Bioengineering,
Journal Year:
2025,
Volume and Issue:
12(6), P. 608 - 608
Published: June 3, 2025
The
rapid
advancement
of
prompt-based
models
in
natural
language
processing
and
image
generation
has
revolutionized
the
field
segmentation.
introduction
Segment
Anything
Model
(SAM)
further
invigorated
this
domain
with
its
unprecedented
versatility.
However,
applicability
to
medical
segmentation
remains
uncertain
due
significant
disparities
between
images,
which
demand
careful
consideration.
This
study
comprehensively
analyzes
recent
efforts
adapt
SAM
for
segmentation,
including
empirical
benchmarking
methodological
refinements
aimed
at
bridging
gap
SAM’s
capabilities
unique
challenges
imaging.
Furthermore,
we
explore
future
directions
field.
While
direct
application
complex,
multimodal,
multi-target
datasets
may
not
yet
yield
optimal
results,
insights
from
these
provide
crucial
guidance
developing
foundational
tailored
intricacies
analysis.
Despite
existing
challenges,
holds
considerable
potential
demonstrate
advantages
robust
near
future.
IEEE Transactions on Circuits and Systems for Video Technology,
Journal Year:
2024,
Volume and Issue:
34(7), P. 5414 - 5423
Published: Jan. 1, 2024
Due
to
the
various
appearance
of
polyps
and
tiny
contrast
between
polyp
area
its
surrounding
background,
accurate
segmentation
has
become
a
challenging
task.
To
tackle
this
issue,
we
introduce
boundary-enhanced
framework
for
segmentation,
called
Focused
on
Boundary
Segmentation
(FoBS)
framework,
that
leverages
multi-level
collaboration
among
sample,
feature,
optimization.
It
places
greater
emphasis
boundary
improve
accuracy
segmentation.
Firstly,
boundary-aware
mixup
method
is
designed
model's
awareness
boundary.
More
importantly,
propose
deformable
laplacian-based
feature
refining
explicitly
strengthen
representation
ability
features.
employs
Laplacian
refinement
function
capture
discriminative
information
from
perceptual
field,
thereby
improving
adapt
variations.
In
addition,
self-adjusting
coefficient
learning
enables
adaptive
control
over
strength
at
each
location.
Furthermore,
develop
location-sensitive
compensation
criterion
assigns
more
importance
degraded
after
during
Extensive
quantitative
qualitative
experiments
four
benchmarks
demonstrate
effectiveness
our
automatic
Our
code
available
https://github.com/TFboys-lzz/
FoBS.
Breast
cancer
is
one
of
the
most
common
cancers
among
women
worldwide,
with
early
detection
significantly
increasing
survival
rates.
Ultrasound
imaging
a
critical
diagnostic
tool
that
aids
in
by
providing
real-time
breast
tissue.
We
conducted
thorough
investigation
Segment
Anything
Model
(SAM)
for
task
interactive
segmentation
tumors
ultrasound
images.
explored
three
pre-trained
model
variants:
ViT_h,
ViT_l,
and
ViT_b,
which
ViT_l
demonstrated
superior
performance
terms
mean
pixel
accuracy,
Dice
score,
IoU
score.
The
significance
prompt
interaction
improving
model's
was
also
highlighted,
substantial
improvements
metrics
when
prompts
were
incorporated.
study
further
evaluated
differential
segmenting
malignant
benign
tumors,
showing
exceptional
proficiency
both
categories,
albeit
slightly
better
tumors.
Furthermore,
we
analyzed
impacts
various
tumor
characteristics--size,
contrast,
aspect
ratio,
complexity--on
performance.
Our
findings
reveal
contrast
size
positively
impact
result,
while
complex
boundaries
pose
challenges.
provides
valuable
insights
using
SAM
as
robust
effective
algorithm
The
purpose
of
this
study
is
to
reduce
radiation
exposure
in
PET
imaging
while
preserving
high-quality
clinical
images.
We
propose
the
Consistency
Model
(PET-CM),
an
efficient
diffusion-model-based
approach,
estimate
full-dose
images
from
low-dose
PETs.
