Digital Diagnostics,
Journal Year:
2024,
Volume and Issue:
unknown
Published: Nov. 21, 2024
Data
processing
methods
using
neural
networks
are
gaining
increasing
popularity
in
a
variety
of
medical
diagnostic
problems.
Most
often,
such
used
the
study
images
human
organs
CT
scan
and
magnetic
resonance
imaging,
ultrasound
other
non-invasive
research
methods.
Diagnosing
pathology
this
case
is
problem
segmenting
image,
that
is,
searching
for
groups
(regions)
pixels
characterize
certain
objects
them.
One
most
successful
solving
U-Net
network
architecture
developed
2015.
This
review
examines
various
modifications
classic
architecture.
The
reviewed
papers
divided
into
several
key
areas:
encoder
decoder,
use
attention
blocks,
combination
with
elements
architectures,
introducing
additional
features,
transfer
learning
approaches
small
sets
real
data.
Various
training
considered,
which
best
values
metrics
achieved
literature
given
(similarity
coefficient
Dice,
intersection
over
union
IoU,
overall
accuracy
some
others).
A
summary
table
provided
indicating
types
analyzed
pathologies
detected
on
Promising
directions
further
to
improve
quality
segmentation
problems
outlined.
can
be
useful
determining
set
tools
identifying
diseases,
primarily
cancers.
presented
algorithms
basis
professional
intelligent
assistants.
Segmenting
organs
in
CT
scan
images
is
a
necessary
process
for
multiple
downstream
medical
image
analysis
tasks.
Currently,
manual
segmentation
by
radiologists
prevalent,
especially
like
the
pancreas,
which
requires
high
level
of
domain
expertise
reliable
due
to
factors
small
organ
size,
occlusion,
and
varying
shapes.
When
resorting
automated
pancreas
segmentation,
these
translate
limited
labeled
data
train
effective
models.
Consequently,
performance
contemporary
models
still
not
within
acceptable
ranges.
To
improve
that,
we
propose
M3BUNet,
fusion
MobileNet
U-Net
neural
networks,
equipped
with
novel
Mean-Max
(MM)
attention
that
operates
two
stages
gradually
segment
from
coarse
fine
mask
guidance
object
detection.
This
approach
empowers
network
surpass
achieved
similar
architectures
achieve
results
are
on
par
complex
state-of-the-art
methods,
all
while
maintaining
low
parameter
count.
Additionally,
introduce
external
contour
as
preprocessing
step
stage
assist
through
standardization.
For
stage,
found
applying
wavelet
decomposition
filter
create
multi-input
enhances
performance.
We
extensively
evaluate
our
NIH
MSD
datasets.
Our
demonstrates
improvement,
achieving
an
average
DSC
value
up
89.53±1.82
IOU
score
81.16±0.03
NIH,
88.60
±1.48
79.90±2.19
MSD.
