926 / 1082
S910
ESTRO 36 2017
_______________________________________________________________________________________________
We based our retrospective study on a total of
N
=122 high-
risk prostate patients treated with radiotherapy, with
inclusion criteria to have a pre-treatment PSA<60 µg/L and
biopsies analyzed a Uppsala University Hospital. The 5-
year local tumor control probability was estimated with
Kaplan Meier analysis to TCP
obs
=94.7% (CI 86.4-98.0%). The
PSA inclusion condition was used to exclude patients with
possible pre-treatment spread. The homogeneous
treatment dose
D
h
was estimated to 91.6 Gy EQD
2
based
on α/β=1.93 for the given proton boost (20Gy in 4
fractions, RBE=1.1) and photon dose (50 Gy in 25
fractions). All patients underwent androgen deprivation
therapy. We parameterized the populations dose-response
TCP
pop
(
D
) with a logistic function with the parameter
γ
50
=2.01 and
D
50
chosen so that TCP
pop
(
D
h
)= TCP
obs
. The
patients’ biopsy statements were used to construct
simulated prostates with voxelized distributions of
Gleason scores
G
varying per voxel.
Voxel specific dose-response functions TCP
vox
(
D
,
G
) were
derived with the logistic parameters γ
50,eff
and
D
50
(
G
) set
so that the average TCP
pat
for all patients equals TCP
obs
at
D
h
, and the average slope for the patients TCP
pat
equals
the slope for TCP
pop
(
D
) at
D
h
. Hence, the voxel specific
dose-response functions are be described by
TCP
vox
(
D
,
G
)=1/(1+(
D
50
(
G
)/
D
)
4γ50,eff
),
where
D
50
(
G
) and γ
50,eff
, for
D
=
D
h
, reconstructs
TCP
vox
(
D
h
,
G
<6)=
C
and
TCP
vox
(
D
h
,
G
≥6)=
C
-
k
×(
G
-6).
For
G
<6 TCP
vox
was set to not vary with Gleason scores
since ADC-MRI likely not distinguish
G
<6 from normal
tissue. We used 3 different values of
C
, a high value
C
high
=1
resulting in zero desired dose for
G
<6 voxels, a low value
C
min
resulting in a homogeneous dose distribution (
k
=0),
and an intermediate
C
im
for a certain minimum dose.
ADC images for a high-risk patient were translated into a
3D-map of Gleason scores based on results published by
Turkbey et al. We used the above functions for dose
painting to minimize the average dose while keeping the
TCP
pat
equal to that for a homogeneous dose of
D
h
.
Results
For the
C
high
scenario the average dose decreased by 9 Gy
(max dose 98 Gy). For the intermediate
C
im
scenario the
average dose decreased by 2 Gy with doses in the range of
74 to 98 Gy. Fig. 1 shows resulting Gleason score to TCP
mappings normalized for a 50cc prostate while Fig. 2
shows a dose painted prostate for the
C
im
scenario.
Fig 1.
TCP vs Gleason scores comprising a 50cc prostate
volume and corresponding dose-response functions for the
intermediate
C
im
scenario.
Fig 2.
ADC in prostate and dose painted prostate for
C
im
scenario.
Conclusion
Gleason driven dose painting for prostate cancer using
ADC-MRI is feasible to reduce the average dose. The
reduction in dose is strongly dependent on the minimum
dose assigned to voxels with
G
<6.
EP-1690 Validating the robustness of PET features in a
phantom in a multicenter setting
T. Konert
1
, M. La Fontaine
2
, S. Van Kranen
2
, W. Vogel
1
, J.
Van de Kamer
2
, J.J. Sonke
2
1
Netherlands Cancer Institute Antoni van Leeuwenhoek
Hospital, Nuclear Medicine, Amsterdam, The
Netherlands
2
Netherlands Cancer Institute Antoni van Leeuwenhoek
Hospital, Radiation Oncology, Amsterdam, The
Netherlands
Purpose or Objective
PET features may have prognostic or predictive value and
could therefore assist treatment decisions. However, PET
features are sensitive to differences in data collection,
reconstruction settings, and image analysis. It is
insufficiently known which features are least affected by
these differences, especially in a multicenter setting.
Therefore, this study investigates the robustness of PET
features in a phantom after repeated measurements
(repeatability), due to varying scanner type
(reproducibility) and their dependence on binning method
and SUV activity.
Material and Methods
PET scans from a NEMA image quality phantom were used
for assessment of PET feature robustness. Scans were
acquired on a Philips, a Siemens and a GE scanner from
three medical centers (see figure 1 and table 1 for more
details). Per sphere, a VOI was created by applying a
threshold of 40% of the SUV
max
. Per VOI, 10 first order
statistics and 10 textural features, often reported in
literature, were extracted. Two common implementations
of image pre-processing, before feature extraction, were
compared: using a fixed bin size (SUV = 1) versus a number
of fixed bins (64 bins). To examine the feature
repeatability, measurements were repeated two or three
times on the same scanner. The reproducibility was
assessed in images by comparing all scanners. The degree
of variation was calculated per VOI with the coefficient of
repeatability (1.96 x SD/mean), normalized to a
percentage (CR
%
). Features were seen as robust with a CR
< 30%, matching the level of uncertainty found in response
of PERCIST criteria. Wilcoxon signed rank tests were used
to estimate the significance of differences due to binning
method and p-values ≤ 0.05 were considered significant.
Results
For an overview of the results, see Table 1. The CR
%
of
SUV
max
in all scans depended on sphere volume, and
ranged from 1.1% (largest sphere) to 15.2% (smallest
sphere). In the repeatability study, 9 out of 10 PET
features were robust with 64 bins in more than one
scanner, and significantly higher (p < 0.05) when
compared to using a fixed bin size, where 7 out of 10 PET
features were robust. Reproducibility was achieved in 3
out of 10 PET features when 64 bins were used. PET
features were not reproducible when using a fixed bin
size.
Dissimilarity (CR
%
: 6.3-24.9), homogeneity 1 (CR
%
:
16.9-22.5), and inertia (CR
%
: 10.2-22.5) were robust to
binning method, scanner type, and SUV activity.
Coarseness, contrast, busyness, energy, correlation were
not robust (CR
%
> 30%).