Indonesian
J
our
nal
of
Electrical
Engineering
and
Computer
Science
V
ol.
36,
No.
3,
December
2024,
pp.
1731
1738
ISSN:
2502-4752,
DOI:
10.11591/ijeecs.v36.i3.pp1731-1738
r
1731
A
utomated
blood
cancer
detection
models
based
on
EfficientNet-B3
ar
chitectur
e
and
transfer
lear
ning
Nawaf
Alshdaifat
1
,
Hamza
Ab
u
Owida
2
,
Zaid
Mustafa
3
,
Ahmad
Ab
ur
omman
4
,
Suhaila
Ab
uo
waida
5
,
Abdullah
Ibrahim
6
,
W
afa
Alsharafat
7
1
F
aculty
of
Information
T
echnology
,
Applied
Science
Pri
v
ate
Uni
v
ersity
,
Amman,
Jordan
2
Department
of
Medical
Engineering,
F
aculty
of
Engineering,
Al-Ahliyya
Amman
Uni
v
ersity
,
Amman,
Jordan
3
Prince
Abdullah
bin
Ghazi
F
aculty
of
Information
and
Communication
T
echnology
,
Al-Balqa
Applied
Uni
v
ersity
,
Al-Salt,
Jordan
4
Department
of
Data
Science
and
AI,
F
aculty
of
Information
T
echnology
,
Zarqa
Uni
v
ersity
,
Zarqa,
Jordan
5
Department
of
Computer
Science,
F
aculty
of
Information
T
echnology
,
Al
al-Bayt
Uni
v
ersity
,
Mafraq,
Jordan
6
F
aculty
of
Engineering,
Zag
azig
Uni
v
ersity
,
Zag
azig,
Egypt
7
Department
of
Information
Systems,
F
aculty
of
Prince
Al-Hussein
bin
Abdullah
II
for
Information
T
echnology
,
Al
al-Bayt
Uni
v
ersity
,
Mafraq,
Jordan
Article
Inf
o
Article
history:
Recei
v
ed
Apr
25,
2024
Re
vised
Aug
12,
2024
Accepted
Aug
26,
2024
K
eyw
ords:
Blood
cancer
Deep
learning
Detection
Ef
ficientNet-B3
T
ransfer
learning
ABSTRA
CT
In
blood
smear
images,
there
are
dif
ficulties
in
diagnosing
blood
cancer
diseases
lik
e
leuk
em
ia
and
lymphoma
because
of
their
v
arious
forms
that
appear
in
the
human
body
.
In
this
paper
,
a
method
for
automatic
detection
of
blood
cancer
is
suggested
t
hat
uses
the
Ef
ficientNet-B3
architecture
along
with
transfer
learning
techniques
to
impro
v
e
accurac
y
and
ef
ficienc
y
.
W
e
first
fine-tuned
the
Ef
ficientNet-B3
model,
which
w
as
pre-trained
on
a
lar
ge
dataset
consisting
of
annotated
bl
ood
smear
images,
to
capture
pertinent
features
link
ed
with
blood
malignant
cells.
T
o
e
xpedite
the
training
process
and
adapt
the
model
to
our
task,
we
use
transfer
learning.
The
proposed
approach’
s
results
from
our
e
xperiments
sho
w
that
it
outperforms
traditional
deep
learning
models
and
state-
of-the-art
methods
in
blood
cancer
de
tection.
Additionally
,
with
high
precision
and
recall
rates,
this
model
also
detects
dif
ferent
types
of
blood
cancers
with
rob
ustness
in
its
performance
since
its
accurac
y
is
o
v
er
99%.
This
means
that
when
used
together
with
the
Ef
ficientNet-B3
architecture,
transfer
learning
can
help
the
de
v
eloped
methods
generalize
among
dif
ferent
types
of
blood
cancers
and
conditions.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Suhaila
Ab
uo
w
aida
Department
of
Computer
Science,
F
aculty
of
Information
T
echnology
,
Al
al-Bayt
Uni
v
ersity
Mafraq,
Jordan
Email:
suhilao
wida@gmail.com
1.
INTR
ODUCTION
Understanding
blood
cancer
is
essential
for
de
v
eloping
accurate
diagnostic
and
treatment
approaches
[1]–[3].
Hematologic
malignanc
y
,
commonly
referred
to
as
blood
cancer
,
arises
in
the
bone
marro
w
or
lym-
phatic
system,
impacting
the
generation
and
operation
of
blood
cells.
These
diseases,
including
leuk
emia,
lymphoma,
and
myeloma,
are
part
of
a
v
aried
group.
Each
disease
has
its
unique
characteristics
and
requires
specific
treatment
methods.
Blood
cancer
de
v
elops
through
genetic
abnormalities
that
cause
unre
gulated
cell
gro
wth
and
a
weak
ened
immune
response.
This
leads
to
a
complicated
interaction
between
the
cancerous
cells
and
the
surrounding
microen
vironment
[4].
J
ournal
homepage:
http://ijeecs.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
1732
r
ISSN:
2502-4752
Exploring
blood
cancer
is
critical
to
tailoring
ef
fecti
v
e
diagnostic
and
therapeutic
strate
gies.
Blood
cancer
,
also
kno
wn
as
hematologic
malignanc
y
,
is
a
disease
of
the
bone
marro
w
and
lymphatic
system
that
ultimately
af
fects
blood
cells’
production
and
functional
capabilities.
The
blood
cancer
-lik
e
coe
xisting
dis-
eases
encompass
leuk
emia,
lymphoma,
and
myeloma,
which
possess
unique
charac
teristics
and
therapeutic
re
gimens.
Genetic
mutations
t
h
a
t
induce
unrestricted
cellular
multiplication
and
impaired
immune
response
are
responsible
for
leading
to
blood
cancer
.
The
tumor
interactome
is
a
comple
x
netw
ork
that
e
v
olv
es
during
the
gro
wth
of
tumors.
So,
understanding
the
tumor
netw
ork
mak
es
it
easier
to
find
the
v
ery
important
signs
of
cancer
and
figure
out
ho
w
the
y
w
ork
at
the
molecular
le
v
el
[4].
