2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC),
Год журнала:
2022,
Номер
unknown, С. 1 - 3
Опубликована: Ноя. 5, 2022
In
this
study,
we
performed
two
experiments
to
explore
radiomic
features
and
multi-modality
medical
image
fusion
(IF).
the
first
experiment,
investigated
performance
of
multiple
IF
algorithms
for
reflecting
from
both
PET
CT
modalities
in
a
single
scan.
second
if
can
serve
as
an
objective
quality
assessment
(QA)
metric.
Experiments
were
on
PET/CT
images
328
histologically
proven
head
neck
tumors
segmented
by
experienced
radiologist.
The
iterative
metal
artifact
reduction
(iMAR)
algorithm
was
applied
images,
their
Hounsfield
Unit
range
clipped
[-500,500],
then
all
resized
isotropic
voxel
size
1
×
mm3,
quantized
normalized
integer
values
[0,
255].
To
have
comprehensive
analysis,
fused
using
11
different
covering
almost
categories,
13
metrics
categories
calculated
each
fusion.
Ninety-four
textural
extracted
regarding
Image
Biomarker
Standardization
Initiative
(IBSI)
guidelines.
For
Spearman
correlation
feature
between
its
fused-set
CT-
PET-sets,
coefficients
higher
than
0.7
considered
significant.
A
"preserved"
it
correlated
with
peer
sets.
QA
significant
0.7.
Among
methods
GFF
(guided
filtering
fusion)
FPDE
(fourth-order
partial
differential
equation)
had
best
results
conserving
22
19
features,
respectively,
showing
ability
reflect
maximum
information
GLCM
least
preserved
across
fusions.
Several
Radiomic
showed
peak
signal-to-noise
ratio
(PSNR)
root
mean
square
error
(RMSE)
metric
methods,
while
no
entropy
(EN),
SSIM
(structural
similarity
index
measure),
AG
(average
gradient),
EI
(edge
intensity),
SD
(standard
deviation),
SF
(spatial
frequency),
Qcv
(Chen-Varshney
metric).
Background
To
assess
the
feasibility
of
a
machine
learning
(ML)
approach
using
radiomics
features
perfusion
defects
on
rest
myocardial
imaging
(MPI)
to
detect
presence
hibernating
myocardium.
Methodology
Data
patients
who
underwent
99mTc-sestamibi
MPI
and
18F-FDG
PET/CT
for
viability
assessment
were
retrieved.
Rest
data
processed
ECToolbox,
polar
maps
saved
NFile
PMap
tool.
The
reference
standard
defining
myocardium
was
mismatched
perfusion-metabolism
defect
with
impaired
contractility
at
rest.
Perfusion
delineated
regions
interest
(ROIs)
after
spatial
resampling
intensity
discretization.
Replicable
random
sampling
allocated
80%
(257)
from
January
2017
September
2022
training
set
remaining
20%
(64)
validation
set.
An
independent
dataset
29
consecutive
October
2023
used
as
testing
model
evaluation.
One
hundred
ten
first
second-order
texture
extracted
each
ROI.
After
feature
normalization
imputation,
14
best-ranked
selected
multistep
selection
process
including
Logistic
Regression
Fast
Correlation-Based
Filter.
Thirteen
supervised
ML
algorithms
trained
stratified
five-fold
cross-validation
validated
Log
Loss
<0.688
<0.672
in
steps
evaluated
Performance
matrices
assessed
included
area
under
curve
(AUC),
classification
accuracy
(CA),
F1
score,
precision,
recall,
specificity.
provide
transparency
interpretability,
SHapley
Additive
exPlanations
(SHAP)
values
depicted
beeswarm
plots.
Results
Two
thirty-nine
(214
males;
mean
age
56
±
11
years)
enrolled
study.
There
371
(321
sets;
50
set).
Based
standard,
168
had
(139
On
cross-validation,
six
AUC
>0.800.
validation,
10
value
<0.672,
among
which
evaluation
models
unseen
set,
nine
>0.800
Gradient
Boosting
Random
Forest
(xgboost)
[GB
RF
(xgboost)]
achieving
highest
0.860
could
21/29
(72.4%)
precision
87.5%
(21/24),
specificity
85.7%
(18/21),
CA
78.0%
(39/50)
Score
0.792.
Four
clear
pattern
interpretability
based
SHAP
These
GB
(xgboost),
(scikit-learn),
Forest.
Conclusion
Our
study
demonstrates
potential
detecting
images.
This
proof-of-concept
underscores
notion
that
capture
nuanced
information
beyond
what
is
perceptible
human
eye,
offering
promising
avenues
improved
assessment.
2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC),
Год журнала:
2022,
Номер
unknown, С. 1 - 3
Опубликована: Ноя. 5, 2022
Glioblastoma
multiforme
(GBM)
is
regarded
as
the
most
prevalent
primary
tumor
of
central
nervous
system
in
brain.
However,
due
to
lack
information,
it
too
hard
understand
underlying
progression
patterns
and
prognosis
patients.
In
this
study,
we
evaluated
overall
survival
predictive
(time
event
analysis)
power
radiomic
features
extracted
from
MRI,
along
with
help
feature
selection
(FS)
machine
learning
(ML)
algorithms.
The
MR
images
119
patients
their
status
were
obtained.
data
randomly
split
into
70%
30%,
indicating
training
testing
datasets,
respectively.
Twelve
preprocessing
methods
(e.g.,
bin
discretization,
Laplacian
Gaussian,
wavelet
transform),
5
FS
Boruta,
Cindex,
Random
Survival
Forest),
7
ML
Glmnet,
CoxBoost)
algorithms
recruited
form
a
total
420
models.
models
C-index
method
showed
more
decent
results
than
others.
highest-achieving
model
(C-index
=
0.72)
was
combination
LOG
sigma
1
mm
preprocessing,
selector,
Coxph
algorithm.
