ESTRO 36 Abstract Book

936 / 1096

S920 ESTRO 36
_______________________________________________________________________________________________
Results
Each of the selected six radiomics features (1 texture and
5 first order statistics), which were normalized to be
comparable, showed higher predictive power compared to
FMISO TMRmax at the moment of predicting outcomes
univariately. AUC-ROC curves demonstrated that a model
created out of only six dominant CT-radiomics features
can discriminate groups better with respect to local
control in HNC using the logistic regression models (AUC =
0.904) than FMISO TMRmax (AUC = 0.800). Nevertheless, a
combination of FMISO-PET TMR max values and only two CT-
radiomics features (Small Zone Emphasis texture and
Minimum Grey Level first-order statistics) can reach an
AUC of 0.886 in our classification model.
Conclusion
CT radiomics proved to have better prognostic power with
respect to local control in HNC than FMISO-PET TMR max.
Nonetheless, a combination of TMR max and the two most
significant features of CT radiomics reaches high
prognostic power with fewer features to assess.
Consequently, analysing tumour heterogeneity using CT
radiomics features may have the power to determine
substitute measures of tumour hypoxia and might
therefore be used as a basis for personalized RT
adaptations in the future.
EP-1686 Diffusion weighted imaging for treatment
response prediction in advanced rectal cancer
H.D. Nissen1
1 Nissen Henrik D., Department of Oncology - Section for
Radiotherapy, Vejle, Denmark
Purpose or Objective
The standard treatment of locally advanced distal rectal
cancer is chemoradiotherapy (CRT) followed by surgery.
Based on pathologic examination of the surgery specimen,
a significant number of patients are found to be without
remaining tumor at the time of surgery. This has led to an
increasing interest in whether, for a select group of
patients, surgery can be replaced by a wait-and-see
strategy. Several recent studies [1, 2] have shown that this
is possible without compromising survival and with
significantly reduced comorbidities. A significant
challenge in this strategy is selecting the patients who are
candidates for this strategy.
We wish to examine whether diffusion weighted MRI (DWI)
can be used as an early biomarker for tumor response to
CRT.
Material and Methods
Here we present data from 25 patients treated for distal
T3 or T4 rectal tumors. Patients were treated with long
course CRT, including a brachytherapy boost to the tumor,
followed by surgery. Patients were DWI scanned before
start of CRT and again after 2 weeks of CRT. The DWI
sequence included 11 b-values from 0 to 1100. Regions of
interest (ROI) were drawn using an algorithm to locate
areas with atypically high signal on high b-value images.
ROI volumes where then analyzed using a radiomics
approach where 67 image features were extracted from
each volume. Tumor response to treatment was
determined by pathologists examining the surgical
specimen and scoring the tumor response using the
Mandard Tumor Regression Grade (TRG). Data were
analyzed both by visual examination of the data and by
applying a decision tree algorithm.
Results
From visual inspection of the data we found that using only
initial entropy and mean values for the ADC image as well
as their change during the first 2 weeks of treatment, we
could correctly classify 24 of the 25 patients as either
major response (TRG 1 or 2) or minor response (TRG 3 or
4). This was confirmed by building a decision tree for the
entire dataset. Applying machine learning techniques
where the data are divided into a training and a test
sample, we were hampered by the small data set, which
meant building a model on only part of the data set and
using the remaining patients to test the model gave large
variations in both the selected parameters and the ability
of the model to correctly predict the response of the
remaining patients (from 40% to 100%).
Conclusion
DWI imaging can provide information on the tumor
response as early as 2 weeks into CRT. Further work is
needed to improve the model and especially testing on an
larger data set is necessary.
[1] High-dose chemoradiotherapy and watchful waiting
for distal rectal cancer: a prospective observational study
Appelt, Ane L et al.
The Lancet Oncology , Volume 16 , Issue 8 , 919 - 927
[2] Watch-and-wait approach versus surgical resection
after chemoradiotherapy for patients with rectal cancer
(the OnCoRe project): a propensity-score matched cohort
analysis
Renehan, Andrew G et al.
The Lancet Oncology , Volume 17 , Issue 2 , 174 - 183
EP-1687 Texture analysis of 18F-FDG PET/CT predicts
local control of stage I NSCLC treated by SBRT
K. Takeda 1, K. Takanami2, Y. Shirata1, T. Yamamoto1, N.
Takahashi1, K. Ito1, K. Takase 2, K. Jingu1
1Tohoku University Graduate School of Medicine,
Radiation Oncology, Sendai, Japan
2Tohoku University Graduate School of Medicine,
Diagnostic Radiology, Sendai, Japan
Purpose or Objective
Recently, there are some reports that texture analysis of
18F-FDG PET/CT has better potential to predict outcome
of radiotherapy than existing PET parameters such as
maximum SUV. We evaluated reproducibility and
predictive value of some texture parameters based on
gradient-based delineation method and existing