Blood
cancer
is
important
for
the
de
v
elopment
of
ef
fecti
v
e
diagnostic
and
therapeutic
approaches
to
treatment.
Blood
cancer
,
or
hematologic
malignanc
y
,
arises
in
the
bone
marro
w
or
the
lymphatic
system
and
disrupts
the
producti
o
n
and
function
of
blood
cells.
This
group
of
diseases
is
v
ery
di
v
erse
and
includes
leuk
emia,
lymphoma,
and
myeloma,
among
others,
which
also
dif
fer
in
characteristics
and
approaches
to
treatment.
Mutations
that
cause
uncheck
ed
cell
di
vision
and
reduce
immune
acti
vity
are
what
lead
to
blood
cancer
.
In
the
conte
xt
of
these
processes,
comple
x
conditions
arise
for
the
interaction
of
the
tumor
cells
[4].
Deep
medical
imaging
learning
has
utterly
changed
the
g
ame
by
pro
viding
diagnost
ics,
treatment
planning,
and
patient
care
that
is
highly
accurate
and
ef
fecti
v
e.
W
ith
amazing
accurac
y
,
v
arious
deep
learn-
ing
architectures
such
as
con
v
olutional
neural
netw
orks
(CNNs)
ha
v
e
been
able
to
help
researchers
find
and
diagnose
dif
ferent
medical
conditions
[5]–[9].
F
or
e
xam
ple,
Ef
ficientNet-B3
architecture
has
sho
wn
good
per
-
formance
in
classifying
subtypes
of
blood
cancer
usi
ng
medical
imaging
data.
T
o
learn
comple
x
patterns
and
features
from
lar
ge
datasets
that
f
acilitate
accurate
predictions
with
high
sensiti
vity
and
specificity
,
deep
learn-
ing
algorithms
are
trained.
This
technological
b
r
eakthrough
can
greatly
enhance
early
detection
rates
as
well
as
treatment
outcomes
for
blood
cancers,
thereby
sa
ving
more
li
v
es
while
impro
ving
the
quality
of
life
among
patients
in
general.
Further
research
in
this
direction
holds
great
promise
for
the
future
of
medical
imaging
in
the
field
of
oncology
[10].
Deep
learning
approaches,
lik
e
the
Ef
ficientNet-B3
architecture,
sho
w
promise
for
analyzing
huge
amounts
of
data
from
genetic
profiling,
diagnostic
tests,
and
treatment
outcomes.
This
will
help
us
learn
more
about
ho
w
blood
cancer
is
dif
ferent
in
each
person
and
pro
vide
more
personalized
care.
By
unra
v
eling
the
intricacies
of
blood
cancer
at
a
molecular
le
v
el,
researchers
can
pa
v
e
the
w
ay
for
tar
geted
therapies
and
impro
v
ed
patient
outcom
es
[11].
Ho
we
v
er
,
it
is
important
to
note
that
the
de
v
elopment
of
such
systems
should
be
done
in
collaboration
with
medical
e
xperts,
who
can
pro
vide
v
aluable
insights
and
guidance
throughout
the
process
[12].
By
utilizing
deep
learning
CNN
algorithms,
blood
cancer
classification
can
be
greatly
impro
v
ed
in
terms
of
accurac
y
and
ef
ficienc
y
.
The
Ef
ficientNet-B3
architecture
in
blood
cancer
de-
tection
represents
a
groundbreaking
approach
that
harnesses
the
po
wer
of
deep
learning
to
enhance
diagnostic
accurac
y
and
ef
ficienc
y
.
Researchers
can
look
at
compli
cated
images
of
blood
cancer
with
a
le
v
el
of
accurac
y
that
has
ne
v
er
been
seen
before,
thanks
to
Ef
ficientNet-B3’
s
superior
feature
e
xtraction
capabilities.
This
al-
lo
ws
for
early
detection
and
personalized
treatment
plans.
Additionally
,
Ef
ficientNet-B3’
s
ability
to
gro
w
and
change
mak
es
it
perfect
for
handling
lar
ge
amounts
of
medical
imaging
data,
which
mak
es
it
easier
to
quickly
and
accurately
diagnose
dif
ferent
types
of
blood
cancer
.
By
inte
grating
this
cutting-edge
deep
learning
model
into
medical
practices,
healthcare
profes
sionals
can
use
sophis
ticated
technology
to
change
the
w
ay
the
y
detect
and
treat
blood
cancer
,
thereby
impro
ving
pa-
tient
outcomes
and
adv
ancing
oncology
.
The
y
ha
v
e
been
ef
fecti
v
e
in
man
y
medical
imaging
tasks
that
relate
to
blood
cancers
b
ut
are
not
limited
to
early
diagnosis
and
personalized
treatment
by
Ef
ficientNet-B3
archi-
tectures,
for
instance.
Th
i
s
study
therefore
intends
to
demonstrate
ho
w
this
particular
architecture
can
be
used
in
analyzi
ng
dif
ferent
kinds
of
blood
cancer
images
to
impro
v
e
their
identificat
ion
and
clas
sification
accurac
y
.
The
main
goal
of
this
research
is
to
in
v
estig
ate
the
Ef
ficientNet-B3
architecture
on
medical
images
for
bet-
ter
detection
abilities
for
v
arious
forms
of
leuk
emia
or
lymphoma.
In
other
w
ords,
what
w
ould
happen
if
we
let
deep
learning
loose
on
all
those
patient
records?
Such
an
e
xamination
w
ould
not
only
push
forw
ard
our
understanding
of
cancer
b
ut
also
help
us
sa
v
e
li
v
es
more
ef
fecti
v
ely
.
There
are
fi
v
e
parts
in
the
manuscript,
including
a
literature
re
vie
w
(section
2),
proposed
w
ork
(section
3),
results
with
rele
v
ant
comments
(section
4),
and
conclusion
(section
5).
2.
RELA
TED
W
ORK
The
man
y
characteristics
of
blood
cancers,
such
as
myeloma,
leuk
emia,
and
lymphoma,
mak
e
diag-
nosis
and
cate
gorization
v
ery
dif
ficult.
