
r
k.
D
at
aset
Inform
at
io
n
an
d
pe
rfor
m
ance
evaluati
on
m
e
tric
s
al
so
res
ul
ts,
and
disc
us
si
on
s
a
re
pr
ese
nt
ed
an
d
descr
i
be
d
in
s
ect
ion
3. I
n
th
e e
nd, t
he
c
onc
lusio
n
is s
how
n
in
s
ect
io
n 4.
2.
MA
TE
RIA
L
S
AND MET
H
ODS
2.1.
C
N
N
m
odel
In
deep
le
ar
ni
ng,
CN
N
is
a
cl
ass
of
deep
neural
netw
ork
s
that
at
tem
pts
to
si
m
ulate
the
process
of
analy
zi
ng
im
a
ges
via
the
vis
ual
co
rtex
(ce
r
ebr
al
c
or
te
x)
in
the
brai
n
[
19]
.
I
n
t
he
pa
st,
m
os
t
researchers
in
com
pu
te
r
visio
n
e
xtracted
th
e
featu
res
by
hand
-
cra
fted
f
or
bette
r
res
ults
in
cl
assifi
cat
ion
[
20
]
.
N
owadays
,
CNN
pe
r
form
s
the
resp
ect
ive
work
of
feat
ure
extracti
on
a
uto
m
at
ic
ally
throu
gh
the
trai
ni
ng
sta
ge
base
d
on
poolin
g
la
ye
rs
and
c
onvoluti
on
la
ye
rs
[
21
]
.
Conv
olu
ti
onal
la
ye
rs
consi
st
of
var
i
ous
ty
pe
s
of
filt
ers
th
at
are
trai
ned
acco
rd
i
ng
to
t
he
cl
ass
ific
at
ion
go
al
.
Wh
il
e
the
po
ol
ing
la
ye
rs
a
re
doin
g
re
duci
ng
the
dim
ensio
n
of
featur
e
e
xtracti
on
a
nd
retai
n
t
he
siz
e
an
d
sh
a
pe
of
an
im
age.
The
re
are
m
a
ny
CNN
m
od
e
ls
popu
la
rly
be
cause
of
their
e
ff
ic
ie
ncy
and
r
ob
us
t
ness
in
the
fiel
d
of
patte
r
n
rec
ogniti
on
[22]
.
I
t
is
us
ed
in
m
a
ny
sco
pes,
es
pe
ci
all
y
in
the
cl
assifi
c
at
ion
of
m
edical
i
m
ages
[23]
.
He
nce,
t
he
VGG
-
19
one
of
t
he
m
od
el
s
us
e
d
in
our
work
is
il
lustrate
d
in
Fi
gure
1.
Figure
1. Mo
d
ifie
d V
GG
-
19
m
od
el
arch
it
ect
ur
e
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
1
,
Febr
uar
y
2021
:
844
-
850
846
VGG
-
19
arc
hitec
ture
co
ns
ist
s
of
co
nvol
ution
al
,
poolin
g,
and
f
ully
con
necte
d
la
ye
rs
.
It
con
ta
i
ns
a
total
of
25
la
ye
rs.
The
in
pu
t
i
m
age
siz
e
is
224×
224
pi
xels.
The
filt
er
siz
e
is
3×3
pix
el
s
for
the
co
nvolut
ion
a
l
la
ye
r
(ReLU
).
The
m
ax
-
po
oling
la
ye
r
is
us
e
d
to
re
duce
the
cost
a
nd
siz
e
of
the
data.
I
n
the
fi
nal
arc
hitec
ture
,
the
la
ye
rs
have
a
fu
ll
y
-
conn
ect
ed
la
ye
r
(F
la
tt
en
and
Re
L
U)
with
a
dro
pout
of
(
0.5
)
that
m
et
ho
d
to
reduce
ov
e
rf
it
ti
ng a
nd an
outp
ut lay
er w
it
h
s
of
tm
ax
act
ivati
on
.
2.2.
Perf
orm
an
ce
e
va
lu
ati
on
m
e
trics
A
var
ie
ty
of
m
et
rics
hav
e
us
e
d
agr
eea
ble
by
the
sci
entifi
c
com
m
un
it
y
to
e
valuate
the
perform
ance
of
the
cl
assifi
cat
ion
syst
em
to
detect
lu
ng
di
sease
[24]
.
T
he
pe
rfor
m
ance
of
this
stu
dy
is
e
valuated
with
the
co
nfusi
on
m
at
rix
base
d
on
the
esse
ntial
pa
ram
et
ers
us
ed:
tr
ue
-
posit
ive
(TP),
tr
ue
-
neg
at
ive
(
TN
),
false
-
po
sit
ive
(
FP
),
and
false
-
neg
a
ti
ve
(F
N
).