Our
findings
represent
be
utilized
prediction
glioblastoma
prognostication
general.
2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC),
Год журнала:
2022,
Номер
unknown, С. 1 - 4
Опубликована: Ноя. 5, 2022
Heart
disease
is
one
of
the
leading
causes
death
worldwide.
Among
various
methods
used
to
assess
heart
function,
MPI
SPECT
method
a
valuable
and
non-invasive
that
brings
high-quality
images
with
low
radiation
exposure.
Radiomics
has
been
developed
extract
quantitative
features
from
medical
images.
These
can
be
predict
diagnosis
treatment
in
science.
To
use
these
clinic,
they
need
reliable;
other
words,
repeatable
reproducible.
Various
factors,
including
different
reconstructions,
affect
repeatability
reproducibility
radiomic
features.
Twenty
patients
who
underwent
stress
rest
were
this
study.
As
result,
40
existing
reconstructed
15
modes.
Finally,
600
unique
reconstructions
obtained,
segmentation
process
was
conducted
using
3D-Slicer
program.
Feature
extraction
done
LIFEx,
finally,
coefficient
variance
(COV)
check
reproducibility.
The
most
robust
FO_Kurtosis,
GLCM_Entropy_log10,
GLCM_Entropy_log2,
GLRLM_SRE,
GLRLM_LRE,
GLRLM_RP,
GLZLM_SZE,
GLZLM_HGZE.
change
order
reconstruction
parameter
only
case
caused
least
feature
variation.
This
study
planned
reliability
over
changes
parameters.
2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC),
Год журнала:
2022,
Номер
unknown, С. 1 - 3
Опубликована: Ноя. 5, 2022
A
body
of
literature
has
reported
the
promising
performance
deep
learning
models
when
applied
to
mammograms
for
different
clinical
tasks.
However,
a
major
pitfall
is
they
are
bearing
high-density
breasts
since
overlapping
high-dense
tissue
can
cover
lesion
and
make
diagnosing
interpreting
it
difficult.
Thus,
analysis
dedicated
trained
low-
be
great
importance.
In
this
study,
we
aimed
develop
deep-learning
classifying
breast
masses
on
into
benign
malignant
cases.
Curated
Breast
Imaging
Subset
Digital
Database
Screening
Mammography
(CBIS-DDSM)
dataset,
including
mammograms,
cropped
masses,
pathologic
diagnoses,
were
adopted
study.
For
models,
dataset
was
split
low-(BI-RADS
density1
2)
(BI-RADS
density3
4)
groups.
Contrast-limited
adaptive
histogram
equalization
(CLAHE)
enhance
contrast
images,
then
all
images
resized
255
×
matrix
size
normalized.
modified
DenseNet-201
neural
network
with
rate
starting
at
0.0001
decreased
in
piecewise
manner
every
epoch
RMSProp
optimizer.
Fifteen
percent
training
data
excluded
validation,
continued
100
epochs.
Data
augmentation,
rotation,
flipping,
scaling,
implemented
prevent
overfitting.
The
model
evaluated
using
test
set.
Accuracy
(ACC),
area
under
receiver
operating
characteristic
curve
(AUC),
sensitivity
(SEN),
specificity
(SPE)
general
0.720,
0.771,
0.732,
0.701,
Low-density
0.788,
0.818,
0.824,
0.742,
High-density
0.712,
0.621,
0.962,
0.180,
respectively.
Our
study
highlights
importance
developing
tailored
nature
improve
overall
accuracy.
We
also
suggest
performing
specific
preprocessing
breasts,
such
as
region-wise
enhancement
regions
high-intensity
values
(with
gamma
filter,
etc.).
2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC),
Год журнала:
2022,
Номер
unknown, С. 1 - 3
Опубликована: Ноя. 5, 2022
In
this
study,
we
performed
two
experiments
to
explore
radiomic
features
and
multi-modality
medical
image
fusion
(IF).
the
first
experiment,
investigated
performance
of
multiple
IF
algorithms
for
reflecting
from
both
PET
CT
modalities
in
a
single
scan.
second
if
can
serve
as
an
objective
quality
assessment
(QA)
metric.
Experiments
were
on
PET/CT
images
328
histologically
proven
head
neck
tumors
segmented
by
experienced
radiologist.
The
iterative
metal
artifact
reduction
(iMAR)
algorithm
was
applied
images,
their
Hounsfield
Unit
range
clipped
[-500,500],
then
all
resized
isotropic
voxel
size
1
×
mm3,
quantized
normalized
integer
values
[0,
255].
To
have
comprehensive
analysis,
fused
using
11
different
covering
almost
categories,
13
metrics
categories
calculated
each
fusion.
Ninety-four
textural
extracted
regarding
Image
Biomarker
Standardization
Initiative
(IBSI)
guidelines.
For
Spearman
correlation
feature
between
its
fused-set
CT-
PET-sets,
coefficients
higher
than
0.7
considered
significant.
A
"preserved"
it
correlated
with
peer
sets.
QA
significant
0.7.
Among
methods
GFF
(guided
filtering
fusion)
FPDE
(fourth-order
partial
differential
equation)
had
best
results
conserving
22
19
features,
respectively,
showing
ability
reflect
maximum
information
GLCM
least
preserved
across
fusions.
Several
Radiomic
showed
peak
signal-to-noise
ratio
(PSNR)
root
mean
square
error
(RMSE)
metric
methods,
while
no
entropy
(EN),
SSIM
(structural
similarity
index
measure),
AG
(average
gradient),
EI
(edge
intensity),
SD
(standard
deviation),
SF
(spatial
frequency),
Qcv
(Chen-Varshney
metric).