Scientific Reports,
Год журнала:
2024,
Номер
14(1)
Опубликована: Дек. 30, 2024
Medical
datasets
are
vital
for
advancing
Artificial
Intelligence
(AI)
in
healthcare.
Yet
biases
these
on
which
deep-learning
models
trained
can
compromise
reliability.
This
study
investigates
stemming
from
dataset-creation
practices.
Drawing
existing
guidelines,
we
first
developed
a
BEAMRAD
tool
to
assess
the
documentation
of
public
Magnetic
Resonance
Imaging
(MRI);
Color
Fundus
Photography
(CFP),
and
Electrocardiogram
(ECG)
datasets.
In
doing
so,
provide
an
overview
that
may
emerge
due
inadequate
dataset
documentation.
Second,
examine
current
state
medical
images
signal
data.
Our
research
reveals
there
is
substantial
variance
image
datasets,
even
though
guidelines
have
been
imaging.
indicates
subject
individual
discretionary
decisions.
Furthermore,
find
aspects
such
as
hardware
data
acquisition
details
commonly
documented,
while
information
regarding
annotation
practices,
error
quantification,
or
limitations
not
consistently
reported.
risks
having
considerable
implications
abilities
users
detect
potential
sources
bias
through
respective
develop
reliable
robust
be
adapted
clinical
practice.
Biomedical Physics & Engineering Express,
Год журнала:
2025,
Номер
11(2), С. 025007 - 025007
Опубликована: Янв. 17, 2025
Abstract
Background
and
Purpose
:
This
study
aimed
to
develop
evaluate
an
efficient
method
automatically
segment
T1-
T2-weighted
brain
magnetic
resonance
imaging
(MRI)
images.
We
specifically
compared
the
segmentation
performance
of
individual
convolutional
neural
network
(CNN)
models
against
ensemble
approach
advance
accuracy
MRI-guided
radiotherapy
(RT)
planning.
Materials
Methods
.
The
evaluation
was
conducted
on
a
private
clinical
dataset
publicly
available
(HaN-Seg).
Anonymized
MRI
data
from
55
cancer
patients,
including
T1-weighted,
T1-weighted
with
contrast,
images,
were
used
in
dataset.
employed
EDL
strategy
that
integrated
five
independently
trained
2D
networks,
each
tailored
for
precise
tumors
organs
at
risk
(OARs)
scans.
Class
probabilities
obtained
by
averaging
final
layer
activations
(Softmax
outputs)
networks
using
weighted-average
method,
which
then
converted
into
discrete
labels.
Segmentation
evaluated
Dice
similarity
coefficient
(DSC)
Hausdorff
distance
95%
(HD95).
model
also
tested
HaN-Seg
public
comparison.
Results
demonstrated
superior
both
datasets.
For
dataset,
achieved
average
DSC
0.7
±
0.2
HD95
4.5
2.5
mm
across
all
segmentations,
significantly
outperforming
yielded
values
≤0.6
≥14
mm.
Similar
improvements
observed
Conclusions
Our
shows
consistently
outperforms
CNN
datasets,
demonstrating
potential
learning
enhance
accuracy.
These
findings
underscore
value
applications,
particularly
RT
Računalniška
tomografija
(CT)
je
slikovna
preiskava,
ki
se
v
klinični
praksi
standardno
zajame
okviru
načrtovanje
radioterapije.
V
primeru
raka
območju
glave
in
vratu
(HaN)
pogosto
tudi
magnetno
resonančne
(MR)
slike
za
natančnejše
orisovanje
tumorjev
kritičnih
organov.
zadnjem
času
vse
bolj
uveljavlja
radioterapija
na
podlagi
MR-samostojnega
pristopa,
odstrani
potrebo
po
zajemu
CT
slik
s
tem
izpostavljenost
ionizirajočemu
sevanju,
vendar
pa
zahteva
rešitev
generiranje
sintetičnih
MR
.
Nedavne
študije
kažejo,
da
difuzijski
modeli
nudijo
realistično
z
natančnimi
anatomskimi
podrobnostmi
manj
artefakti
kot
generativne
nasprotniške
mreže.
tej
študiji
smo
razvili
model
pretvorbo
sintetične
HaN
področje.
Naš
pristop,
ovrednoten
zbirki
podatkov
HaN-Seg,
vključuje
pare
istih
bolnikov,
doseže
indeks
strukturne
podobnosti
92,2
%,
vršno
razmerje
signal-šum
33,1
dB
ter
povprečno
absolutno
napako
35,3
HU.
Model
dodatno
ovrednotimo
segmentacijo
Rezultati
potrjujejo
potencial
uporabe
difuzijskih
modelov
pri
načrtovanju