By
these
par
am
et
ers,
it
can
be
c
al
culat
ed
valid
it
y
m
et
rics,
su
ch
as
accuracy,
se
nsi
ti
vity
,
sp
eci
fici
ty
,
F1
scor
e,
pr
eci
sio
n.
Al
so
,
ot
her
valu
es
false
-
ne
gati
ve
rate
(FNR)
,
false
po
sit
ive
rate
(
FPR),
fa
lse
dis
cov
e
ry
rate
(FDR),
false
om
i
ssion
rate
(FO
R),
m
atthew
s
correla
ti
on
c
oe
ff
ic
ie
nt
(MCC
),
bo
okm
aker
inf
or
m
edn
e
ss
(BM)
and
m
ark
ed
ne
ss
(MK
)
are
al
so
com
pu
te
d.
The
m
at
hem
at
ic
a
l
form
ulae o
f
t
he
se m
easur
es c
an be e
xpresse
d
as
[
25
]
:
=
+
+
+
+
(1)
=
+
(2)
=
+
(3)
1
=
2
2
+
+
(4)
=
+
(5)
=
+
(6)
=
+
(
7
)
=
+
(
8
)
=
+
(
9
)
=
+
(
10
)
=
×
−
×
√
(
+
)
(
+
)
(
+
)
(
+
)
(1
1
)
=
+
−
1
(1
2
)
=
+
−
1
(1
3
)
3.
RESU
LT
S
A
ND D
I
SCUS
S
ION
S
In
this
sect
io
n,
the
res
ults
are
pr
ese
nted
f
or
l
ung
cl
assi
ficat
ion.
At
fi
rst,
th
e
dataset
us
e
d
is
presente
d,
and
i
nfor
m
at
ion
it
s,
then
,
the
m
et
rics
us
ed
ar
e
sh
ow
n
f
or
pe
rfor
m
ance
eval
uation,
as
well
as
detai
l
the
re
su
lt
s
of the im
ple
m
e
nted, al
so com
par
e
w
it
h ot
her relat
ed w
orks.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A d
ee
p
le
ar
ni
ng fr
am
ew
or
k t
o detec
t
C
ovi
d
-
19
disease
…
(
Mo
hamm
e
d Y.
Kam
il
)
847
3.1.
D
atase
t
i
nf
or
ma
tio
n
This
wo
rk
,
a
database
of
lun
g
disease
with
chest
X
-
ray
or
CT
im
ages
is
us
ed,
wh
ic
h
is
pu
blicl
y
avail
able
in
Re
f.
[2
6]
and
Re
f.
[23]
.
The
dataset
con
ta
ins
10
0
0
im
ages,
80
5
im
ages
of
no
rm
al
,
and
19
5
im
ages
of
C
ov
id
-
19
.
The
no
rm
al
as
Chest
X
-
ray
im
ages,
wh
il
e
C
ov
id
-
19
im
ages
con
sist
of
17
2
Chest
X
-
ray
and
2
3
Lun
g
CT
im
ages,
sh
ow
n
in
the
Table
1
.
All
im
ages
fo
r
C
ov
id
-
19
with
chest
X
-
ray
or
CT
im
ages,
are av
ai
la
bl
e
in 2
4
-
bit
RGB
-
scal
e in
JPEG
f
or
m
at
, w
it
h
a d
iffer
ent
siz
e. Chest
X
-
ray no
rm
al
(CXRN)
im
ages
wer
e
sel
ect
ed
fr
om
Gu
ang
zho
u
W
om
en
and
Chil
dr
en’
s
Me
dical
Ce
nter.
All
CXRN
im
aging
was
per
fo
rm
ed
as
a
m
ajo
r
a
sp
ect
of
patie
nts’
daily
care.
Be
fo
re
trai
nin
g
the
sy
ste
m
,
the
real
diagn
os
es
fo
r
the
im
ages
wer
e
evaluated
by
three
exp
ert
rad
iolog
ist
s.
The
CXRN
im
ages
are
avail
able
in
JPEG
fo
rm
at
,
and
a d
iffer
ent size
o
f
abo
ut 2
02
2
×21
29
to
10
88
×82
4.
I
m
ages s
a
m
ples u
sed
can b
e sh
ow
n
in Fig
ur
e
2.
(a)
(b)
(c)
Figure
2. Sam
ple o
f
lu
ng im
ag
e
,
(a
)
C
hest X
-
r
ay
norm
al
, (
b)
Chest
X
-
r
ay
C
ov
i
d
-
19
,
(c)
CT
im
age
C
ov
i
d
-
19
Table
1
.
Ca
te
gorize al
l i
m
ages th
at
hav
e
test
ed.