PET-CM
delivers
synthetic
comparable
quality
state-of-the-art
diffusion-based
methods
but
with
significantly
higher
efficiency.
process
involves
adding
Gaussian
noise
PETs
through
a
forward
diffusion
and
then
using
U-shaped
network
(PET-Unet)
for
denoising
reverse
process,
conditioned
on
corresponding
In
experiments
one-eighth
dose
images,
achieved
MAE
1.321±0.134%,
PSNR
33.587±0.674
dB,
SSIM
0.960±0.008,
NCC
0.967±0.011.
scenarios
reducing
1/4
full
dose,
further
showcased
its
capability
1.123±0.112%,
35.851±0.871
0.975±0.003,
0.990±0.003.
IEEE Transactions on Image Processing,
Journal Year:
2024,
Volume and Issue:
33, P. 6204 - 6215
Published: Jan. 1, 2024
Medical
image
segmentation
is
a
critical
task
in
clinical
applications.
Recently,
the
Segment
Anything
Model
(SAM)
has
demonstrated
potential
for
natural
segmentation.
However,
requirement
expert
labour
to
provide
prompts,
and
domain
gap
between
medical
images
pose
significant
obstacles
adapting
SAM
images.
To
overcome
these
challenges,
this
paper
introduces
novel
prompt
generation
method
named
EviPrompt.
The
proposed
requires
only
single
reference
image-annotation
pair,
making
it
training-free
solution
that
significantly
reduces
need
extensive
labelling
computational
resources.
First,
prompts
are
automatically
generated
based
on
similarity
features
of
target
images,
evidential
learning
introduced
improve
reliability.
Then,
mitigate
impact
gap,
committee
voting
inference-guided
in-context
employed,
generating
primarily
human
prior
knowledge
reducing
reliance
extracted
semantic
information.
EviPrompt
represents
an
efficient
robust
approach
We
evaluate
across
broad
range
tasks
modalities,
confirming
its
efficacy.
source
code
available
at
https://github.com/SPIresearch/EviPrompt.
Remote Sensing,
Journal Year:
2025,
Volume and Issue:
17(2), P. 342 - 342
Published: Jan. 20, 2025
With
the
continuous
advancement
of
deep
neural
networks,
salient
object
detection
(SOD)
in
natural
images
has
made
significant
progress.
However,
SOD
optical
remote
sensing
(ORSI-SOD)
remains
a
challenging
task
due
to
diversity
objects
and
complexity
backgrounds.
The
primary
challenge
lies
generating
robust
features
that
can
effectively
integrate
both
global
semantic
information
for
localization
local
spatial
details
boundary
reconstruction.
Most
existing
ORSI-SOD
methods
rely
on
pre-trained
CNN-
or
Transformer-based
backbones
extract
from
ORSIs,
followed
by
multi-level
feature
aggregation.
Given
differences
between
ORSIs
used
pre-training,
generalization
capability
these
backbone
networks
is
often
limited,
resulting
suboptimal
performance.
Recently,
prompt
engineering
been
employed
enhance
ability
Segment
Anything
Model
(SAM),
an
emerging
vision
foundation
model
achieved
remarkable
success
across
various
tasks.
Despite
its
success,
directly
applying
SAM
without
prompts
manual
interaction
unsatisfactory.
In
this
paper,
we
propose
novel
progressive
self-prompting
based
SAM,
termed
PSP-SAM,
which
generates
internal
external
network
overcome
limitations
ORSI-SOD.
Specifically,
domain-specific
prompting
modules,
consisting
block-shared
block-specific
adapters,
are
integrated
into
learn
visual
within
backbone,
facilitating
adaptation
Furthermore,
introduce
decoder
module
performs
prompt-guided
integration
stage-wise
mask
progressively,
enabling
prompt-based
decoders
outside
predict
saliency
maps
coarse-to-fine
manner.
entire
trained
end-to-end
with
parameter-efficient
fine-tuning.