Con
v
entional
approaches
rely
significantly
on
laborious
and
subjecti
v
e
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
3,
December
2024:
1731–1738
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1733
histopathological
analysis,
genetic
profiling,
and
clinical
data
processing.
Deep
learning
has
recently
come
into
its
o
wn
as
an
ef
fecti
v
e
tool
for
blood
cancer
cate
gorization,
with
the
ability
to
enhance
precision,
ef
ficac
y
,
and
indi
vidualized
treatment
approaches.
While
in
v
estig
ating
se
v
eral
deep
learning
appli
cations,
this
section
highlights
k
e
y
approaches,
breakthroughs,
and
challenges
in
blood
cancer
cate
gorization.
The
researchers
[11],
[13],
[14]
used
CNNs
to
analyze
microscopic
images
of
blood
samples,
one
of
the
most
widely
recognized
deep
learning
applications
in
blood
cancer
cate
gorization.
CNNs’
e
xceptional
ability
to
capture
comple
x
characteris-
tics
and
patt
erns
within
these
images
enables
accurate
subtype
classification
of
blood
cancers
based
on
cellular
morphology
.
T
o
better
cate
gorize
blood
cancers,
researchers
ha
v
e
de
v
eloped
CNN
architectures
tailored
to
the
task.
T
o
train
algorithms
that
can
distinguish
between
normal
and
diseased
cells
based
on
minor
v
ariations,
the
y
used
massi
v
e
annotated
datasets.
These
models
pro
vide
a
f
aster
and
more
objecti
v
e
alternati
v
e
to
standard
manual
e
v
aluation
methods,
and
the
y
ha
v
e
demonstrated
promising
resul
ts
in
automated
diagnosis
and
subtype
identification
[15]–[17].
The
a
v
ailability
of
annotated
data
presents
a
problem
in
blood
cancer
cate
gorization,
particularly
for
rare
subtypes
or
less
pre
v
alent
presentations.
One
useful
strate
gy
for
dealing
with
the
data
limits
of
deep
learning
is
transfer
learning,
which
in
v
olv
es
taking
what
you’
v
e
learned
and
applying
it
to
a
ne
w
related
task.
W
e
can
fine-tune
pre-trained
models
to
respond
to
specific
classification
tasks
using
smaller
annotated
blood
cancer
datasets.
These
models
are
typically
trained
on
lar
ge-scale
image
datasets,
lik
e
Image-Net.
When
you
combine
features
from
dif
ferent
datasets
with
sparsely
annotated
data,
this
is
called
transfer
learning.
It
helps
deep
learning
models
classify
blood
cancer
better
across
dif
ferent
datasets
[18],
[19].
In
addition
to
using
indi
vidual
deep
learning
models,
certain
research
papers
ha
v
e
de
v
eloped
h
ybrid
deep
learning
frame
w
orks
for
leuk
emia
cancer
diagnosis.
Y
osinski
et
al.
[18]
introduced
a
h
ybrid
approach
that
utilizes
ResNet50,
V
GG16,
and
V
GG19
which
are
adv
anced
CNNs,
to
automate
cell
identification.
The
y
compare
their
proposed
method
with
K-nearest
neighbors
(KNN),
logistic
re
gression
(LR),
support
v
ector
ma-
chine
(SVM),
and
decision
tree
(DT).
The
proposed
technique
achie
v
es
88.50%
accurac
y
in
cell
identification.
Y
osinski
et
al.
[18]
created
a
h
ybrid
model
using
tw
o
CNN
architectures
to
enhance
accurac
y
in
deep
learning.
T
esting
the
suggested
procedure
on
636
blood
samples
with
health
y
and
acute
lymphoblastic
leuk
emia
(ALL)
cells
yields
an
accurac
y
of
89.70%.
T
ogether
,
the
vision
transformer
and
CNN-based
model
e
xtract
man
y
cell
features
in
di
v
erse
w
ays,
impro
ving
classification
results.
The
y
impro
v
ed
an
une
v
en
dataset
using
enhancement-
random
sampling
(DERS).
The
algorithm
achie
v
es
99.03%
accurac
y
,
pro
ving
its
ef
fecti
v
eness
as
a
computer
aided
design
(CAD)
system
for
ALL.
Kasani
et
al.
[20]
de
v
eloped
a
h
ybrid
method
that
combines
V
GG16
and
MobileNet
to
e
xtract
the
profound
characteristics,
which
are
then
used
for
the
cate
gorization
of
leuk
emic
B-lymphoblast
[20].
Their
proposed
methodology
incorporates
multiple
data
augmentation
techniques
and
achie
v
ed
an
accurac
y
of
96.17%,
a
sensiti
vity
of
95.17%,
and
a
specificity
of
98.58%.
In
addition,
the
re-
searcher
in
[21]
combined
the
N
ASNetLar
ge
and
V
GG19
d
e
ep
machine
learning
models
to
classify
leuk
emic
B-lymphoblast
cells
and
normal
B-lymphoid
precursor
cells,
achi
e
ving
a
detection
accurac
y
of
96.58%.
Their
suggested
model
accurately
detects
acute
lymphoblastic
leuk
emia,
and
the
y
demonstrate
that
ensemble
learning
is
significantly
superior
to
a
single
model.
T
o
a
d
dr
ess
the
problem
of
classifying
blood
cancer
,
our
suggested
deep
learning
model
mak
es
use
of
the
Ef
ficientNet-B3
model
and
transfer
learning.
By
using
this
method,
the
traditional
deep
learning
model’
s
accurac
y
is
impro
v
ed,
and
the
images
are
classified
as
either
positi
v
e
or
ne
g
ati
v
e
blood
samples.
3.
METHOD
Ef
ficientNet-B3
architecture,
an
adv
anced
neural
netw
ork
model
[22],
has
sho
wn
promising
results
in
v
arious
deep
learning
applications,
including
medical
image
analysis
for
detecting
blood
cancer
.
This
architec-
ture,
characterized
by
its
increased
depth
and
breadth
in
comparison
to
pre
vious
models,
achie
v
es
a
harmonious
equilibrium
between
the
comple
xity
of
the
model
and
its
computational
ef
ficienc
y
.