Dataset
C
lass
Nu
m
b
e
r
o
f
i
m
ag
es
Lun
g
CT
Co
v
id
-
19
23
Ch
est X
-
ra
y
Co
v
id
-
19
172
No
r
m
al
805
Total
1000
3.2.
F
ine
-
tuni
ng
th
e
V
GG
19 m
od
el
The
deep
le
arn
ing
syst
em
was
im
plem
ented
in
a
per
so
nal
com
pu
te
r
with
an
In
te
l
Core
i7
-
7700HQ
CPU
@
2.
81
GH
z,
Nv
idia
GeFo
rce
GTX
10
50
-
Ti
gr
aph
ic
card
s,
and
16
GB o
f
RAM, wo
rk
ing
o
n
a W
ind
ow
s
10
(6
4
-
bit)
op
erati
ng
sy
ste
m
,
and
im
plem
ented
fu
ll
y
in
Pyt
ho
n
la
ng
uag
e
via
Ker
as
li
br
ary
with
Op
en
CV
and
Tenso
rf
low
as
back
-
end
.
The
VG
G1
9
m
od
el
giv
en
in
Figu
re
1
was
trai
ned
at
80
%
and
validat
ed
on
20
%
fo
r
al
l
dataset
avail
able
i
m
ages.
Ba
sed
on
the
abo
ve,
20
0
im
ages
hav
e
us
ed
fo
r
the
validat
ion
set
and
the r
em
ai
nin
g
80
0
im
ages
fo
r
the
trai
nin
g
set
.
Ther
e w
ere
19
5
C
ov
id
-
19
and
8
05
no
rm
al
im
ages.
So
, th
e
rati
o
of
COVI
D
-
19
to n
or
m
al
im
ages in
the total datase
t was ar
ou
nd
2
4%.
The
e
xp
erim
ental
wo
rk
in
this
stud
y
is
div
ided
into
three
-
sta
ge.
In
the
first
ste
p,
al
l
im
ages
hav
e
pr
epr
ocessed
via
con
ver
ti
ng
it
to
the
RGB scal
e and
r
esi
zi
ng
it
to
22
4×2
24
p
ixels so
that t
he
im
ages ar
e read
y
as
an
inp
ut
to
the
CNN
m
od
el
.
Then
,
the
data
(im
ag
e
intensit
y)
wer
e
no
rm
al
iz
ed
by
con
ver
ti
ng
it
to
the
ran
ge
(0
,
1)
.
In
the
secon
d
ste
p,
since
the
nu
m
ber
of
trai
nin
g
im
ages
(d
at
a)
us
ed
in
ou
r
wo
rk
is
no
t
su
ff
ic
ie
nt
and
to
ensu
re
that
m
od
el
gen
erali
zes,
data
aug
m
entat
ion
has
per
fo
rm
ed
via
set
ti
ng
the
im
ag
e
ro
ta
ti
on
to
15
deg
rees
cl
ock
wise
ran
do
m
ly
.
In
the
third
ste
p,
transf
er
le
arn
ing
is
the
pr
ocess
of
ta
kin
g
a
netwo
rk
pr
e
-
trai
ned
on
a
dataset
and
util
iz
ing
it
to
reco
gn
iz
e
im
age
or
ob
j
ect
cat
ego
ries
it
was
no
t
trai
ned
on
.
W
hile
fine
-
tun
ing
req
uires
that
re
trai
nin
g
the
head
of
CNN
arch
it
ect
ur
e
to
reco
gn
iz
e
new
ob
j
ect
cl
asses
it
was
no
t
pr
im
aril
y
pr
epar
ed
fo
r.
In
this
wo
rk
has
us
ed
Fine
-
tun
ing
us
ing
Ker
as
via
a
m
ulti
-
ste
p
pr
ocess.
Firstl
y,
al
l
la
ye
rs
below
the
head
are
fr
ozen
in
the
netwo
rk
ensu
ring
that
the
back
war
d
pass
in
back
pr
op
agati
on
do
es
no
t
reach
it
.
Secon
dly,
the
fu
ll
y
con
nected
no
des
are
rem
ov
ed
at
the
end
of
the
netwo
rk
and
rep
la
ced
it
with
new
ly
init
ia
li
zed o
nes.
Th
en,
training
is starte
d
on
ly
f
or
the f
ully
co
nn
ect
ed
la
ye
r
head
s.
Figu
re
2
il
lustra
te
s
the
sam
ple
te
st
im
ages
of
chest
X
-
r
ay
and
lun
g
CT
with
a
no
rm
al
case
or
C
ov
id
-
19
disease.
This
dataset
con
sist
s
of
var
iou
s
dim
ension
s
of
im
ages.
