Informatyka Automatyka Pomiary w Gospodarce i Ochronie Środowiska,
Год журнала:
2023,
Номер
13(4), С. 27 - 33
Опубликована: Дек. 20, 2023
Transfer
Learning
(TL)
is
a
popular
deep
learning
technique
used
in
medical
image
analysis,
especially
when
data
limited.
It
leverages
pre-trained
knowledge
from
State-Of-The-Art
(SOTA)
models
and
applies
it
to
specific
applications
through
Fine-Tuning
(FT).
However,
fine-tuning
large
can
be
time-consuming,
determining
which
layers
use
challenging.
This
study
explores
different
strategies
for
five
SOTA
(VGG16,
VGG19,
ResNet50,
ResNet101,
InceptionV3)
on
ImageNet.
also
investigates
the
impact
of
classifier
by
using
linear
SVM
classification.
The
experiments
are
performed
four
open-access
ultrasound
datasets
related
breast
cancer,
thyroid
nodules
salivary
glands
cancer.
Results
evaluated
five-fold
stratified
cross-validation
technique,
metrics
like
accuracy,
precision,
recall
computed.
findings
show
that
15%
last
ResNet50
InceptionV3
achieves
good
results.
Using
classification
further
improves
overall
performance
6%
two
best-performing
models.
research
provides
insights
into
importance
transfer
Biomedicines,
Год журнала:
2022,
Номер
10(11), С. 2971 - 2971
Опубликована: Ноя. 18, 2022
Breast
cancer,
which
attacks
the
glandular
epithelium
of
breast,
is
second
most
common
kind
cancer
in
women
after
lung
and
it
affects
a
significant
number
people
worldwide.
Based
on
advantages
Residual
Convolutional
Network
Transformer
Encoder
with
Multiple
Layer
Perceptron
(MLP),
this
study
proposes
novel
hybrid
deep
learning
Computer-Aided
Diagnosis
(CAD)
system
for
breast
lesions.
While
backbone
residual
network
employed
to
create
features,
transformer
utilized
classify
according
self-attention
mechanism.
The
proposed
CAD
has
capability
recognize
two
scenarios:
Scenario
A
(Binary
classification)
B
(Multi-classification).
Data
collection
preprocessing,
patch
image
creation
splitting,
artificial
intelligence-based
lesion
identification
are
all
components
execution
framework
that
applied
consistently
across
both
cases.
effectiveness
AI
model
compared
against
three
separate
models:
custom
CNN,
VGG16,
ResNet50.
Two
datasets,
CBIS-DDSM
DDSM,
construct
test
system.
Five-fold
cross
validation
data
used
evaluate
accuracy
performance
results.
suggested
achieves
encouraging
evaluation
results,
overall
accuracies
100%
95.80%
binary
multiclass
prediction
challenges,
respectively.
experimental
results
reveal
could
identify
benign
malignant
tissues
significantly,
important
radiologists
recommend
further
investigation
abnormal
mammograms
provide
optimal
treatment
plan.
Diagnostics,
Год журнала:
2023,
Номер
13(13), С. 2242 - 2242
Опубликована: Июнь 30, 2023
This
study
aims
to
develop
an
efficient
and
accurate
breast
cancer
classification
model
using
meta-learning
approaches
multiple
convolutional
neural
networks.
Breast
Ultrasound
Images
(BUSI)
dataset
contains
various
types
of
lesions.
The
goal
is
classify
these
lesions
as
benign
or
malignant,
which
crucial
for
the
early
detection
treatment
cancer.
problem
that
traditional
machine
learning
deep
often
fail
accurately
images
due
their
complex
diverse
nature.
In
this
research,
address
problem,
proposed
used
several
advanced
techniques,
including
ensemble
technique,
transfer
learning,
data
augmentation.
Meta-learning
will
optimize
model's
process,
allowing
it
adapt
new
unseen
datasets
quickly.
Transfer
leverage
pre-trained
models
such
Inception,
ResNet50,
DenseNet121
enhance
feature
extraction
ability.
Data
augmentation
techniques
be
applied
artificially
generate
training
images,
increasing
size
diversity
dataset.
Meta
combine
outputs
CNNs,
improving
accuracy.
work
investigated
by
pre-processing
BUSI
first,
then
evaluating
CNNs
different
architectures
models.
Then,
a
algorithm
CNN.
Additionally,
evaluation
results
indicate
highly
effective
with
high
Finally,
performance
compared
state-of-the-art
in
other
existing
systems'
accuracy,
precision,
recall,
F1
score.
Diagnostics,
Год журнала:
2022,
Номер
12(10), С. 2541 - 2541
Опубликована: Окт. 20, 2022
Brain
tumors
(BTs)
are
deadly
diseases
that
can
strike
people
of
every
age,
all
over
the
world.
