Oral
cancer
is
one
of
the
most
commonly
found
cancers
worldwide.
Epithelial
Dysplasia
(OED)
an
Potentially
Malignant
Disorder
(OPMD)
that
can
be
characterized
for
preventive
oral
screening.
The
standard
OED
histological
grading
conducted
via
epithelial
regions
tissue
biopsies.
However,
this
procedure
laborious,
time-consuming,
and
subjective;
consequently,
it
prone
to
variability
due
fatigue
limited
expertise.
Therefore,
study
aims
explore
potential
using
Convolutional
Neural
Network
(CNN)
Transformer
models
automated
epithelium
segmentation
algorithm
directly
from
Whole
Slide
Images
(WSIs).
This
approach
reduce
manual
process
support
pathologists
in
activities.
Accordingly,
candidate
architectures
based
on
CNN
are
selected:
UNet,
ResNet50-UNet,
VGG19-UNet,
Swin-UNet,
MISSFormer.
These
trained
patch-based
mitigate
high
computational
cost
caused
by
processing
WSIs.
results
indicate
optimized
with
ADAM
optimizer,
demonstrates
best
performance
Intersection
over
Union
(IoU)
0.82
Dice-Similarity
Coefficient
(DSC)
0.87.
Furthermore,
model
achieves
highest
IoU
DSC
tissue-level
prediction,
scoring
0.88
0.94,
respectively.
According
experiment,
overlapping
non-overlapping
patching
strategies
perform
similarly
selected
architectures.
latter
approach,
hence,
suggested
efficiency.
enhancing
provide
a
reliable
tool
assisting
pathologists.
Medical Physics,
Journal Year:
2025,
Volume and Issue:
unknown
Published: March 20, 2025
Accurate
segmentation
of
diffuse
large
B-cell
lymphoma
(DLBCL)
lesions
is
challenging
due
to
their
complex
patterns
in
medical
imaging.
Traditional
methods
often
struggle
delineate
these
accurately.
This
study
aims
develop
a
precise
method
for
DLBCL
using
18F-fluorodeoxyglucose
(18F-FDG)
positron
emission
tomography
(PET)
and
computed
(CT)
images.
We
propose
3D
based
on
an
encoder-decoder
architecture.
The
encoder
incorporates
dual-branch
design
the
shifted
window
transformer
extract
features
from
both
PET
CT
modalities.
To
enhance
feature
integration,
we
introduce
multi-scale
information
fusion
(MSIF)
module
that
performs
cross-attention
mechanisms
with
framework.
A
gated
neural
network
within
MSIF
dynamically
adjusts
weights
balance
contributions
each
modality.
model
optimized
dice
similarity
coefficient
(DSC)
loss
function,
minimizing
discrepancies
between
prediction
ground
truth.
Additionally,
total
metabolic
tumor
volume
(TMTV)
performed
statistical
analyses
results.
was
trained
validated
private
dataset
165
patients
publicly
available
(autoPET)
containing
145
PET/CT
scans
patients.
Both
datasets
were
analyzed
five-fold
cross-validation.
On
dataset,
our
achieved
DSC
0.7512,
sensitivity
0.7548,
precision
0.7611,
average
surface
distance
(ASD)
3.61
mm,
Hausdorff
at
95th
percentile
(HD95)
15.25
mm.
autoPET
0.7441,
0.7573,
0.7427,
ASD
5.83
HD95
21.27
outperforming
state-of-the-art
(p
<
0.05,
t-test).
For
TMTV
quantification,
Pearson
correlation
coefficients
0.91
(private
dataset)
0.86
observed,
R2
values
0.89
0.75,
respectively.
Extensive
ablation
studies
demonstrated
module's
contribution
enhanced
accuracy.
presents
effective
automatic
leverages
complementary
strengths
demonstrates
robust
performance
datasets,
ensuring
its
reliability
generalizability.
