Semi-supervised
learning
techniques
leverage
both
labeled
and
unlabeled
images
to
enhance
classification
performance
in
scenarios
where
are
limited.
However,
challenges,
such
as
determining
appropriate
thresholds,
integrating
incorrect
pseudo-labels,
establishing
effective
consistency
augmentations,
hinder
the
effectiveness
of
existing
methods.
Additionally,
label
prediction
fluctuations
on
low-confidence
their
impact
generalization
pose
further
limitations.
This
research
introduces
a
novel
framework
named
interpolation
for
bad
semi-supervised
generative
adversarial
networks
(ICBSGAN)
which
addresses
limitations
through
utilization
new
loss
function.
The
proposed
model
combines
training,
fusion
techniques,
regularization
learning.
ICBSGAN
incorporates
three
types
training:
fake
images,
real
images.
improve
generation
diverse
support
vectors
low-density
areas.
It
demonstrates
linear
behavior
at
interpolation,
reducing
predictions,
improving
stability,
identification
decision
boundaries.
Experimental
evaluations
CIFAR-10,
CINIC-10,
MNIST,
SVHN
datasets
showcase
compared
state-of-the-art
approach
achieves
notable
improvements
error
rate
from
2.87
1.47
3.89
3.13
SVHN,
15.48
9.59
CIFAR-10
using
1000
training
it
reduces
22.11
18.40
CINIC-10
when
700
per
class.
code
can
be
found
following
GitHub
repository:
https://github.com/ms-iraji/ICBSGAN
↗.
Wiley Interdisciplinary Reviews Data Mining and Knowledge Discovery,
Год журнала:
2025,
Номер
15(1)
Опубликована: Янв. 20, 2025
ABSTRACT
Medical
image
segmentation
is
prerequisite
in
computer‐aided
diagnosis.
As
the
field
experiences
tremendous
paradigm
changes
since
introduction
of
foundation
models,
technicality
deep
medical
model
no
longer
a
privilege
limited
to
computer
science
researchers.
A
comprehensive
educational
resource
suitable
for
researchers
broad,
different
backgrounds
such
as
biomedical
and
medicine,
needed.
This
review
strategically
covers
evolving
trends
that
happens
fundamental
components
emerging
multimodal
datasets,
updates
on
learning
libraries,
classical‐to‐contemporary
development
models
latest
challenges
with
focus
enhancing
interpretability
generalizability
model.
Last,
conclusion
section
highlights
future
worth
further
attention
investigations.
Journal of bone oncology,
Год журнала:
2023,
Номер
42, С. 100498 - 100498
Опубликована: Авг. 17, 2023
The
objective
of
this
study
was
to
investigate
the
use
contrast-enhanced
magnetic
resonance
imaging
(CE-MRI)
combined
with
radiomics
and
deep
learning
technology
for
identification
spinal
metastases
primary
malignant
bone
tumor.The
region
growing
algorithm
utilized
segment
lesions,
two
parameters
were
defined
based
on
interest
(ROI).
Deep
algorithms
employed:
improved
U-Net,
which
CE-MRI
parameter
maps
as
input,
used
10
layers
CE
images
input.
Inception-ResNet
model
extract
relevant
features
disease
construct
a
diagnosis
classifier.The
diagnostic
accuracy
0.74,
while
average
U-Net
0.98,
respectively.
PA
our
is
high
98.001%.
findings
indicate
that
have
potential
assist
in
differential
tumor.CE-MRI
can
potentially
tumor,
providing
promising
approach
clinical
diagnosis.
Machine Learning with Applications,
Год журнала:
2024,
Номер
16, С. 100549 - 100549
Опубликована: Апрель 5, 2024
Infrared
(IR)
spectroscopic
imaging
is
of
potentially
wide
use
in
medical
applications
due
to
its
ability
capture
both
chemical
and
spatial
information.
This
complexity
the
data
necessitates
using
machine
intelligence
as
well
presents
an
opportunity
harness
a
high-dimensionality
set
that
offers
far
more
information
than
today's
manually-interpreted
images.
While
convolutional
neural
networks
(CNNs),
including
well-known
U-Net
model,
have
demonstrated
impressive
performance
image
segmentation,
inherent
locality
convolution
limits
effectiveness
these
models
for
encoding
IR
data,
resulting
suboptimal
performance.
In
this
work,
we
propose
INfrared
Spectroscopic
imaging-based
TRAnsformers
Segmentation
(INSTRAS).
novel
model
leverages
strength
transformer
encoders
segment
breast
images
effectively.
Incorporating
skip-connection
encoders,
INSTRAS
overcomes
issue
pure
models,
such
difficulty
capturing
long-range
dependencies.
To
evaluate
our
existing
conducted
training
on
various
encoder-decoder
dataset
INSTRAS,
utilizing
9
spectral
bands
achieved
remarkable