The
Ef
ficientNet-B3
model
uses
ef
ficient
b
uilding
blocks,
lik
e
mobile
in
v
erted
bottleneck
con
v
olution
and
squeeze-and-e
xcitation
blocks,
to
impro
v
e
performance
while
k
eeping
the
cost
of
computing
lo
w
.
Furthermore,
i
t
emplo
ys
com
poun
d
scaling
to
ef
fecti
v
ely
increase
the
netw
ork’
s
depth,
width,
and
resolution
at
the
sam
e
time,
hence
impro
ving
its
ability
to
repre
sent
information
without
a
substantial
increase
in
resources.
The
Ef
ficientNet-B3
design
e
xhibits
ro-
b
ustness
and
ef
fecti
v
eness,
rendering
it
a
f
a
v
orable
option
for
the
precis
e
classification
of
blood
cancer
jobs.
Furthermore,
it
e
xhibits
superior
performance
when
used
with
e
xtensi
v
e
collections
of
images
[22].
The
in
v
es-
tig
ation
and
application
of
this
structure
emphasize
its
capacity
for
enhancing
deep
learning
techniques
for
the
interpretation
of
medical
images,
specifically
in
the
field
of
oncology
research.
A
utomated
blood
cancer
detection
models
based
on
Ef
ficientNet-B3
...
(Nawaf
Alshdaifat)
Evaluation Warning : The document was created with Spire.PDF for Python.
1734
r
ISSN:
2502-4752
3.1.
The
dataset
W
e
obtained
the
dataset
for
this
w
ork
from
supplementary
data
sources,
specifically
Kaggle
[23].
The
collection
comprises
3,262
authentic
photographs
of
blood
v
essel
stains.
Out
of
the
89
subjects
whose
photographs
were
included,
64
were
thought
to
ha
v
e
ALL,
while
the
remaining
25
were
assessed
to
be
normal.
W
e
initially
di
vided
the
dataset
into
tw
o
identical
cate
gories:
benign
and
malignant
,
before
or
g
anizing
it
into
four
primary
groups
with
three
subtypes
of
cancer
(benign,
early
pre-B,
pre-B,
and
pro-B).
The
photographs
were
captured
with
a
Zeiss
camera
mounted
on
a
microscope
at
a
magnificat
ion
le
v
el
of
100
and
sa
v
ed
in
JPG
format.
An
e
xpert
conducted
a
comprehensi
v
e
and
conclusi
v
e
in
v
estig
ation
using
flo
w
c
ytometry
to
determine
the
types
and
subtypes
of
these
images.
3.2.
T
ransfer
lear
ning
T
ransfer
learning,
or
domain
adaptation,
is
a
c
o
nc
eptual
frame
w
ork
that
le
v
erages
kno
wledge
from
one
area
to
address
related
tasks
[5],
[6],
[24].
In
thi
s
re
g
ard,
the
study
authors
utilized
their
parameters
on
the
ImageNet
dataset,
which
holds
great
significance
in
deep
learning
applications.
T
ransfer
learning
is
a
type
of
semi-supervised
learning
method
that
reduces
the
need
for
labeled
data,
especially
in
situations
where
there
is
a
limited
amount
of
data
a
v
ailable
for
collection
[25].
An
outstanding
adv
antage
of
this
strate
gy
is
its
remarkable
capacity
to
ef
fecti
v
ely
generalize
across
v
arious
jobs.
The
study
ef
fecti
v
ely
emplo
yed
transfer
learning
and
fine-tuning
strate
gies
to
de
v
elop
Ef
ficientNet-B3,
sho
wcasing
its
ef
ficac
y
in
tackling
similar
jobs.
3.3.
Pr
oposed
EfficientNet-B3
model
Once
t
he
transfer
learning
model
is
utilized.
The
primary
focus
of
this
study
w
as
on
the
Ef
ficientNet
f
amily
of
models,
with
particular
emphasis
on
Ef
ficientNet-B3,
as
sho
wn
in
Figure
1.
The
goal
of
Ef
ficientNet
models
is
to
find
a
good
balance
between
accurac
y
and
ef
ficienc
y
in
CNNs
by
using
a
unique
method
called
compound
scaling.
This
method
equally
adjusts
the
width,
depth,
and
quality
characteristics
by
applying
pre-
determined
f
actors.
Ef
ficientNet-B3,
as
a
member
of
this
group,
e
x
emplifies
this
equilibrium
by
emplo
ying
a
customized
architecture
that
pro
vides
significant
generalizability
.
Figure
1.
Architecture
of
the
proposed
w
ork
[22]
F
or
our
model
creation,
we
utilized
the
K
eras
libra
ry
to
creat
e
an
inst
ance
o
f
Ef
fici
entNet-B3
as
the
foundational
model,
which
w
as
already
trained
on
ImageNet
and
had
pre-e
xisting
weights.
T
o
optimize
the
use
of
the
already
trained
features,
we
immobilized
the
initial
half
of
the
layers.
Aft
erw
ard,
we
incorporated
more
customized
layers
into
the
e
xisting
model.
A
global
a
v
erage
pooling
layer
w
as
used
to
mak
e
the
space
smaller
,
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
3,
December
2024:
1731–1738
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1735
and
then
tw
o
dense
layers
(128
and
64
pix
els)
used
rectifie
d
linear
unit
(ReLU)
acti
v
ation
functions.
Dropout
layers
with
a
dropout
rate
of
0.5
were
selecti
v
ely
included
to
impl
ement
re
gularization
and
mitig
ate
the
risk
of
o
v
erfitting
[26],
[27].
The
last
layer
,
a
dense
layer
with
softmax
acti
v
ation,
w
as
set
up
for
our
multi-class
classification
problem
with
4
classes.
This
model,
characterized
by
its
unique
design,
has
pro
v
en
to
be
highly
ef
fecti
v
e
in
accurately
capturing
comple
x
patterns,
particularly
in
photographs
of
superior
quality
.
Using
the
compound
scaling
method,
Ef
ficientNet-B3
w
as
able
to
find
a
good
balance
between
the
model’
s
parameters,
accurac
y
,
and
ho
w
quickly
it
could
do
its
w
ork.
