Medical Physics,
Journal Year:
2024,
Volume and Issue:
unknown
Published: Dec. 5, 2024
Abstract
Background
Accurate
and
automatic
segmentation
of
pericardial
adipose
tissue
(PEAT)
in
cardiac
magnetic
resonance
(MR)
images
is
essential
for
the
diagnosis
treatment
cardiovascular
diseases.
Precise
challenging
due
to
high
costs
need
specialized
knowledge,
as
a
large
amount
accurately
annotated
data
required,
demanding
significant
time
medical
resources.
Purpose
In
order
reduce
burden
annotation
while
maintaining
accuracy
tasks,
this
paper
introduces
semi‐supervised
learning
method
solve
limitations
current
PEAT
methods.
Methods
paper,
we
propose
difference‐guided
collaborative
mean
teacher
(DCMT)
method,
designed
from
DCMT
consists
two
main
components:
framework
with
difference
fusion
strategy
backbone
network
MCM‐UNet
using
Mamba‐CNN
mixture
(MCM)
blocks.
The
differential
effectively
utilizes
uncertain
areas
unlabeled
data,
encouraging
model
reach
consensus
predictions
across
these
difficult‐to‐segment
yet
information‐rich
areas.
addition,
considering
sparse
scattered
distribution
MR
images,
which
makes
it
segment,
our
framework.
This
not
only
enhances
processing
ability
global
information,
but
also
captures
detailed
local
features
image,
greatly
improves
segmentation.
Results
Our
experiments
conducted
on
MRPEAT
dataset
show
that
outperforms
existing
state‐of‐the‐art
methods
terms
accuracy.
These
findings
underscore
effectiveness
approach
handling
specific
challenges
associated
Conclusions
significantly
images.
By
utilizing
enhancing
feature
capture
MCM‐UNet,
demonstrates
superior
performance
offers
promising
solution
image
can
alleviate
extensive
requirements
typically
necessary
training
accurate
models
imaging.
Computerized Medical Imaging and Graphics,
Journal Year:
2024,
Volume and Issue:
115, P. 102383 - 102383
Published: April 17, 2024
Semi-supervised
learning
has
made
significant
progress
in
medical
image
segmentation.
However,
existing
methods
primarily
utilize
information
from
a
single
dimensionality,
resulting
sub-optimal
performance
on
challenging
magnetic
resonance
imaging
(MRI)
data
with
multiple
segmentation
objects
and
anisotropic
resolution.
To
address
this
issue,
we
present
Hybrid
Dual
Mean-Teacher
(HD-Teacher)
model
hybrid,
semi-supervised,
multi-task
to
achieve
effective
semi-supervised
HD-Teacher
employs
2D
3D
mean-teacher
network
produce
labels
signed
distance
fields
the
hybrid
captured
both
dimensionalities.
This
mechanism
allows
features
2D,
3D,
or
dimensions
as
needed.
Outputs
teacher
models
are
dynamically
combined
based
confidence
scores,
forming
prediction
estimated
uncertainty.
We
propose
regularization
module
encourage
student
results
close
uncertainty-weighted
further
improve
their
feature
extraction
capability.
Extensive
experiments
of
binary
multi-class
conducted
three
MRI
datasets
demonstrated
that
proposed
framework
could
(1)
significantly
outperform
state-of-the-art
(2)
surpass
fully-supervised
VNet
trained
substantially
more
annotated
data,
(3)
perform
par
human
raters
muscle
bone
task.
Code
will
be
available
at
https://github.com/ThisGame42/Hybrid-Teacher.
Medical Image Analysis,
Journal Year:
2024,
Volume and Issue:
95, P. 103183 - 103183
Published: April 21, 2024
Automated
segmentation
is
a
challenging
task
in
medical
image
analysis
that
usually
requires
large
amount
of
manually
labeled
data.
However,
most
current
supervised
learning
based
algorithms
suffer
from
insufficient
manual
annotations,
posing
significant
difficulty
for
accurate
and
robust
segmentation.
In
addition,
semi-supervised
methods
lack
explicit
representations
geometric
structure
semantic
information,
restricting
accuracy.
this
work,
we
propose
hybrid
framework
to
learn
polygon
vertices,
region
masks,
their
boundaries
weakly/semi-supervised
manner
significantly
advances
representations.
Firstly,
multi-granularity
constraints
via
vertices
(PolyV)
pixel-wise
(PixelR)
masks
manner.
Secondly,
eliminating
boundary
ambiguity
by
using
an
contrastive
objective
discriminative
feature
space
contours
at
the
pixel
level
with
limited
annotations.
Thirdly,
exploit
task-specific
clinical
domain
knowledge
differentiate
function
assessment
end-to-end.
The
ground
truth
assessment,
on
other
hand,
can
serve
as
auxiliary
weak
supervision
PolyV
PixelR
learning.
We
evaluate
proposed
two
tasks,
including
optic
disc
(OD)
cup
(OC)
along
vertical
cup-to-disc
ratio
(vCDR)
estimation
fundus
images;
left
ventricle
(LV)
end-diastolic
end-systolic
frames
ejection
fraction
(LVEF)
two-dimensional
echocardiography
images.
Experiments
nine
large-scale
datasets
tasks
under
different
label
settings
demonstrate
our
model's
superior
performance
assessment.
Scientific Reports,
Journal Year:
2024,
Volume and Issue:
14(1)
Published: Sept. 20, 2024
The
analysis
and
interpretation
of
cytopathological
images
are
crucial
in
modern
medical
diagnostics.
However,
manually
locating
identifying
relevant
cells
from
the
vast
amount
image
data
can
be
a
daunting
task.
This
challenge
is
particularly
pronounced
developing
countries
where
there
may
shortage
expertise
to
handle
such
tasks.
acquiring
large
amounts
high-quality
labelled
remains,
many
researchers
have
begun
use
semi-supervised
learning
methods
learn
unlabeled
data.
Although
current
models
partially
solve
issue
limited
data,
they
inefficient
exploiting
samples.
To
address
this,
we
introduce
new
AI-assisted
scheme,
Reliable-Unlabeled
Semi-Supervised
Segmentation
(RU3S)
model.
model
integrates
ResUNet-SE-ASPP-Attention
(RSAA)
model,
which
includes
Squeeze-and-Excitation
(SE)
network,
Atrous
Spatial
Pyramid
Pooling
(ASPP)
structure,
Attention
module,
ResUNet
architecture.
Our
leverages
effectively,
improving
accuracy
significantly.
A
novel
confidence
filtering
strategy
introduced
make
better
samples,
addressing
scarcity
Experimental
results
show
2.0%
improvement
mIoU
over
state-of-the-art
segmentation
ST,
demonstrating
our
approach's
effectiveness
solving
this
problem.