Every
year,
thousands
die
brain
tumors.
Brain-related
diagnoses
require
caution,
and
even
smallest
error
in
diagnosis
have
negative
repercussions.
Medical
errors
tumor
common
frequently
result
higher
patient
mortality
rates.
Magnetic
resonance
imaging
(MRI)
is
widely
used
for
evaluation
detection.
However,
MRI
generates
large
amounts
data,
making
manual
segmentation
difficult
laborious
work,
limiting
use
accurate
measurements
clinical
practice.
As
a
result,
automated
dependable
methods
required.
Automatic
early
detection
tasks
computer
vision
due
to
their
high
spatial
structural
variability.
Therefore,
or
treatment
critical.
Various
traditional
Machine
learning
(ML)
techniques
been
detect
various
types
The
main
issue
with
these
models
features
were
manually
extracted.
To
address
aforementioned
insightful
issues,
this
paper
presents
hybrid
deep
transfer
(GN-AlexNet)
model
BT
tri-classification
(pituitary,
meningioma,
glioma).
proposed
combines
GoogleNet
architecture
AlexNet
by
removing
five
layers
adding
ten
model,
which
extracts
classifies
them
automatically.
On
same
CE-MRI
dataset,
was
compared
(VGG-16,
AlexNet,
SqeezNet,
ResNet,
MobileNet-V2)
ML/DL.
outperformed
current
terms
accuracy
sensitivity
(accuracy
99.51%
98.90%).
Diagnostics,
Год журнала:
2022,
Номер
13(1), С. 89 - 89
Опубликована: Дек. 28, 2022
Early
detection
of
breast
cancer
is
an
essential
procedure
to
reduce
the
mortality
rate
among
women.
In
this
paper,
a
new
AI-based
computer-aided
diagnosis
(CAD)
framework
called
ETECADx
proposed
by
fusing
benefits
both
ensemble
transfer
learning
convolutional
neural
networks
as
well
self-attention
mechanism
vision
transformer
encoder
(ViT).
The
accurate
and
precious
high-level
deep
features
are
generated
via
backbone
network,
while
used
diagnose
probabilities
in
two
approaches:
Approach
A
(i.e.,
binary
classification)
B
multi-classification).
To
build
CAD
system,
benchmark
public
multi-class
INbreast
dataset
used.
Meanwhile,
private
real
images
collected
annotated
expert
radiologists
validate
prediction
performance
framework.
promising
evaluation
results
achieved
using
mammograms
with
overall
accuracies
98.58%
97.87%
for
approaches,
respectively.
Compared
individual
networks,
model
improves
6.6%
4.6%
approaches.
hybrid
shows
further
improvement
when
ViT-based
network
8.1%
6.2%
diagnosis,
For
validation
purposes
images,
system
provides
encouraging
97.16%
89.40%
has
capability
predict
lesions
single
mammogram
average
0.048
s.
Such
could
be
useful
helpful
assist
practical
applications
providing
second
supporting
opinion
distinguishing
various
malignancies.
Diagnostics,
Год журнала:
2022,
Номер
12(11), С. 2815 - 2815
Опубликована: Ноя. 16, 2022
Blood
cells
carry
important
information
that
can
be
used
to
represent
a
person's
current
state
of
health.
The
identification
different
types
blood
in
timely
and
precise
manner
is
essential
cutting
the
infection
risks
people
face
on
daily
basis.
BCNet
an
artificial
intelligence
(AI)-based
deep
learning
(DL)
framework
was
proposed
based
capability
transfer
with
convolutional
neural
network
rapidly
automatically
identify
eight-class
scenario:
Basophil,
Eosinophil,
Erythroblast,
Immature
Granulocytes,
Lymphocyte,
Monocyte,
Neutrophil,
Platelet.
For
purpose
establishing
dependability
viability
BCNet,
exhaustive
experiments
consisting
five-fold
cross-validation
tests
are
carried
out.
Using
strategy,
we
conducted
in-depth
comprehensive
BCNet's
architecture
test
it
three
optimizers
ADAM,
RMSprop
(RMSP),
stochastic
gradient
descent
(SGD).
Meanwhile,
performance
directly
compared
using
same
dataset
state-of-the-art
models
DensNet,
ResNet,
Inception,
MobileNet.
When
employing
optimizers,
demonstrated
better
classification
ADAM
RMSP
optimizers.
best
evaluation
achieved
optimizer
terms
98.51%
accuracy
96.24%
F1-score.
Compared
baseline
model,
clearly
improved
prediction
1.94%,
3.33%,
1.65%
RMSP,
SGD,
respectively.
model
outperformed
AI
DenseNet,
MobileNet
testing
time
single
cell
image
by
10.98,
4.26,
2.03,
0.21
msec.