W
her
efo
re,
the
im
ages
resizi
ng
is
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
1
,
Febr
uar
y
2021
:
844
-
850
848
pr
ocessed
to
red
uce
the
dim
ension
to
22
4
×22
4
×3
pix
el
s.
Fu
rther
,
im
age
a
ug
m
entat
ion
is
im
plem
ented
to
gr
ow
up
the
nu
m
ber
of
trai
nin
g
im
ages.
In
it
ia
ll
y,
a
pr
e
-
trai
ned
VG
G1
6,
VG
G1
9,
Xcep
ti
on
,
Re
sNet50
V2
,
Mob
il
eNetV2
,
NA
SN
et
Mob
il
e,
Re
sNet10
1V
2,
and
In
cepti
on
V3
is
us
ed
to
analy
ze
the
dataset
us
ed
and
com
par
ed
am
on
g
them
as
in
Tabl
e
2.
Fu
rther
m
or
e,
the
per
fo
rm
ance
of
al
l
DL
m
od
el
s
is
then
validat
ed
with
the o
ther
pr
edict
ive an
al
yt
ic
s p
aram
et
ers,
as sh
ow
n
in Tab
le
3
.
Table
2
.
Re
s
ults o
f
perform
ance ev
al
ua
ti
on
m
et
rics
Ap
p
roach
TP
TN
FP
FN
Accurac
y
Sen
sitiv
ity
Sp
ecif
icity
F1
sco
re
P
recisio
n
VGG1
6
36
159
2
3
0
.97
5
0
.92
3
0
.98
8
0
.93
5
0
.94
7
VGG1
9
38
160
1
1
0
.99
0
0
.97
4
0
.99
4
0
.97
4
0
.97
4
Xcepti
o
n
27
158
3
12
0
.92
5
0
.69
2
0
.98
1
0
.78
3
0
.90
0
Res
Net5
0
V2
16
157
4
23
0
.86
5
0
.41
0
0
.97
5
0
.54
2
0
.80
0
Mob
ileNetV2
15
156
5
24
0
.85
5
0
.38
5
0
.96
9
0.
508
0
.75
0
NASNet
Mob
ile
17
157
4
22
0
.87
0
0
.43
6
0
.97
5
0
.56
7
0
.81
0
Res
Net1
0
1
V2
21
158
3
18
0
.89
5
0
.53
8
0
.98
1
0
.66
7
0
.87
5
Incep
tio
n
V3
4
154
7
35
0
.79
0
0
.10
3
0
.95
7
0
.16
0
0
.36
4
Table
3
.
Re
s
ults o
f pr
e
dicti
ve a
naly
ti
cs p
ara
m
et
ers
Ap
p
roach
NPV
FNR
FPR
FDR
F
OR
MCC
BM
MK
VGG1
6
0
.98
1
0
.07
7
0
.01
2
0
.05
3
0
.01
9
0
.41
1
0
.91
1
0
.92
9
VGG1
9
0
.99
4
0
.02
6
0
.00
6
0
.02
6
0
.00
6
0
.43
0
0
.96
8
0
.96
8
Xcepti
o
n
0
.92
9
0
.30
8
0
.01
9
0
.10
0
0
.07
1
0
.34
3
0
.67
4
0
.82
9
Res
Net5
0
V2
0
.87
2
0
.59
0
0
.02
5
0
.20
0
0
.12
8
0
.24
2
0
.38
5
0
.67
2
Mob
ileNetV2
0.
867
0
.61
5
0
.03
1
0
.25
0
0
.13
3
0
.22
3
0
.35
4
0
.61
7
NASNet
Mob
ile
0
.87
7
0
.56
4
0
.02
5
0
.19
0
0
.12
3
0
.25
2
0
.41
1
0
.68
7
Res
Net1
0
1
V2
0
.89
8
0
.46
2
0
.01
9
0
.12
5
0
.10
2
0
.29
6
0
.52
0
0
.77
3
Incep
tio
n
V3
0
.81
5
0
.89
7
0
.04
3
0
.63
6
0
.18
5
0
.05
1
0
.05
9
0
.17
8
Figu
re
3
sh
ow
s
trai
nin
g
and
accuracy
fo
r
the
pr
e
-
trai
ned
VG
G1
9
m
od
el
.
The
m
od
el
was
fine
-
tun
ed
accord
ing
to
the
par
am
et
ers
exp
la
ined
abo
ve.
The
hig
hest
accuracy
value
was
ob
ta
ined
com
par
ed
to
oth
er
m
od
el
s,
see
T
able
1.
The
resu
lt
s
sh
ow
that
pr
e
-
trai
ned
m
od
el
s
can
ou
tpu
t
hig
h
acc
ur
acy
per
fo
rm
ances,
as
sh
ow
n
in
T
able
2.