Our
provides
clinicians
more
delineation,
which
can
improve
accuracy
diagnostic
interpretations
assist
treatment
planning
code
proposed
https://github.com/chenzhao2023/lymphoma_seg.
Bioengineering,
Journal Year:
2025,
Volume and Issue:
12(4), P. 390 - 390
Published: April 5, 2025
Rheumatoid
arthritis
(RA)
is
a
chronic
autoimmune
disease
that
can
cause
severe
joint
damage
and
functional
impairment.
Ultrasound
imaging
has
shown
promise
in
providing
real-time
assessment
of
synovium
inflammation
associated
with
the
early
stages
RA.
Accurate
segmentation
region
quantification
inflammation-specific
biomarkers
are
crucial
for
assessing
grading
However,
automatic
3D
ultrasound
challenging
due
to
ambiguous
boundaries,
variability
shape,
inhomogeneous
intensity
distribution.
In
this
work,
we
introduce
novel
network
architecture,
Swin
Transformers
Deep
Attentive
Features
(SwinDAF3D),
which
integrates
into
framework.
The
developed
architecture
leverages
hierarchical
structure
shifted
windows
capture
rich,
multi-scale
attentive
contextual
information,
improving
modeling
long-range
dependencies
spatial
hierarchies
images.
six-fold
cross-validation
study
images
RA
patients’
finger
joints
(n
=
72),
our
SwinDAF3D
model
achieved
highest
performance
Dice
Score
(DSC)
0.838
±
0.013,
an
Intersection
over
Union
(IoU)
0.719
0.019,
Surface
(SDSC)
0.852
0.020,
compared
UNet
(DSC:
0.742
0.025;
IoU:
0.589
0.031;
SDSC:
0.661
0.029),
DAF3D
0.813
0.017;
0.689
0.022;
0.817
0.013),
UNETR
0.808
0.678
0.032;
0.822
0.039),
UNETR++
0.810
0.014;
0.684
0.018;
0.829
0.027)
TransUNet
0.818
0.013;
0.692
0.815
0.016)
models.
This
ablation
demonstrates
effectiveness
combining
feature
pyramid
deep
attention
mechanism,
accuracy
ultrasound.
advancement
shows
great
enabling
more
efficient
standardized
screening
using
imaging.
Medical Physics,
Journal Year:
2025,
Volume and Issue:
unknown
Published: April 24, 2025
Abstract
Background
The
magnetic
resonance
(MR)
image
translation
model
is
designed
to
generate
MR
images
of
required
sequence
from
the
existing
sequence.
However,
generalization
performance
generation
models
on
external
datasets
tends
be
unsatisfactory
due
inconsistency
in
data
distribution
across
different
centers
or
scanners.
Purpose
aim
this
study
propose
a
cross‐sequence
synthesis
that
could
high‐quality
synthetic
with
high
transferability
for
small‐sized
datasets.
Methods
We
proposed
dual‐way
using
transformer‐based
adversarial
network
(DMTrans)
sequences.
It
integrates
generative
architecture
an
innovative
discriminator
design.
shifted
window‐based
multi‐head
self‐attention
mechanism
DMTrans
enables
efficient
capture
global
and
local
features
images.
sequential
dual‐scale
distinguish
generated
at
multi‐scale.
Results
pre‐trained
bi‐directional
T1/T2‐weighted
dataset
comprising
4229
slices.
demonstrates
superior
baseline
methods
both
qualitative
quantitative
measurements.
SSIM,
PSNR,
MAE
metrics
T1
based
T2
are
0.91
±
0.04,
25.30
2.40,
24.65
10.46,
while
metric
values
0.90
24.72
1.62,
23.28
7.40
opposite
direction.
Fine‐tuning
then
utilized
adapt
another
public
T1/T2/proton‐weighted
(PD)
images,
so
only
6
patients
500
slices
adaptation
achieve
T1/T2,
T1/PD,
T2/PD
results.
