International Journal of Imaging Systems and Technology,
Год журнала:
2024,
Номер
34(3)
Опубликована: Апрель 30, 2024
Abstract
Colorectal
cancer
is
a
common
gastrointestinal
malignancy.
Early
screening
and
segmentation
of
colorectal
polyps
are
great
clinical
significance.
Colonoscopy
the
most
effective
method
to
detect
polyps,
but
some
may
be
missed
in
detection
process.
On
this
basis,
use
computer‐aided
diagnosis
technology
particularly
important
for
polyp
segmentation.
To
improve
rate
intestinal
under
colonoscopy,
network
(MobileRaNet)
based
on
lightweight
model
reverse
attention
(RA)
mechanism
was
proposed
accurately
segment
colonoscopy
images.
The
coordinated
module
used
MobileNetV3
make
it
backbone
(CaNet).
Second,
part
output
high‐level
feature
from
passed
into
parallel
axial
receptive
field
(PA_RFB)
extract
global
dependency
representation
without
losing
details.
Third,
map
generated
combined
as
initial
boot
area
subsequent
components.
Finally,
RA
mine
target
region
boundary
clues
accuracy.
verify
effectiveness
performance
algorithm,
five
challenging
datasets,
including
CVC‐ColonDB,
CVC‐300,
Kvasir,
paper.
In
six
indexes,
MeanDice,
MeanIoU,
MAE,
compared
with
seven
typical
models
such
PraNet
TransUnet,
accuracy,
FLOPs,
parameters,
FPS
were
compared.
experimental
results
show
that
MobileRaNet
paper
has
improved
datasets
varying
degrees,
especially
MeanDice
MeanIOU
indexes
Kvasir
dataset
reach
91.2%
85.6%,
which
are,
respectively,
increased
by
1.4%
1.6%
PraNet.
Compared
PraNet,
FLOPs
parameters
decreased
83.3%
76.7%,
respectively.
Diagnostics,
Год журнала:
2022,
Номер
12(10), С. 2316 - 2316
Опубликована: Сен. 26, 2022
The
first
step
in
the
diagnosis
of
gastric
abnormalities
is
detection
various
human
gastrointestinal
tract.
Manual
examination
endoscopy
images
relies
on
a
medical
practitioner’s
expertise
to
identify
inflammatory
regions
inner
surface
length
alimentary
canal
and
large
volume
obtained
from
endoscopic
procedures
make
traditional
methods
time
consuming
laborious.
Recently,
deep
learning
architectures
have
achieved
better
results
classification
images.
However,
visual
similarities
between
different
portions
tract
pose
challenge
for
effective
disease
detection.
This
work
proposes
novel
system
by
focusing
feature
mining
through
convolutional
neural
networks
(CNN).
model
presented
built
combining
state-of-the-art
architecture
(i.e.,
EfficientNet
B0)
with
custom-built
CNN
named
Effimix.
proposed
Effimix
employs
combination
squeeze
excitation
layers
self-normalising
activation
precise
diseases.
Experimental
observations
HyperKvasir
dataset
confirm
effectiveness
yields
an
accuracy
97.99%,
F1
score,
precision,
recall
97%,
98%,
respectively,
which
significantly
higher
compared
existing
works.
2021 IEEE/CVF International Conference on Computer Vision (ICCV),
Год журнала:
2023,
Номер
unknown, С. 21617 - 21627
Опубликована: Окт. 1, 2023
Zero-shot
point
cloud
segmentation
aims
to
make
deep
models
capable
of
recognizing
novel
objects
in
that
are
unseen
the
training
phase.
Recent
trends
favor
pipeline
which
transfers
knowledge
from
seen
classes
with
labels
without
labels.
They
typically
align
visual
features
semantic
obtained
word
embedding
by
supervision
classes'
annotations.
However,
contains
limited
information
fully
match
features.
In
fact,
rich
appearance
images
is
a
natural
complement
textureless
cloud,
not
well
explored
previous
literature.
Motivated
this,
we
propose
multi-modal
zero-shot
learning
method
better
utilize
complementary
clouds
and
for
more
accurate
visual-semantic
alignment.
Extensive
experiments
performed
two
popular
benchmarks,
i.e.,
SemanticKITTI
nuScenes,
our
outperforms
current
SOTA
methods
52%
49%
improvement
on
average
class
mIoU,
respectively.
This
paper
presents
a
comprehensive
study
into
the
field
of
polyp
segmentation
using
novel
use
two
renowned
deep
learning
architectures:
U-Net
and
ENet.
The
paper,
which
focuses
on
essential
issue
identifying
areas
in
medical
imaging,
defines
distinct
application
both
ENet
algorithms,
followed
by
careful
comparison
their
various
results.
examines
each
algorithm's
unique
strengths,
limitations,
overall
effectiveness
analyzing
data
acquired
from
method.
Essentially,
this
provides
thorough
examination
utilizing
ENet,
opening
door
for
improved
image
analysis
informed
decision-making
clinical
terms.
2021 IEEE/CVF International Conference on Computer Vision (ICCV),
Год журнала:
2023,
Номер
unknown
Опубликована: Окт. 1, 2023
Point-,
voxel-,
and
range-views
are
three
representative
forms
of
point
clouds.
All
them
have
accurate
3D
measurements
but
lack
color
texture
information.
RGB
images
a
natural
complement
to
these
cloud
views
fully
utilizing
the
comprehensive
information
benefits
more
robust
perceptions.
In
this
paper,
we
present
unified
multi-modal
LiDAR
segmentation
network,
termed
UniSeg,
which
leverages
cloud,
accomplishes
semantic
panoptic
simultaneously.
Specifically,
first
design
Learnable
cross-Modal
Association
(LMA)
module
automatically
fuse
voxel-view
range-view
features
with
image
features,
utilize
rich
calibration
errors.
Then,
enhanced
transformed
space,
where
further
fused
adaptively
by
cross-View
(LVA).
Notably,
UniSeg
achieves
promising
results
in
public
benchmarks,
i.e.,
SemanticKITTI,
nuScenes,
Waymo
Open
Dataset
(WOD);
it
ranks
1st
on
two
challenges
including
challenge
nuScenes
SemanticKITTI.
Besides,
construct
OpenPCSeg
codebase,
is
largest
most
outdoor
codebase.
It
contains
popular
algorithms
provides
reproducible
implementations.
The
codebase
will
be
made
publicly
available
at
https://github.com/PJLab-ADG/PCSeg.
