HaN-Seg: The head and neck organ-at-risk CT and MR segmentation challenge
Radiotherapy and Oncology,
Год журнала:
2024,
Номер
198, С. 110410 - 110410
Опубликована: Июнь 24, 2024
To
promote
the
development
of
auto-segmentation
methods
for
head
and
neck
(HaN)
radiation
treatment
(RT)
planning
that
exploit
information
computed
tomography
(CT)
magnetic
resonance
(MR)
imaging
modalities,
we
organized
HaN-Seg:
The
Head
Neck
Organ-at-Risk
CT
MR
Segmentation
Challenge.
challenge
task
was
to
automatically
segment
30
organs-at-risk
(OARs)
HaN
region
in
14
withheld
test
cases
given
availability
42
publicly
available
training
cases.
Each
case
consisted
one
contrast-enhanced
T1-weighted
image
same
patient,
with
up
corresponding
reference
OAR
delineation
masks.
performance
evaluated
terms
Dice
similarity
coefficient
(DSC)
95-percentile
Hausdorff
distance
(HD95),
statistical
ranking
applied
each
metric
by
pairwise
comparison
submitted
using
Wilcoxon
signed-rank
test.
While
23
teams
registered
challenge,
only
seven
their
final
phase.
top-performing
team
achieved
a
DSC
76.9
%
HD95
3.5
mm.
All
participating
utilized
architectures
based
on
U-Net,
winning
leveraging
rigid
registration
combined
network
entry-level
concatenation
both
modalities.
This
simulated
real-world
clinical
scenario
providing
non-registered
images
varying
fields-of-view
voxel
sizes.
Remarkably,
segmentation
surpassing
inter-observer
agreement
dataset.
These
results
set
benchmark
future
research
this
dataset
paired
multi-modal
general.
Язык: Английский
A Method for Sensitivity Analysis of Automatic Contouring Algorithms Across Different MRI Contrast Weightings Using SyntheticMR
medRxiv (Cold Spring Harbor Laboratory),
Год журнала:
2025,
Номер
unknown
Опубликована: Янв. 12, 2025
Currently,
a
majority
of
institution-specific
automatic
MRI-based
contouring
algorithms
are
trained,
tested,
and
validated
on
one
contrast
weighting
(i.e.,
T2-weighted),
however
their
actual
performance
within
this
across
different
repetition
times,
TR,
echo
TE)
is
under-investigated
poorly
understood.
As
result,
external
institutions
with
scan
protocols
for
the
same
may
experience
sub-optimal
performance.
The
purpose
study
was
to
develop
method
evaluate
robustness
varying
MRI
weightings.
One
healthy
volunteer
patient
scanned
using
SyntheticMR
MR-Simulation
device.
parotid
submandibular
glands
in
these
subjects
were
contoured
an
algorithm
trained
T2-weighted
MRIs.
For
ground
truth
manual
contours,
two
radiation
oncology
residents
pre-resident
physician
recruited
STAPLE
consensus
determined.
A
total
216
TR
TE
combinations
simulated
T1-,
T2-,
PD-weighted
ranges
SyntheticMR's
post-processing
software,
SyMRI.
Comparisons
between
contours
determined
Dice
similarity
coefficient
(DSC)
95
th
percentile
Hausdorff
distance
(HD95).
Notable
differences
model's
seen
contrast-weighted
range,
even
range.
Further,
some
models
performed
as
well
or
better
subsets
T1-weighted
range
saw
worst
discrepancy
DSC
HD95
exceeded
0.2
3.66
mm,
respectively,
structures.
In
region
where
model
100%,
40%,
24%,
57%
left
parotid,
right
submandibular,
gland,
interobserver
variability.
This
demonstrates
variable
combinations.
methodology
could
be
applied
future
studies
evaluating
sensitivity,
out
distribution
detection
ability,
drift.
Язык: Английский
A general model for head and neck auto‐segmentation with patient pre‐treatment imaging during adaptive radiation therapy
Medical Physics,
Год журнала:
2025,
Номер
unknown
Опубликована: Март 7, 2025
During
head
and
neck
(HN)
radiation
therapy,
patients
may
undergo
anatomical
changes
due
to
tumor
shrinkage
or
weight
loss.
For
these
patients,
adaptive
therapy
(ART)
is
required
correct
treatment
plans
ensure
that
the
prescribed
dose
delivered
while
minimizing
surrounding
organs-at-risk
(OARs).
Patient
pre-treatment
images
segmentation
labels
are
always
available
during
ART
be
incorporated
into
deep
learning
(DL)
auto-segmentation
models
improve
performance
on
mid-treatment
images.
Existing
DL
methods
typically
incorporate
data
training.
In
this
work,
we
investigated
whether
including
at
inference
time
would
affect
model
performance,
as
inference-time
inclusion
eliminate
requirement
for
costly
retraining
new
patient
cohorts.
We
developed
a
general
(GAM)
included
through
additional
input
channels.
compared
GAM
with
patient-specific
(PSM),
which
training,
reference
(RM),
did
not
include
data,
rigid
image
registration
(RIR)
method.
Models
were
using
large
dataset
of
pre-
computed
tomography
(primary
gross
volume
[GTVp]
16
OARs)
110
who
underwent
HN
cancer.
The
showed
improved
over
PSM
RM
several
structures,
largest
differences
in
dice
similarity
coefficient
difficult-to-segment
structures:
GTVp
(RM:
0.17,
PSM:
0.36,
GAM:
0.61,
RR:
0.65)
left/right
brachial
plexus
0.38/0.35,
0.43/0.43,
0.49/0.49,
0.36/0.38).
attained
similar
RR
all
structures
except
brainstem
(GAM:
0.82,
0.74),
mandible
0.88,
0.68),
spinal
cord
0.76,
0.51),
performed
higher.
can
ART,
particular
high
variability
low
contrast.
Including
give
improvements
standard
OARs,
eliminating
need
However,
provides
most
OARs.
Язык: Английский
A method for sensitivity analysis of automatic contouring algorithms across different contrast weightings using synthetic magnetic resonance imaging
Physics and Imaging in Radiation Oncology,
Год журнала:
2025,
Номер
unknown, С. 100790 - 100790
Опубликована: Май 1, 2025
Язык: Английский