Altho
ug
h
that,
the
VG
G1
9
was
achieved
the
top
aver
age
accuracy
thro
ug
h
this
valida
ti
on
at 99
% f
or
1
00
ep
och
s.
Figu
re
4
sh
ow
s the lo
ss accur
acy
att
ai
ned
f
or
the V
GG
19
m
od
el
.
Fig
ure
3. Acc
uracy
curve
to
t
r
ai
n
ing an
d vali
dation o
f
the VG
G19
m
od
el
f
or
100 ep
ochs
Fig
ure
4. Lo
ss
accuracy c
urve
to
trai
ning a
nd
validat
io
n of
t
he
VGG1
9
m
odel
Most
of
the
deep
le
arn
ing
m
od
el
s ar
e
the f
ocu
s o
f
m
edical
im
age d
ia
gn
os
is on
ly
o
n
accuracy. D
o
no
t
fo
cus
on
the
ti
m
e
per
fo
r
m
ance
of
m
od
el
s,
m
ay
be
m
easur
em
ent
of
cl
assifi
cat
ion
or
trai
nin
g
ti
m
e
cou
ld
be
decep
ti
ve.
Be
cause
it
dep
end
s
on
the
har
dw
are
com
pu
te
r
li
ke
CPU,
and
li
br
aries
us
ed.
In
this
wo
rk
,
the
har
dw
are,
li
br
aries,
and
la
ng
uag
e
us
ed
are
m
entioned
at
the
beg
inn
ing
of
this
sect
ion
.
Figu
re
5
sh
ow
s
the
per
fo
rm
ance
of
the
pr
edict
ion
ti
m
e
fo
r
al
l
m
od
el
s
us
ed.
It
ob
serv
ed
fr
om
Figu
res
5
and
6
that
the
hig
hest
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A d
ee
p
le
ar
ni
ng fr
am
ew
or
k t
o detec
t
C
ovi
d
-
19
disease
…
(
Mo
hamm
e
d Y.
Kam
il
)
849
accuracy
value
on
the
VG
G1
9
m
od
el
with
a
pr
edict
ion
ti
m
e
is
the
m
idd
le
of
the
scal
e
fo
r
al
l
m
od
el
s.
Con
ver
sel
y,
oth
e
r
resu
lt
s
are
ob
ta
ined
hig
h
pr
edict
ion
ti
m
e,
and
low
accuracy
fo
r
cl
assifi
cat
ion
in
ou
r
dataset
C
ov
id
-
19
im
ages
was
util
iz
ed.
The
cause
m
ay
be
the
nu
m
ber
of
hid
den
la
ye
rs
and
neu
ro
ns
per
hid
den
la
ye
r,
so
the level o
f
dr
op
ou
t.
Fig
ure
5
.
Rel
at
ive pre
dicti
on
t
i
m
e u
sing C
PU f
or
all
m
od
el
s u
sed
Fig
ure
6. Acc
uracy
v
al
ue
r
esu
lt
s
for
al
l t
he
m
od
el
s
i
m
ple
m
ented
Desp
it
e
the
no
velty
of
the
cov
id1
9
disease
wh
ic
h
we
hav
e
wo
rk
ed
on
,
there
is
a
sta
te
-
of
-
the
-
art
m
et
ho
d
im
plem
ented
on
cov
id1
9
im
ages.
Table
4
sh
o
ws
the
com
par
ison
of
oth
er
m
et
ho
ds
with
ou
r
wo
rk
.
The
com
par
ison
is
con
fined
to
the
ty
pe
of
m
od
el
and
the
nu
m
ber
of
im
ages
that
hav
e
been
trai
ned
and
te
ste
d.
As
well
as
cal
culat
e
the
m
os
t
pr
om
inent
par
am
et
ers
as
area
un
der
cur
ve
(A
UC),
accuracy
(A
CC
),
sen
sit
ivit
y
(S
EN)
,
and
sp
eci
fici
ty
(S
PE).
Ob
vio
us
ly
,
the
hig
hest
accuracy
in
oth
er
li
te
ratur
e
is
97
.2
7%,
wh
ic
h
was
ann
ou
nced
by
Bhand
ary
[15]
.
In
com
par
ison
with
the
li
te
ratur
e
resu
lt
s
by
diff
eren
t
autho
rs,
it
ob
serv
es
ou
r
tun
ing
m
od
el
h
as
pr
od
uce
d
a m
axim
um
accuracy of
9
9%.
Based o
n
the r
esults, we
cou
ld say that
o
ur
V
GG
19
m
od
el
o
utp
erf
or
m
ed
the o
ther
m
od
el
s in
cor
on
a
viru
s d
ise
ase diagn
os
is.
