ESTRO 36 Abstract Book

950 / 1082

S934 ESTRO 36 2017
_______________________________________________________________________________________________
compared to NTCP models that predict the risk of
pneumonitis.
Material and Methods
A cohort of 9 NSCLC patients made up of 4 patients (ID 71
to 74) that developed pneumonitis after radiotherapy and
5 patients that remained asymptomatic after radiotherapy
(ID 10 to 14) was selected. Radiotherapy planning CT
images were acquired at 3 mm slice thickness with a pixel
resolution of 1 mm. 6 patients were treated with 55 Gy in
20 fractions and 3 patients with 60 Gy in 30 fractions.
Treatment plans were produced on Eclipse using a pencil
beam convolution dose calculation algorithm. 7
radiobiological models were used to calculate NTCP on the
whole, right and left lungs. Image texture analysis was
used to calculate 99 unique features on 32x32 and 20x20
pixel 2 subimages within the whole lung volume. Redundant
texture features were removed and a neural network (NN)
trained to classify the results.
Results
The predicted NTCP values are shown in Figure 1 for the
analysis of the whole lung volume (normal lung tissue
excluding the GTV). Similar results were obtained for the
right and left lungs. Although model 5, symptomatic or
radiographic pneumonitis <=6 months, showed high NTCP
values this was not specific to patients with confirmed
pneumonitis. Similar values were obtained in patients
showing no signs of pneumonitis. The image texture
analysis results identified the risk of pneumonitis most
notably in the right lung (87.49%).
Figure 1: NTCP results on the 9 patients (ID 10-14
asymptomatic, 71-74 symptomatic).
Table 1: Texture analysis classification on the whole lung
and right and left lung volumes.
Conclusion
These preliminary results show that it is possible to predict
radiation-induced pneumonitis, both prior to treatment
and independently of dosimetric evaluation, using image
texture analysis of the radiotherapy planning CT images.
However further validation on a larger patient cohort is
required.
EP-1726 Efficacy of vendor supplied distortion
correction algorithms for a variety of MRI scanners
E.P. Pappas1, I. Seimenis2, D. Dellios2, A. Moutsatsos1, E.
Georgiou 1, P. Karaiskos1
1 National and Kapodistrian University of Athens, Medical
Physics Laboratory - Medical School, Athens, Greece
2Democritus University of Thrace, Medical Physics
Laboratory - Medical School, Alexandroupolis, Greece
Purpose or Objective
Although inherently distorted, Magnetic Resonance Images
(MRIs) are being increasingly used in stereotactic
radiosurgery (SRS) treatment planning in order to take
advantage of the superior soft tissue contrast they exhibit.
MR scanner manufacturers have equipped their units with
distortion correction algorithms to mainly compensate for
gradient nonlinearity induced spatial inaccuracies. The
purpose of this study is to assess the accuracy of these
algorithms by comparing distortion maps deduced with
and without the optional distortion correction schemes
enabled for a variety of MRI scanners.
Material and Methods
A custom acrylic-based phantom was designed and
constructed in-house. Its external dimensions were limited
to approximately 17x16x16 cm3 in order to accurately fit
in a typical head coil while extending to the edges of the
available space. On eleven axial planes, a total of 1978
holes were drilled, the centers of which serve as control
points (CPs) for distortion detection. Center-to-center CP
distance is 10 mm on x and y axis and 14 mm on z axis,
resulting in adequately high CP density. The phantom was
filled with copper sulfate solution and MR scanned at 1.5T
(SIEMENS Avanto, Philips Achieva) and 3.0T (SIEMENS
Skyra) using the corresponding standard clinical MR
protocol for SRS treatment planning. All scans were
repeated after disabling the vendor supplied distortion
correction scheme. The phantom was emptied and CT
scanned to provide the reference CP distribution. In-house
MATLAB routines were developed for distortion
assessment. Reference and evaluated CP distributions
were spatially registered and compared to derive 3D
distortion maps. This methodology does not consider
uniform geometric distortion as it cancels out during the
spatial registration step. This results in omitting uniform
susceptibility-induced CP dispositions and thus mainly
takes into account machine-related distortions.
Results
At central slices, around the scanners’ isocenters
minimum distortion was detected even with the correction
algorithms disabled. However, at the edges of the
available space distortion magnitude greatly increases
(figure 1) and efficacy of algorithm becomes paramount.
Maximum detected distortion reaches 3 mm for the
SIEMENS 3.0T scanner but is reduced to less than 1.5 mm
if the correction algorithm is enabled. For the 1.5T
scanners slightly lower corresponding values were
observed.