2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM),
Год журнала:
2022,
Номер
unknown, С. 3015 - 3022
Опубликована: Дек. 6, 2022
Due
to
the
shortage
of
training
data,
transfer
learning
is
frequently
used
in
constructing
medical
imaging
models.
In
this
study,
we
perform
pre-training
dataset
and
fine-tuning
effect
analysis
cancer
histopathology
by
evaluating
three
popular
deep
neural
network
algorithms
on
target
datasets
under
various
configurations.
Pre-training
models
with
image
appear
worse
or
not
better
than
ImageNet
random
initialization.
Furthermore,
study
demonstrates
that
performance
pre-trained
improves
increase
images
fine-tuning,
which
was
previously
overlooked.
Recently,
deep
learning
techniques
have
been
widely
used
in
content-based
image
retrieval
(CBIR)
due
to
their
success
bridging
the
"semantic
gap"
issue.
Nevertheless,
high-dimensional
features
extracted
through
models
usually
lead
a
high
computational
cost.
This
issue
has
major
impact
especially
when
it
involves
of
medical
images,
where
short
response
time
is
very
important.
To
address
this
issue,
we
propose
paper
an
effective
optimization
approach
for
reducing
dimension
from
colon
histology
images.
Specifically,
extract
features,
first
apply
transfer
on
pre-trained
ResNet18
and
GoogLeNet
networks.
Then,
use
sine
cosine
algorithm
(SCA)
optimize
combined
features.
Experiments
conducted
Kather-5k
colorectal
cancer
dataset
demonstrate
effectiveness
proposed
dimensionality
reduction
method
feature
dimension,
with
gain
up
50%
better,
while
keeping
good
performance.
Cancers,
Год журнала:
2024,
Номер
16(22), С. 3879 - 3879
Опубликована: Ноя. 20, 2024
Lung
and
colon
cancers
are
among
the
leading
causes
of
cancer-related
mortality
worldwide.
Early
accurate
detection
these
is
crucial
for
effective
treatment
improved
patient
outcomes.
False
or
incorrect
harmful.
Accurately
detecting
cancer
in
a
patient's
tissue
to
their
treatment.
While
analyzing
samples
complicated
time-consuming,
deep
learning
techniques
have
made
it
possible
complete
this
process
more
efficiently
accurately.
As
result,
researchers
can
study
patients
shorter
amount
time
at
lower
cost.
Much
research
has
been
conducted
investigate
models
that
require
great
computational
ability
resources.
However,
none
had
100%
rate
life-threatening
malignancies.
Misclassified
falsely
very
harmful
consequences.
This
proposes
new
lightweight,
parameter-efficient,
mobile-embedded
model
based
on
1D
convolutional
neural
network
with
squeeze-and-excitation
layers
efficient
lung
detection.
proposed
diagnoses
classifies
squamous
cell
carcinomas
adenocarcinoma
from
digital
pathology
images.
Extensive
experiment
demonstrates
our
achieves
accuracy
lung,
colon,
histopathological
(LC25000)
datasets,
which
considered
best
around
0.35
million
trainable
parameters
6.4
flops.
Compared
existing
results,
architecture
shows
state-of-the-art
performance
2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM),
Год журнала:
2022,
Номер
unknown, С. 3015 - 3022
Опубликована: Дек. 6, 2022
Due
to
the
shortage
of
training
data,
transfer
learning
is
frequently
used
in
constructing
medical
imaging
models.
In
this
study,
we
perform
pre-training
dataset
and
fine-tuning
effect
analysis
cancer
histopathology
by
evaluating
three
popular
deep
neural
network
algorithms
on
target
datasets
under
various
configurations.
Pre-training
models
with
image
appear
worse
or
not
better
than
ImageNet
random
initialization.
Furthermore,
study
demonstrates
that
performance
pre-trained
improves
increase
images
fine-tuning,
which
was
previously
overlooked.