Ther
e
are
m
ulti
ple
CNN
m
od
el
s
us
ed
by
diff
eren
t
autho
rs
in
the
diagn
os
is
of
diseases
by
m
edical
im
aging
.
These
m
et
ho
ds
hav
e
so
m
e
li
m
it
at
ion
s.
The
m
os
t
im
po
rtant
li
m
it
at
ion
in
this
wo
rk
;
the
few
nu
m
ber
s
of
im
ages th
at
the m
od
el
w
as trained
on
, d
ue
to the d
ifficult
y i
n
ob
ta
ining
it
as the
pu
blic at
p
resen
t.
Table
4
.
C
om
par
iso
n of dif
fere
n
t m
od
el
s to
e
xisti
ng li
te
rature
wo
rk
d
ataset
i
m
ag
es
Mod
el
AUC
ACC
SEN
SPE
Li
[
1
4
]
p
ri
v
ate
4356
Res
tNet5
0
0
.96
-
90
96
Bh
an
d
ary
[
1
5
]
LI
DC
-
IDRI
1018
Alex
Net
0
.99
6
9
7
.27
9
8
.09
9
5
.63
W
an
g
[
1
6
]
p
rivate
1119
Incep
tio
n
0
.81
7
9
.3
83
67
Xu
[
1
7
]
p
rivate
618
Res
n
et1
8
-
8
6
.7
9
3
.1
90
Dh
u
rgh
a
m
[
1
8
]
p
rivate
470
FFT
-
G
ab
o
r
-
9
5
.37
9
5
.99
9
4
.76
p
rop
o
sed
p
rivate
1000
VGG1
9
-
99
9
7
.4
9
9
.4
4.
CONCL
US
I
O
N
The
early
dia
gnos
is
of
C
ovid
-
19
has
bee
n
co
ns
ide
re
d
chall
eng
i
ng
due
to
the
c
onseq
uen
ces
of
the
disease
spr
ead
to
so
ci
et
y.
Deep
le
ar
ning
te
chn
iq
ues
a
nd
so
ft
com
pu
ti
ng
sk
il
ls
would
ai
de
in
the
accuracy
and
a
cc
el
erati
on
of
the
dia
gnos
ti
c
proces
s.
In
t
his
stu
dy,
we
ha
d
offer
e
d
a
tu
ne
d
V
G
G19
m
od
el
th
at
cou
l
d
help
dia
gnos
i
s
Cov
id
-
19
a
uto
m
at
ic
ally.
All
i
m
ple
m
en
te
d
m
od
el
s
ha
ve
giv
e
n
good
resu
lt
s
at
chest
rad
i
ography.
But
the
tu
ne
d
VGG
19
m
od
el
has
produc
ed
bette
r
accu
racy
res
ults
th
an
oth
er
m
et
h
od
s
a
nd
ou
t
perform
ed
t
he
present
li
te
ratur
e
.
The
refo
re,
m
od
el
s
with
a
fine
-
t
un
i
ng
cou
ld
be
a
co
m
m
i
tt
ed
co
m
pu
te
r
-
ai
ded
diag
nosti
c syst
e
m
f
or cli
nical
docto
rs
a
nd contri
bu
te
t
o
the
contr
ol
of the
pandem
ic
.
REFERE
NCE
S
[1]
W
HO
,
"Coronavi
rus di
sea
se
201
9
(
COV
ID
-
19
)
Situa
t
ion
R
epor
t
-
94,
"
[Onlin
e]
,
Available:
http:/
/ww
w.w
ho.int/
,
[2]
C.
Huang
et
al.,
"Clini
c
al
fe
at
ur
e
s
of
pat
ie
n
ts
infect
ed
wi
th
2019
n
ovel
cor
on
avi
rus
in
W
uhan,
Chin
a,
"
T
he
Lancet,
vol.
395
,
no
.
102
23,
pp
.
497
-
506
,
2020.
[3]
N.
Lurie,
M.
Savil
le,
R
.
Ha
t
che
t
t,
and
J.
Halt
on,
"D
eve
l
oping
COV
ID
-
19
vacci
nes
at
pande
m
ic
spe
ed,
"
New
Eng
land
Jo
urnal
of
M
edicin
e,
2020
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
1
,
Febr
uar
y
2021
:
844
-
850
850
[4]
T.
Ai
,
e
t
al.
,
"Correl
a
ti
on
of
ch
est
CT
and
RT
-
PC
R
te
sting
in
cor
onavi
rus
dise
ase
2019
(
COV
ID
-
19
)
in
Chin
a
:
a
rep
or
t
of
1014
ca
ses,"
Radi
olog
y,
2020
.
[5]
Y.
Fang
et
al.