4.
RESUL
TS
AND
DISCUSSION
This
section
analyses
the
results
from
e
xperiments
using
deep
learning
models.
This
section
gi
v
es
an
e
v
aluation
analysis
using
a
state-of-the-art
model.
The
suggested
model
w
as
e
v
aluated
using
v
arious
per
-
formance
matrices,
including
accurac
y
(A),
precision
(P),
recall
(R),
and
F1-score
with
state-of-the-art
model
[28]–[30].
T
able
1
depicts
that
this
highly
adv
anced
neural
netw
ork
architecture,
the
Ef
ficientNet-B3
archi-
tecture,
yields
good
performance
in
medical
image
analysis
tar
geted
to
identify
blood
cancer
and
can
be
used
for
a
wide
range
of
deep
learning
applications.
It
is
deeper
and
wider
compared
to
the
model
on
which
it
is
b
uilt,
while
still
balancing
between
the
comple
xity
of
the
model
and
the
computational
ef
ficienc
y
.
The
Ef
ficientNet-B3
model
uses
ef
ficient
b
uilding
blocks-e.g.,
mobile
in
v
erted
bottleneck
con
v
olution
blocks
and
squeeze-and-e
xcitation
blocks-to
enhance
performance
while
k
eeping
the
cost
of
computing
v
ery
lo
w
.
Com-
pound
scaling
is
also
adopted
to
scale
the
depth,
width,
and
resolut
ion
of
a
netw
ork
all
at
the
same
time
in
a
resource-ef
ficient
manner
.
T
able
1.
The
o
v
erall
performance
of
dif
ferent
Pre-trained
CNN
models
CNN
model
Precision
recall
F1-score
V
GG19
96.64
96.87
96.75
ResNet50
98.28
98.32
98.30
Inception
V3
77.82
75.68
75.76
Inception
+
Xception
87.71
84.27
85.06
Proposed
research
99.00
99.00
99.00
Figure
2
illustrates
the
projected
netw
ork’
s
accurac
y
of
99%.
Figure
2(a)
displays
an
accurac
y
curv
e
that
sho
ws
the
relationship
between
the
number
of
epochs
(horizontal
axis)
and
the
range
of
accurac
y
(v
ertical
axis).
Augmenting
the
number
of
epochs
results
in
enhanced
detection
accurac
y
.
According
to
Figure
2(b),
increasing
the
number
of
epochs
decreases
the
loss
of
the
detection
netw
ork.
(a)
(b)
Figure
2.
The
plots
diagram
of
proposed
architecture
(a)
training
and
v
alidation
accurac
y
and
(b)
training
and
v
alidation
loss
The
Ef
ficientNet-B3
architecture’
s
rob
ustness
and
ef
ficienc
y
mak
e
it
a
promising
candidate
for
ac-
curate
blood
cancer
classification
tasks,
of
fering
impro
v
ed
performance
on
comple
x
image
datasets.
The
e
x-
A
utomated
blood
cancer
detection
models
based
on
Ef
ficientNet-B3
...
(Nawaf
Alshdaifat)
Evaluation Warning : The document was created with Spire.PDF for Python.
1736
r
ISSN:
2502-4752
ploration
and
utilization
of
this
architecture
highlight
its
potential
for
adv
ancing
deep
learning
methodologies
for
medical
image
analysis,
particularly
in
oncology
research.
In
addition,
transfer
learning
has
had
transfor
-
mati
v
e
ef
fects
on
blood
detection
deep
learning
models
by
impro
ving
accurac
y
,
reducing
data
requirements,
enhancing
generalization
and
domain
adaptation,
and
promoti
n
g
accessibility
for
medical
professionals.
It
has
f
acilitated
adv
ancements
in
research
and
clinical
applications,
pa
vi
ng
the
w
ay
for
more
ef
fecti
v
e
and
ef
ficient
blood
detection
in
healthcare
settings.
5.
CONCLUSION
As
a
result,
the
conclusion
sho
ws
tha
t
the
Ef
ficientNet-B3
architecture
w
orks
well
for
deep
learning
models
that
use
transfer
learning
to
find
and
classify
blood
cancer
.
The
high
accurac
y
and
ef
ficienc
y
of
the
model
in
discriminating
between
dif
ferent
types
of
blood
cancer
cells
indicate
that
this
method
may
impro
v
e
diagnostic
procedures
in
actual
clinical
settings.
Besides,
it
can
be
ar
gued
that
the
model’
s
ability
to
deal
with
dif
ferent
types
of
datasets
could
be
e
xtremely
comple
x
and
di
v
er
gent,
which
re
v
eals
its
rob
ustness
and
scalabil-
ity
.
T
o
further
our
kno
wledge,
there
must
be
additional
w
ork
to
e
xplore
the
full
potential
and
limitations
of
this
architecture
as
a
means
for
detecting
blood
cancer
,
as
well
as
refining
it
for
real-time
applications.
Deep
learn-
ing
inte
grates
with
inno
v
ati
v
e
architectures,
which
allo
w
us
to
push
forw
ard
medical
imaging
and
diagnosis
to
realize
better
patient
outcomes.
T
o
conclude,
deep
learning
combined
with
cutting-edge
architectural
designs
will
re
v
olutionize
medical
imaging,
thereby
enhancing
patient
results
and
healthcare
deli
v
ery
.
REFERENCES
[1]
H.
M.
Rai,
“Cancer
detection
and
se
gmentation
using
machine
learning
and
deep
learning
technique
s:
a
re
vie
w
,
”
Multimedia
T
ools
and
Applications
,
v
ol.
83,
no.
9,
pp.
27001–27035,
Aug.
2023,
doi:
10.1007/s11042-023-16520-5.
[2]
A.
J.
Aboug
arair
et
al.
,
“Blood
cells
cancer
detection
based
on
deep
learning,
”
J
ournal
of
Advances
in
Artificial
Intellig
ence
,
v
ol.
2,
no.
1,
pp.
108–121,
2024,
doi:
10.18178/jaai.2024.2.1.108-121.