Extensive
experiments
three
benchmark
datasets
demonstrate
our
proposed
achieves
state-of-the-art
International Journal of Imaging Systems and Technology,
Journal Year:
2025,
Volume and Issue:
35(2)
Published: Feb. 6, 2025
ABSTRACT
Current
colorectal
polyps
detection
methods
often
struggle
with
efficiency
and
boundary
precision,
especially
when
dealing
of
complex
shapes
sizes.
Traditional
techniques
may
fail
to
precisely
define
the
boundaries
these
polyps,
leading
suboptimal
rates.
Furthermore,
flat
small
blend
into
background
due
their
low
contrast
against
mucosal
wall,
making
them
even
more
challenging
detect.
To
address
challenges,
we
introduce
SCABNet,
a
novel
deep
learning
architecture
for
efficient
polyps.
SCABNet
employs
an
encoder‐decoder
structure
three
blocks:
Feature
Enhancement
Block
(FEB),
Channel
Prioritization
(CPB),
Spatial‐Gradient
Boundary
Attention
(SGBAB).
The
FEB
applies
dilation
spatial
attention
high‐level
features,
enhancing
discriminative
power
improving
model's
ability
capture
patterns.
CPB,
alternative
traditional
channel
blocks,
assigns
prioritization
weights
diverse
feature
channels.
SGBAB
replaces
conventional
mechanisms
solution
that
focuses
on
map.
It
Jacobian‐based
approach
construct
learned
convolutions
both
vertical
horizontal
components
This
allows
effectively
understand
changes
in
map
across
different
locations,
which
is
crucial
detecting
complex‐shaped
These
blocks
are
strategically
embedded
within
network's
skip
connections,
capabilities
without
imposing
excessive
computational
demands.
They
exploit
enhance
features
at
levels:
high,
mid,
low,
thereby
ensuring
wide
range
has
been
trained
Kvasir‐SEG
CVC‐ClinicDB
datasets
evaluated
multiple
datasets,
demonstrating
superior
results.
code
available
on:
https://github.com/KhaledELKarazle97/SCABNet
.
Machine Learning and Knowledge Extraction,
Journal Year:
2025,
Volume and Issue:
7(1), P. 22 - 22
Published: Feb. 24, 2025
We
introduce
a
weakly
supervised
segmentation
approach
that
leverages
class
activation
maps
and
the
Segment
Anything
Model
to
generate
high-quality
masks
using
only
classification
data.
A
pre-trained
classifier
produces
that,
once
thresholded,
yield
bounding
boxes
encapsulating
regions
of
interest.
These
prompt
SAM
detailed
masks,
which
are
then
refined
by
selecting
best
overlap
with
automatically
generated
from
foundational
model
intersection
over
union
metric.
In
polyp
case
study,
our
outperforms
existing
zero-shot
methods,
achieving
mean
0.63.
This
method
offers
an
efficient
general
solution
for
image
tasks
where
data
scarce.
Journal of Medical Imaging,
Journal Year:
2025,
Volume and Issue:
12(02)
Published: Feb. 27, 2025
PurposeSegmentation
of
ultrasound
images
for
medical
diagnosis,
monitoring,
and
research
is
crucial,
although
existing
methods
perform
well,
they
are
limited
by
specific
organs,
tumors,
image
devices.
Applications
the
Segment
Anything
Model
(SAM),
such
as
SAM-med2d,
use
a
large
number
datasets
that
contain
only
small
fraction
images.ApproachIn
this
work,
we
proposed
SAM-MedUS
model
generic
segmentation
utilizes
latest
publicly
available
dataset
to
create
diverse
containing
eight
site
categories
training
testing.
We
integrated
ConvNext
V2
CM
blocks
in
encoder
better
global
context
extraction.
In
addition,
boundary
loss
function
used
improve
fuzzy
boundaries
low-contrast
images.ResultsExperimental
results
show
outperforms
recent
on
multiple
datasets.
For
more
easily
adult
kidney,
it
achieves
87.93%
IoU
93.58%
dice,
whereas
complex
ones
infant
vein,
dice
reach
62.31%
78.93%,
respectively.ConclusionsWe
collected
collated
an
different
types
achieve
uniform
images.
additional
auxiliary
branches
block
enhances
ability
extract
information
allows
exhibit
robust
performance
excellent
generalization
ability.