,
"S
ensit
ivi
t
y
of ch
est
CT
for
COV
I
D
-
19
:
compari
so
n
to
R
T
-
PC
R,
"
R
adiol
ogy
,
2020
.
[6]
N.
Zha
ng
et
a
l.
,
"Rec
ent
adv
ances
in
the
det
e
ct
i
on
of
respir
at
or
y
virus
inf
e
ct
ion
in
hum
ans,
"
Journal
of
Me
dic
a
l
Vi
rology
,
2020
.
[7]
M.
Chung
et
al.
,
"CT
imaging
fea
tur
es
of
201
9
novel
cor
ona
virus
(2019
-
nCoV),"
Radi
olog
y,
vol.
295
,
no
.
1
,
pp.
202
-
207
,
20
20.
[8]
V.
M.
Corm
an
e
t
al.
,
"D
e
te
c
ti
on
of
2019
novel
co
rona
virus
(2019
-
nCoV)
b
y
re
a
l
-
tim
e
RT
-
PC
R,
"
E
uros
urve
il
lance,
vol.
25
,
no
.
3
,
20
20.
[9]
X.
Xie,
Z.
Zho
ng,
W
.
Zha
o
,
C
.
Zhe
ng
,
F.
W
a
ng,
and
J.
Li
u
,
"Chest
CT
for
t
y
p
ical
2019
-
n
CoV
pneumonia:
rel
a
ti
onship
to
n
ega
t
ive
R
T
-
PC
R
te
st
ing,
"
Radi
ol
ogy,
2020
.
[10]
J.
George
,
S.
S
kar
ia,
and
V.
V
aru
n,
"U
si
ng
Y
OLO
base
d
de
e
p
learni
ng
n
et
w
ork
for
re
al
ti
m
e
de
te
c
ti
on
and
loc
a
li
z
at
ion
of
l
ung
nodu
le
s
from
low
dose
CT
sca
ns,"
Inte
rnat
ional
Society
fo
r
Optic
s
and
Phot
onic
s,
Me
d
ic
a
l
Imaging
2018:
Computer
-
Ai
ded
Diagnosis
,
vol
.
10575,
2018
.
[11]
Q.
Song,
L.
Zha
o,
X.
Luo
,
and
X.
Dou,
"U
sing
dee
p
le
arn
ing
for
cl
assific
a
ti
o
n
of
lung
nodule
s
on
computed
tomograph
y
ima
ges,
"
J
ournal
of
heal
th
care
en
gin
ee
ring,
vol
.
201
7,
no
.
1
,
pp
.
1
-
7
,
2017.
[12]
S.
Naqi,
M.
Sharif,
M.
Yasm
in,
and
S.
L.
Ferna
ndes,
"Lung
nodule
det
e
ction
using
pol
y
g
on
ap
proximati
on
and
h
y
brid
fe
at
ur
es
f
rom
CT
images,"
Curr
ent
Me
d
ical
Imaging,
vol
.
14,
no
.
1
,
pp
.
10
8
-
117,
2018
.
[13]
J.
C.
Souza,
J.
O.
B.
Diniz,
J.
L.
Ferre
ira,
G.
L.
F
.
da
Silva
,
A.
C.
Silva
,
and
A.
C.
de
Paiva
,
"A
n
au
tomati
c
m
et
hod
for
lung
segm
e
nta
ti
on
an
d
r
ec
o
nstruct
ion
in
ch
est
X
-
ra
y
using
dee
p
n
eur
a
l
ne
tworks,"
Computer
methods
and
progr
ams
in
biomedi
ci
n
e,
vol. 1
77,
pp
.
285
-
296
,
2019.
[14]
L.
Li
et
a
l.
,
"A
rti
ficia
l
in
te
l
li
g
en
ce
d
isti
nguishes
Covid
-
19
from
comm
unity
a
cq
uire
d
pn
eumonia
on
che
st
CT
,
"
Radi
ology
,
vo
l.
296,
2020
.
[15]
A.
Bhanda
r
y
,
e
t
al
.
,
"D
ee
p
-
l
ea
rni
ng
fra
m
ework
to
det
ect
lung
abn
orm
al
ity
–
A
study
with
ch
est
X
-
Ra
y
and
lung
CT
sca
n
images
,
"
P
att
ern
Re
cogn
it
i
on
Letters
,
vol
.
1
29,
pp
.
271
-
278
,
2020.
[16]
S.
W
ang
,
et
al
.
,
"A
dee
p
le
arn
in
g
al
gorit
hm
usin
g
CT
imag
es
to
scre
en
for
Coro
na
Virus
Disea
se
(
COV
ID
-
19
),
"
medR
xiv
,
2020.
[17]
X.
Xu
,
et
al
.