[3]
F
.
M.
T
alaat
and
S.
A.
Gamel,
“Machine
learning
in
detection
and
classification
of
leuk
emia
using
C-NMC
Leuk
emia,
”
Multimedia
T
ools
and
Applications
,
v
ol.
83,
no.
3,
pp.
8063–8076,
Jan.
2024,
doi:
10.1007/s11042-023-15923-8.
[4]
D.
Gao
and
G.
Ge,
“Exploring
the
underlying
biology
of
cancer
and
potential
therapeutic
strate
gies:
a
special
issue
fo-
cused
on
mechanism-based
studies,
”
Acta
Bioc
himica
et
Biophysica
Sinica
,
v
ol.
55,
no.
6,
pp.
891–892,
Jun.
2023,
doi:
10.3724/abbs.2023114.
[5]
H.
A.
Owida,
B.
A.
Moh’
d,
N.
T
urab,
J.
Al-Nab
ulsi,
and
S.
Ab
uo
w
aida,
“The
e
v
olution
and
reliability
of
machine
learning
techniques
for
oncology
,
”
International
J
ournal
of
Online
and
Biomedical
Engineering
(iJOE)
,
v
ol.
19,
no.
08,
pp.
110–129,
Jun.
2023,
doi:
10.3991/ijoe.v19i08.39433.
[6]
N.
Alshdaif
at,
M.
A.
Osman,
and
A.
Z.
T
alib,
“
An
impro
v
ed
multi-object
instance
se
gmentation
based
on
deep
learning,
”
K
uwait
J
ournal
of
Science
,
v
ol.
49,
no.
2,
pp.
1–15,
Dec.
2022,
doi:
10.48129/kjs.10879.
[7]
H.
Ab
u
Owida,
“Recent
biomimetic
approaches
for
articular
cartilage
tissue
engineering
and
their
clinical
applications:
narrati
v
e
re
vie
w
of
the
literature,
”
Advances
in
Orthopedics
,
v
ol.
2022,
pp.
1–9,
Apr
.
2022,
doi:
10.1155/2022/8670174.
[8]
A
.
A.
Nafea,
M.
S.
Ibrahim,
M.
M.
Shw
aysh,
K.
Abdul-Kadhim,
H.
R.
Almamoori,
and
M.
M.
AL-Ani,
“
A
deep
learning
algorithm
for
lung
cancer
detection
using
Ef
ficientNet-B3,
”
W
asit
J
ournal
of
Computer
and
Mathematics
Science
,
v
ol.
2,
no.
4,
pp.
68–76,
Dec.
2023,
doi:
10.31185/wjcms.209.
[9]
S.
Ab
uo
w
aida
and
H.
Y
.
Chan,
“Impro
v
ed
deep
learning
architecture
for
depth
estimation,
”
J
or
danian
J
ournal
of
Computer
s
and
Information
T
ec
hnolo
gy
,
v
ol.
06,
no.
04,
pp.
434–445,
2020,
doi:
10.5455/jjcit.71-1593368945.
[10]
N.
Ruza,
S.
Hussan,
S.
Mohamed,
and
M.
Arzmi,
“Early
detection
of
breast
cancer
in
mammograms
using
the
lightweight
modifi-
cation
of
ef
ficientNet
B3,
”
M
´
etodos
num
´
ericos
par
a
c
´
alculo
y
dise
˜
no
en
ing
enier
´
ıa:
Re
vista
internacional
,
v
ol.
39,
no.
3,
pp.
1–11,
2023,
doi:
10.23967/j.rimni.2023.08.002.
[11]
Y
.
Zhu
et
al.
,
“
Application
of
con
v
olutional
neural
netw
ork
in
the
diagnosis
of
the
in
v
asion
depth
of
g
astric
cancer
based
on
con
v
en-
tional
endoscop
y
,
”
Gastr
ointestinal
Endoscopy
,
v
ol.
89,
no.
4,
pp.
806–815,
Apr
.
2019,
doi:
10.1016/j.gie.2018.11.011.
[12]
A.
Ab
ug
abah,
A.
Al
Smadi,
and
A.
Ab
uqabbeh,
“Data
mining
in
health
care
sector
,
”
in
Pr
oceedings
of
the
2019
2nd
International
Confer
ence
on
Computational
Intellig
ence
and
Intellig
ent
Systems
,
No
v
.
2019,
pp.
63–68,
doi:
10.1145/3372422.3372451.
[13]
Q.
D.
V
u
et
al.
,
“Methods
for
se
gmentation
and
classification
of
digital
microscop
y
tissue
images,
”
F
r
ontier
s
in
Bioengineering
and
Biotec
hnolo
gy
,
v
ol.
7,
no.
APR,
p.
433738,
2019,
doi:
10.3389/fbioe.2019.00053.
[14]
S.
H.
Kassani,
P
.
H.
Kassani,
M.
J.
W
esolo
wski,
K.
A.
Schneider
,
and
R.
Deters,
“
A
h
ybrid
deep
learning
architecture
for
leuk
emic
B-lymphoblast
classification,
”
in
2019
International
Confer
ence
on
Information
and
Communication
T
ec
hnolo
gy
Con
ver
g
ence
(ICTC)
,
Oct.
2019,
pp.
271–276,
doi:
10.1109/ICTC46691.2019.8939959.
[15]
A.
Este
v
a
et
al.
,
“
A
guide
to
deep
learning
in
healthcare,
”
Natur
e
Medicine
,
v
ol.
25,
no.
1,
pp.
24–29,
Jan.
2019,
doi:
10.1038/s41591-
018-0316-z.
[16]
E.
J.
Hw
ang
et
al.
,
“De
v
elopment
and
v
alidation
of
a
deep
learning–based
automated
detection
algorithm
for
major
thoracic
diseases
on
chest
radiographs,
”
J
AMA
Network
Open
,
v
ol.
2,
no.
3,
p.
e191095,
Mar
.
2019,
doi:
10.1001/jamanetw
ork
open.2019.1095.
[17]
H.-C.
Shin
et
al.
