93 / 1082
S80
ESTRO 36 2017
_______________________________________________________________________________________________
presence of metal artefacts. To eliminate left-right bias,
each combination was shown twice.
Results
VMDE images reconstructed at energies in the range 60 to
70 keV showed improved CNR for all soft tissue regions
when compared to the standard CBCT. On average, the
reconstruction energy corresponding to the maximum CNR
improvement is 65.5 ± 2.4 keV. The increase in maximum
CNR varied from 29% to 78%.
The clinical observer comparison gave a series of rankings
for each image series for each patient (see table 1). Using
signed rank Wilcoxon comparison test, the observers found
the VMDE images at 65 keV preferable to the standard
CBCT image. The p-value was found to be < 0.01, where p
< 0.05 is considered significant. An estimate of inter-
observer variability test was done with Fleiss’ kappa and
found to be moderate with a κ-value of 0.47, where values
above 0.4 is considered acceptable and 1 is perfect
agreement. Occasionally, an observer ranked the 75 keV
reconstruction as the most preferable image while the
overall preferred image was the 65 keV reconstruction.
Except in the case of patient one where the standard CBCT
image was ranked the highest of all.
Conclusion
VMDE images can increase soft tissue contrast and improve
clinical image quality for image-guided radiotherapy
compared to the standard CBCT protocol.
OC-0160 Radiomics Features Harmonization for CT
and CBCT in Rectal Cancer
R. Luo
1
, J. Wang
1
, H. Zhong
1
, J. Gan
1
, P. Hu
1
, L. Shen
1
,
W. Hu
1
, Z. Zhang
1
1
Fudan University Cancer Hospital, Radiation Oncology,
Shanghai, China
Purpose or Objective
Inter-scanner variability can lead to degradation of
radiomics
model
quality.
Therefore,
feature
harmonization is necessary for consistent findings in
radiomics studies, especially for multi-institution studies.
The purpose of this study is
to establish harmonization
methods for CT and CBCT radiomics features in rectal
cancer, and to provide a harmonization evaluation
method.
Material and Methods
Three harmonization strategies were tested in this study,
including no correction, simple correction and phantom
based correction. 50 rectal cancer patients with both
planning CT images and positioning CBCT images before
the first fraction of treatment were collected for
harmonization performance evaluation. 203 features were
extracted from CT and CBCT images. For the phantom
based correction, a texture phantom comprised of 30
different materials was designed for features selection
and nonlinear functions generation for normalizing CT and
CBCT features.The Main workflow was shown in Figure 1.
Mixed datasets consisting of CT and CBCT features were
generated for harmonization performance evaluation
using cluster analysis. The harmonization performance
was evaluated by Chi-square testing between clustering
results and scanner machines, and the clustering
consistency with original CT feature. These tests were
repeated for 50 times with randomized sample
generation.
Figure 1. Main Workflow. Four steps of this study:(I)
Feature selection by features range comparison.
(II)Feature selection by spearman correlation test. (III)
Nonlinear mapping function generation using texture
phantom. (IV)Correction methods performance evaluation
on patients.
Results
41 of the 203 radiomics features were selected by range
comparison and spearman correlation test. Among 50
randomized sampling processes, all clustering (100%)
results without any correction showed high correlation
with imaging machine (p>0.05, χ^
2
test), while this
probability reduced to 0 % and 42% respectively when
simple correction or phantom based correction were
applied. Average accuracy and Kappa index increased
significantly (p<0.05, t-test), respectively to 0.71±0.07
and 0.42±0.12 for simple correction method and 0.68±0.06
and 0.36±0.14 for phantom based correction method, from
0.61±0.06 and 0.23±0.13 without any correction.
Table1. Performance evaluation result for different
harmonization strategies.
Conclusion
This is the first study focused on feature harmonization for
CT images. Two proposed correction methods, simple
correction and phantom based correction, were verified
to be feasible for CT and CBCT harmonization, which could
significantly improve the modeling consistency.
Proffered Papers: Novelties in image guidance
OC-0161 patient tolerance of stereotactic MR-guided
adaptive radiation therapy: an assessment using PRO’s