Conclusions
achieves
state‐of‐the‐art
conversion,
which
provide
more
information
assisting
clinical
diagnosis
treatment.
also
offered
versatile
solution
needs
data‐scarce
conditions
centers.
Tomography,
Journal Year:
2023,
Volume and Issue:
9(5), P. 1933 - 1948
Published: Oct. 18, 2023
Convolutional
neural
networks
(CNNs)
have
a
proven
track
record
in
medical
image
segmentation.
Recently,
Vision
Transformers
were
introduced
and
are
gaining
popularity
for
many
computer
vision
applications,
including
object
detection,
classification,
Machine
learning
algorithms
such
as
CNNs
or
subject
to
an
inductive
bias,
which
can
significant
impact
on
the
performance
of
machine
models.
This
is
especially
relevant
segmentation
applications
where
limited
training
data
available,
model’s
bias
should
help
it
generalize
well.
In
this
work,
we
quantitatively
assess
two
CNN-based
(U-Net
U-Net-CBAM)
three
popular
Transformer-based
network
architectures
(UNETR,
TransBTS,
VT-UNet)
context
HNC
lesion
volumetric
[F-18]
fluorodeoxyglucose
(FDG)
PET
scans.
For
assessment,
272
FDG
PET-CT
scans
clinical
trial
(ACRIN
6685)
utilized,
includes
total
650
lesions
(primary:
secondary:
378).
The
used
highly
diverse
representative
use.
analysis,
several
error
metrics
utilized.
achieved
Dice
coefficient
ranged
from
0.833
0.809
with
best
being
by
approaches.
U-Net-CBAM,
utilizes
spatial
channel
attention,
showed
advantages
smaller
compared
standard
U-Net.
Furthermore,
our
results
provide
some
insight
regarding
features
specific
application.
addition,
highlight
need
utilize
primary
well
secondary
derive
clinically
estimates
avoiding
biases.
Cancers,
Journal Year:
2024,
Volume and Issue:
16(14), P. 2538 - 2538
Published: July 14, 2024
We
introduce
an
innovative,
simple,
effective
segmentation-free
approach
for
survival
analysis
of
head
and
neck
cancer
(HNC)
patients
from
PET/CT
images.
By
harnessing
deep
learning-based
feature
extraction
techniques
multi-angle
maximum
intensity
projections
(MA-MIPs)
applied
to
Fluorodeoxyglucose
Positron
Emission
Tomography
(FDG-PET)
images,
our
proposed
method
eliminates
the
need
manual
segmentations
regions-of-interest
(ROIs)
such
as
primary
tumors
involved
lymph
nodes.
Instead,
a
state-of-the-art
object
detection
model
is
trained
utilizing
CT
images
perform
automatic
cropping
anatomical
area,
instead
only
lesions
or
nodes
on
PET
volumes.
A
pre-trained
convolutional
neural
network
backbone
then
utilized
extract
features
MA-MIPs
obtained
72
multi-angel
axial
rotations
cropped
These
extracted
multiple
projection
views
volumes
are
aggregated
fused,
employed
recurrence-free
cohort
489
HNC
patients.
The
outperforms
best
performing
target
dataset
task
analysis.
circumventing
delineation
malignancies
FDG
PET-CT
dependency
subjective
interpretations
highly
enhances
reproducibility
method.
code
this
work
publicly
released.
Quantitative Imaging in Medicine and Surgery,
Journal Year:
2023,
Volume and Issue:
14(1), P. 1122 - 1140
Published: Dec. 29, 2023
Automatic
tumor
segmentation
is
a
critical
component
in
clinical
diagnosis
and
treatment.
Although
single-modal
imaging
provides
useful
information,
multi-modal
more
comprehensive
understanding
of
the
tumor.
Multi-modal
has
been
an
essential
topic
medical
image
processing.
With
remarkable
performance
deep
learning
(DL)
methods
analysis,
based
on
DL
attracted
significant
attention.
This
study
aimed
to
provide
overview
recent
DL-based
methods.