,
“Deep
con
v
olutional
neural
netw
orks
for
computer
-aided
detection:
CNN
architectures,
dataset
charac-
teristics
and
transfer
learning,
”
IEEE
T
r
ansactions
on
Medical
Ima
ging
,
v
ol.
35,
no.
5,
pp.
1285–1298,
May
2016,
doi:
10.1109/TMI.2016.2528162.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
3,
December
2024:
1731–1738
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1737
[18]
J.
Y
osinski
,
J.
Clune,
Y
.
Bengio,
and
H.
Lipson,
“Ho
w
transferable
are
features
in
deep
neural
netw
orks?,
”
Advances
in
Neur
al
Information
Pr
ocessing
Systems
,
v
ol.
4,
no.
January
,
pp.
3320–3328,
2014.
[19]
Z.
Xiao,
Y
.
Jiang,
T
.
Sun,
Y
.
W
u,
and
Y
.
T
ang,
“Refining
po
wer
con
v
erter
loss
e
v
aluation:
a
transfer
learning
approach,
”
IEEE
T
r
ansactions
on
P
ower
Electr
onics
,
v
ol.
39,
no.
4,
pp.
4313–4324,
Apr
.
2024,
doi:
10.1109/TPEL.2023.3349178.
[20]
P
.
H.
Kasani,
S.-W
.
P
ark,
a
nd
J.-W
.
Jang,
“
An
aggre
g
ated-based
deep
learning
method
for
leuk
emic
B-lymphoblast
classification,
”
Dia
gnostics
,
v
ol.
10,
no.
12,
p.
1064,
Dec.
2020,
doi:
10.3390/diagnostics10121064.
[21]
B.
Zoph,
V
.
V
asude
v
an,
J.
Shlens,
and
Q.
V
.
Le,
“Learning
transferable
architectures
for
scalable
image
recognition,
”
in
Pr
o-
ceedings
of
the
IEEE
Computer
Society
Confer
ence
on
Computer
V
ision
and
P
attern
Reco
gnition
,
2018,
pp.
8697–8710,
doi:
10.1109/CVPR.2018.00907.
[22]
M.
T
an
and
Q.
V
.
Le,
“Ef
ficientNet:
rethinking
model
scaling
for
con
v
olutional
neural
netw
orks,
”
in
36th
International
Confer
ence
on
Mac
hine
Learning
,
ICML
2019
,
2019,
v
ol.
2019-June,
pp.
10691–10700.
[23]
M.
Mohseni,
H.
Uludag,
and
J.
M.
B
randwein,
“
Adv
ances
in
biology
of
acute
lymphoblastic
leuk
emia
(ALL)
a
nd
therapeutic
implications.,
”
American
journal
of
blood
r
esear
c
h
,
v
ol.
8,
no.
4,
pp.
29–56,
2018.
[24]
S.
F
.
A.
Ab
uo
w
aida,
H.
Y
.
Chan,
N.
F
.
F
.
Alshdaif
at,
and
L.
Ab
ualig
ah,
“
A
no
v
el
i
nstance
se
gmentation
algorithm
based
on
impro
v
ed
deep
learning
algorithm
for
multi-object
images,
”
J
or
danian
J
ournal
of
Computer
s
and
Information
T
ec
hnolo
gy
,
v
ol.
7,
no.
1,
pp.
74–88,
2021,
doi:
10.5455/jjcit.71-1603701313.
[25]
H.
A.
Owida,
H.
S.
Migdadi,
O.
S.
M.
Hemied,
N.
F
.
F
.
Alshdaif
at,
S.
F
.
A.
Ab
uo
w
aida,
and
R.
S.
Alkha
w
aldeh,
“Deep
learning
algorithms
to
impro
v
e
CO
VID-19
classification
based
on
CT
images,
”
Bulletin
of
Electrical
Engineering
and
Informatics
,
v
ol.
11,
no.
5,
pp.
2876–2885,
Oct.
2022,
doi:
10.11591/eei.v11i5.3802.
[26]
A.
AL
Smadi,
A.
Ab
ug
abah,
A.
Mehmood,
and
S.
Y
ang,
“Brain
image
fusion
approach
based
on
side
windo
w
filtering,
”
Pr
ocedia
Computer
Science
,
v
ol.
198,
pp.
295–300,
2022,
doi:
10.1016/j.procs.2021.12.243.
[27]
O.
T
ara
wneh
et
al.
,
“The
Ef
fect
of
pre-processing
on
a
con
v
olutional
neural
netw
ork
model
for
dorsal
hand
v
ein
recognition,
”
International
J
ournal
of
Advanced
Computer
Science
and
Applications
,
v
ol.
15,
no.
3,
2024,
doi:
10.14569/IJ
A
CSA.2024.01503126.
[28]
O.
Alidma
t
et
al.
,
“Simulation
of
cro
wd
e
v
acuation
in
asymmetrical
e
xit
layout
based
on
impro
v
ed
dynamic
parameters
model,
”
IEEE
Access
,
pp.
1–1,
2024,
doi:
10.1109/A
CCESS.2024.3420068.
[29]
G.
Samara
et
al.
,
“Intelligent
edge:
the
intersection
of
artificial
intelligence
and
digital
communicati
on-a
surv
e
y
,
”
Studies
in
Systems,
Decision
and
Contr
ol
,
v
ol.
528,
pp.
1011–1022,
2024,
doi:
10.1007/978-3-031-56586-1
73.
[30]
D.
I.
A.
B.
A.
Aziz,
M.
Y
usof
f,
N.
Ibrahim,
and
R.
Alazaidah,
“P
addy
diseases
multi-class
classification
using
CNN
v
ariants,
”
in
2023
24th
International
Ar
ab
Confer
ence
on
Information
T
ec
hnolo
gy
(A
CIT)
,
Dec.
2023,
pp.
1–8,
doi:
10.1109/A
CIT58888.2023.10453746.
BIOGRAPHIES
OF
A
UTHORS
Nawaf
Alshdaifat
recei
v
ed
the
B.Sc.
in
Computer
Science
from
AL
al-Bayt
Uni
v
ersity
,
Jordan,
M.Sc.
de
gree
in
Computer
Science
from
Uni
v
ersity
of
Jordan,
Jordan,
in
2002
and
2011,
respecti
v
ely
and
the
Ph.D.
de
gree
in
Computer
Sciences
from
Uni
v
ersiti
Sains
Malaysia,
Malaysia,
in
2023.