In
comparison
most
recent
models,
could
able
generate
encouraging
outcomes.
It
for
advancement
healthcare
facilities
have
such
recognition
rate
improving
detection
cells.
Computers, materials & continua/Computers, materials & continua (Print),
Год журнала:
2023,
Номер
76(2), С. 2201 - 2216
Опубликована: Янв. 1, 2023
Breast
cancer
is
a
major
public
health
concern
that
affects
women
worldwide.
It
leading
cause
of
cancer-related
deaths
among
women,
and
early
detection
crucial
for
successful
treatment.
Unfortunately,
breast
can
often
go
undetected
until
it
has
reached
advanced
stages,
making
more
difficult
to
treat.
Therefore,
there
pressing
need
accurate
efficient
diagnostic
tools
detect
at
an
stage.
The
proposed
approach
utilizes
SqueezeNet
with
fire
modules
complex
bypass
extract
informative
features
from
mammography
images.
extracted
are
then
utilized
train
support
vector
machine
(SVM)
image
classification.
SqueezeNet-guided
SVM
model,
known
as
SNSVM,
achieved
promising
results,
accuracy
94.10%
sensitivity
94.30%.
A
10-fold
cross-validation
was
performed
ensure
the
robustness
mean
standard
deviation
various
performance
indicators
were
calculated
across
multiple
runs.
This
model
also
outperforms
state-of-the-art
models
in
all
indicators,
indicating
its
superior
performance.
demonstrates
effectiveness
diagnosis
using
makes
tool
diagnosis.
may
have
significant
implications
reducing
mortality
rates.
Decision Analytics Journal,
Год журнала:
2023,
Номер
7, С. 100240 - 100240
Опубликована: Май 4, 2023
With
the
extensive
applicability
of
machine
learning
classification
algorithms
to
a
wide
spectrum
domains,
feature
selection
(FS)
becomes
relevant
data
preprocessing
technique
due
high
dimensionality
used
in
these
domains.
While
efforts
have
been
made
study
various
filters
for
ranking
features,
scholars
paid
little
attention
developing
unified
framework
that
can
be
as
an
interface
any
filter.
The
development
such
would
formalize
understanding
filter-based
FS.
This
helps
put
same
perspective
when
analyzing
new
FS
algorithms.
proposes
based
on
best–worst
multi-attribute
decision-making
method.
proposed
algorithm
is
compared
two
control
groups:
(a)
no
and
(b)
randomized
algorithm.
Furthermore,
blocking
variables
are
considered:
(i)
classifier
(ii)
training
dataset.
performance
classifiers
was
measured
using
area
under
curve
(AUC)
receiver
operating
characteristics
(ROC)
curve.
A
three-way
analysis
variance
(ANOVA)
compare
approach
groups
considering
variables.
paper
offers
several
contributions
literature.
For
one
thing,
it
few
works
forward
performing
To
best
authors'
knowledge,
first
provide
empirical
evidence
about
interaction
between
factors
considered
literature
evaluating
Diagnostics,
Год журнала:
2023,
Номер
13(10), С. 1753 - 1753
Опубликована: Май 17, 2023
Breast
cancer
is
the
second
most
common
type
of
among
women,
and
it
can
threaten
women's
lives
if
not
diagnosed
early.
There
are
many
methods
for
detecting
breast
cancer,
but
they
cannot
distinguish
between
benign
malignant
tumors.
Therefore,
a
biopsy
taken
from
patient's
abnormal
tissue
an
effective
way
to
challenges
facing
pathologists
experts
in
diagnosing
including
addition
some
medical
fluids
various
colors,
direction
sample,
small
number
doctors
their
differing
opinions.
Thus,
artificial
intelligence
techniques
solve
these
help
clinicians
resolve
diagnostic
differences.
In
this
study,
three
techniques,
each
with
systems,
were
developed
diagnose
multi
binary
classes
datasets
types
40×
400×
factors.
The
first
technique
dataset
using
neural
network
(ANN)
selected
features
VGG-19
ResNet-18.
by
ANN
combined
ResNet-18
before
after
principal
component
analysis
(PCA).
third
analyzing
hybrid
features.
handcrafted;
handcrafted.
handcrafted
mixed
extracted
Fuzzy
color
histogram
(FCH),
local
pattern
(LBP),
discrete
wavelet
transform
(DWT)
gray
level
co-occurrence
matrix
(GLCM)
methods.
With
data
set,
reached
precision
95.86%,
accuracy
97.3%,
sensitivity
96.75%,
AUC
99.37%,
specificity
99.81%
images
at
magnification
factor
400×.
Whereas
99.74%,
99.7%,
100%,
99.85%,
100%