928 / 1082
S912
ESTRO 36 2017
_______________________________________________________________________________________________
dose rate image-guided uterovaginal brachytherapy of 15
Gy.
The primary tumor was delineated on the PET images using
a fixed threshold (40% of SUV
max
) within a manually drawn
volume of interest (VOI), called VOI-T. A 73 mL sphere was
drawn in the healthy liver considered as homogeneous
(VOI-L). For each VOI, 5 conventional indices (SUV
mean
,
SUV
max
, SUV
peak
in a 1 mL sphere, metabolic volume, tumor
lesion glycolysis) and 6 3D textural indices were calculated
after resampling the VOI SUV between 0 and 40 using 128
gray levels: Homogeneity and Entropy from the Gray-Level
Co-occurrence Matrix, Short-Run Emphasis (SRE), Long-
Run Emphasis (LRE), Low Gray-level Zone Emphasis (LGZE)
and High Gray-level Zone Emphasis (HGZE).
Wilcoxon’s tests were performed between G1 and G2 in
VOI-L to determine to what extent technological
differences and image properties influence radiomic
feature values. A stepwise model selection using the
Akaike Information Criterion was applied to determine the
best 4-feature signature for local recurrence prediction in
both groups, used successively for training and validation.
Delong’s tests between AUC were performed to evaluate
if the signature was more powerful than SUV
max
only.
Results
SUV
max
, SUV
peak
, Homogeneity and SRE computed in VOI-L
were significantly different between the two devices
(p<0.05). These p-values suggested that data coming from
the two PET devices can therefore not be gathered.
In G1, the best 4-feature signature was a combination of
Entropy, SUV
mean
, SUV
max
and SRE (AUC=0.77) and in G2, a
combination of SUV
peak
, Homogeneity, LGZE, HGZE
(AUC=0.86). G2 signature was validated in G1 with
AUC=0.76 and was significantly more powerful than SUV
max
according to Delong’s test (p=0.02). G1 signature was not
validated in G2, yielding to an AUC less than that obtained
with SUV
max
only.
Conclusion
Some conventional and textural features are strongly
dependent on the PET device and acquisition parameters
such as voxel size. A robustness analysis should be
performed before each multi-centric radiomic study, to
evaluate the possibility of gathering data from different
devices. Multivariate analysis showed that radiomic
features can predict LACC local recurrence with a better
accuracy than SUV
max
for recent PET devices. The creation
of an external validation cohort is in progress to confirm
the results.
EP-1693 Functional MRI to individualize PTV margins
to seminal vesicles with suspected cancer involvement
S. Damkjaer
1
, J. Thomsen
1
, S. Petersen
1
, J. Bangsgaard
1
,
M. Aznar
2
, I. Vogelius
1
, P. Petersen
1
1
Rigshospitalet, Department of Oncology- Section for
Radiotherapy - 3994, København, Denmark
2
University of Oxford, Clinical Trial Service Unit- Richard
Doll Building, Oxford, United Kingdom
Purpose or Objective
For external beam radiotherapy of prostate cancer
patients, the information from pre-treatment MRIs can
give patient specific and visual evaluation of suspected
pathologically involved volumes in the seminal vesicles
(SV) as an important addition to probability based
nomograms [1]. We investigate the impact of