His
research
interests
include
deep
lear
ning
and
tracking
objects.
He
can
be
contacted
at
email:
n
alshdaif
at@asu.edu.jo.
Hamza
Ab
u
Owida
has
completed
his
Ph.D.
from
K
eele
Uni
v
ersity
,
UK.
He
w
as
a
post-doctoral
Research
Associate:
De
v
eloping
x
eno-free
nanofibrous
scaf
fold
methodology
for
hu-
man
pluripotent
stem
cell
e
xpansion,
dif
ferentiation,
and
implantation
to
w
ards
the
therapeutic
prod-
uct,
K
eele
Uni
v
ersity
,
Institute
for
Science
a
nd
T
echnology
in
Medicine
(ISTM),
Staf
fordshire/UK.
He
is
an
assistant
professor
in
the
me
dical
engineering
department
at
Al-Ahliyya
Amman
Uni
v
er
-
sity
.
He
has
published
more
than
10
papers
in
reputed
journals.
He
can
be
contacted
at
email:
hamza.o
wida@gmail.com.
Zaid
Mustafa
recei
v
ed
his
Ph.D.
De
gree
in
Computer
Information
Systems
and
Computer
Graphics,
from
Santiago
de
Compostela
Uni
v
ersity
-Spain
(2014-2019).
Pre
viously
M.Sc,
in
Com-
puter
Information
Systems
(2012)
from
Middle
East
Uni
v
ersity
and
a
B.Sc
in
Sofw
are
Engineering
(2007)
from
Al-Balqa
Applied
Uni
v
ersity
,
Al-Salt,
Hashemite
Kingdom
of
Jordan.
Currently
,
he
serv
es
as
an
Assistant
Professor
at
Al-Balqa
Applied
Uni
v
ersity
,
Prince
Abdullah
bin
Ghazi
F
ac-
ulty
of
Information
and
Communication
T
echnology
,
Jordan
(2019-present),
and
is
also
Head
of
the
Computer
Information
Systems
depart
ment
(2022–present).
His
re
search
interests
include
machine
learning,
human-computer
interaction,
de
v
eloping
geographic
information
system
(GIS)
tools,
virtual
reality
,
augmented
reality
,
internet
of
things
(IoT),
and
smart
cities.
He
can
be
contacted
at
email:
z
lami@bau.edu.jo.
A
utomated
blood
cancer
detection
models
based
on
Ef
ficientNet-B3
...
(Nawaf
Alshdaifat)
Evaluation Warning : The document was created with Spire.PDF for Python.
1738
r
ISSN:
2502-4752
Ahmad
Ab
ur
omman
is
currently
an
Assistant
Professor
of
Artificial
Intelligence
in
the
Department
of
Data
Science
and
Artificial
Intelligence
in
the
Zaraqa
Uni
v
ersity
.
He
recei
v
ed
a
B.Sc.
de
gree
in
information
technology
from
Al-Balqa
Applied
Uni
v
ersity
(B
A
U),
Jordan,
in
2005,
and
an
M.Sc.
de
gree
in
information
technology
(e-Services)
from
Grif
fith
Uni
v
ersity
,
Australia,
in
2009.
He
got
a
Ph.D.
de
gree
in
Intelligence
Systems
at
the
Research
Center
in
Intelligent
T
echnologies,
Uni
v
ersity
of
Santiago
de
Compostela
(CiTIUS).
His
research
interests
include
machine
learning,
deep
learning.
He
can
be
contacted
at
email:
ahmad.ab
uroman@gmail.com.
Suhaila
Ab
uo
waida
recei
v
ed
the
B.Sc.
de
grees
in
Computer
Information
System,
M.Sc.
de
grees
in
Computer
Science
from
AL
alBayt
Uni
v
ersity
,
Jordan,
in
2012
and
2015,
respecti
v
ely
,
and
the
Ph.D.
de
gree
in
Computer
Science
from
Uni
v
ersiti
Sains
Malaysia,
Malaysia,
in
2023.
Dr
.
Ab
uo
w
aida
is
currently
w
orking
as
an
Assist
ant
Professor
with
the
Compute
r
Science
Department,
Zarqa
Uni
v
ersity
,
Jordan.
Her
rese
arch
interests
include
deep
learning,
depth
estimation,
point
cloud,
and
computer
vision.
She
can
be
contacted
at
email:
suhilao
wida@gmail.com.
Abdullah
Ibrahim
recei
v
ed
the
B.Sc.
de
grees
in
Medical
Informatics
from
Zag
azig
Uni-
v
ersity
,
Egypt,
in
2023.
Eng.
W
agih
is
currently
w
orking
as
an
AI
Engineer
at
Inno
v
ateX
Solutions
compan
y
,
Giza,
Egypt.
Also,
he
is
an
AI
trainer
at
AMIT
Learning
Compan
y
in
Cairo,
Egypt.
He
can
be
contacted
at
email:
abdulla
w
agih1@gmail.com.
W
afa
Alsharafat
recei
v
ed
the
B.S.
de
gree
in
Computer
Science
from
Hashemite
Uni
v
er
-
sity
,
Zarqa,
Jordan,
in
2001,
and
the
M.S.
and
Ph.D.
de
grees
in
computer
information
systems
from
Arab
Academy
,
Amman,
Jordan,
in
2004
and
2009,
respecti
v
ely
.
She
i
s
currently
an
Associate
Pro-
fessor
at
the
Information
Systems
Department,
Al
al-Bayt
Uni
v
ersity
,
Mafraq,
Jordan.
Her
research
interests
include
artificial
intelligence,
wireless
sensor
netw
orks,
genetic
algorithms,
optimization
algorithms,
and
netw
ork
security
.
She
can
be
contacted
at
email:
w
af
a@aab
u.edu.jo.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
3,
December
2024:
1731–1738
Evaluation Warning : The document was created with Spire.PDF for Python.