,
"D
ee
p
l
ea
rning
s
y
st
em
to
scre
en
co
r
onavi
rus
disea
se
2019
pneumonia,
"
arXiv
pr
eprint
arXiv
:
.
09334
,
2020.
[18]
A.
-
K.
Dhurgha
m
,
A.
-
Z.
Shakir,
P.
Nisree
n
,
and
J.
Sabah,
"M
ac
h
ine
L
ea
rn
i
ng
An
aly
s
is
of
Chest
CT
Scan
Im
ag
es
as
a
Com
ple
m
ent
a
r
y
D
igi
t
al
Te
st
of
Coronavi
rus
(
COV
ID
-
19
)
Pati
e
nts,"
medR
xi
v
,
2
020.
[19]
G.
Han,
X.
Li
u,
G.
Zhe
ng,
M.
W
ang,
and
S.
Hua
ng,
"A
utomati
c
r
ec
ogni
ti
on
of
3D
GG
O
CT
imaging
signs
through
the
fusion
of
h
y
b
rid
resa
m
pli
ng
a
nd
lay
e
r
-
wise
f
in
e
-
tuni
ng
CNN
s,"
Me
di
cal
&
biol
ogic
al
engi
n
ee
ri
ng
&
co
mputing,
vol.
56
,
no
.
12
,
p
p.
2201
-
2212
,
2
018.
[20]
O.
Russ
akovsk
y
,
e
t
al
.
,
"Im
ag
e
net
la
rge
sca
le
visual
r
ec
ogni
tion
challe
ng
e,
"
I
nte
rnational
jou
rnal
of
computer
vi
sion,
vol
.
115
,
no.
3
,
pp
.
211
-
2
52,
2015
.
[21]
G.
Zhong,
S.
Ya
n,
K.
Huang
,
Y.
Cai
,
and
J.
Dong,
"Reduc
ing
and
stret
ch
ing
de
ep
convol
uti
on
al
activa
t
ion
f
ea
tur
es
for
accurate
ima
ge
c
la
ss
ifi
c
ation,
"
Cognit
i
ve Computat
ion,
vol
.
10
,
no
.
1
,
pp
.
179
-
186,
2018
.
[22]
Y.
Jiao,
Y.
Zh
an
g,
Y.
W
ang,
B
.
W
ang,
J.
Jin,
an
d
X.
W
ang,
"A
novel
m
ult
ila
y
er
cor
relati
on
m
ax
imiza
ti
on
m
odel
for
improving
CCA
-
base
d
fre
q
u
ency
r
ec
ogn
it
i
on
in
SS
VEP
b
rai
n
-
computer
i
nte
rfa
ce,"
Int
ernati
onal
journal
of
neural
systems,
vol.
28
,
no
.
4
,
20
18.
[23]
D.
S.
Kerm
any
,
et
al.
,
"Ident
i
f
y
i
ng
m
edi
ca
l
dia
g
noses
and
tre
at
a
ble
disea
ses
b
y
i
m
age
-
base
d
deep
le
arn
ing
,
"
Cel
l
,
v
ol.
172
,
no
.
5
,
p
p.
1122
-
1131
,
2
018.
[24]
G.
L.
da
Silva
,
et
al.
,
"Lung
nodule
s
dia
gnosis
base
d
on
evol
utionar
y
convo
lut
i
onal
neur
a
l
netw
ork,
"
Mult
ime
dia
Tools
and
Appli
cat
ions,
vol
.
76
,
no.
18
,
pp
.
1903
9
-
19055,
2017
.
[25]
T.
Fawc
ett,
"A
n
int
roduc
t
ion to R
OC a
naly
sis,"
Pat
te
rn
rec
ogni
t
ion
l
et
t
ers,
vol
.
2
7,
no
.
8
,
pp
.
861
-
874,
2006
.
[26]
J.
P.
Coh
en, P.
Morrison,
and
L
.
Dao, "
COV
ID
-
19
image
data co
ll
e
ct
ion
,
"
arXi
v preprint
arXi
v:.11597,
202
0.
BIOGRA
PH
Y
OF A
UTHOR
Mohamme
d
Y.
Kami
l
obta
ine
d
his
M.
Sc
.
in
Optic
s
from
Mus
ta
nsiriy
a
Univer
sity
,
Coll
ege
of
Scie
nce
,
Phy
sics
depa
rtment,
Ira
q,
in
2005.
He
rec
ei
ved
his
Ph.D.
in
digi
ta
l
image
proc
essing
from
Mus
ta
nsiriy
a
Univer
sity
in
2011.
He
is
int
ere
sted
in
m
edi
ca
l
image
proc
essing,
computer
vision,
and
Artif
ic
ia
l
Inte
ll
ige
nce
.
He
is a
m
ember
of
the
IEE
E
Ira
q
sec
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.