Abstract Book

S281

ESTRO 37

variability under all examined conditions. The choice of interpolator had generally a negligible impact on features. Furthermore, certain features were found to depend on the tumour volume, which limits their predictive value. Conclusion This study showed that radiomic features can generally be affected by inter-observer variability and image pre- processing. Intensity histogram features were the most broadly reproducible. Shape and textural features, due to their reliance on more nuanced image properties, expressed significantly lower reproducibility.

Conclusion Development of harmonization processes is crucial for radiomic analysis on larger cohorts. We proposed a methodology showing promising results on a large multicenter cohort. These methods will contribute to suppress the intrinsic limits of large scale, inter-center radiomics studies. PV-0529 Reproducibility of radiomic feature es in apparent-diffusion coefficient images of rectal cancer A. Traverso 1 , M. Kazmierski 1 , P. Kalendralis 1 , Z. Shi 1 , M. Welch 2 , H. Dahl Nissen 3 , A. Dekker 1 , L. Wee 1 1 Maastricht Radiation Oncology MAASTRO clinic, Radiotherapy, Maastricht, The Netherlands 2 University of Toronto, Department of Medical Biophysics, Toronto, Canada 3 Vejle Hospital, Department of Medical Physics, Vejle, Denmark Purpose or Objective Changes in Apparent Diffusion Coefficient (ADC) values derived from magnetic resonance imaging (MRI) have been shown to potentially predict therapy response in rectal cancer (Monguzzi 2013). Mentioned models could be improved by adding radiomic features. However, to avoid the risk of false positive rates, the stability of radiomic features in ADC maps must be examined. This preliminary study of MRI-radiomic features addresses the stability of ADC features with respect to inter-observer variability, image pre-processing and conversion to binary mask. Material and Methods We used a data set of ADC images of 23 patients taken from the THUNDER rectal cancer trial. Three independent observers manually delineated tumours. Radiomic features were computed with an open source software (pyradiomics) library for first-order intensity statistics (FO), shape metrics (SM) and textural analyses (TA) specifically grey-level co-occurrence matrix (GLCM) and grey level size-zone matrix (GLSZM). Reproducibility was assessed using concordance and intra-class correlation coefficients (CCC and ICC, respectively). The difference between drawn contours was quantified with Dice similarity. Sensitivity to image pre-processing was assessed by (i) in-slice spatial resolution down-sampling by up to half, (ii) histogram bin width quantization from 20 to 100 in steps of 20 and (iii) effect of different interpolation during down-sampling. The effect of filtration (Gaussian, curvature flow smoothing filters, Laplacian edge detection and Gaussian noise) and different conversions to binary mask were tested. Features computed from each combination of parameters were compared to features computed in the raw image Results FO features were the most reproducible, independent of resampling (78% reproducible at highest value) and quantization (89%), but stability was adversely affected by filtering and inter-observer variability. SM were highly sensitive to inter-observer variability, but insensitive to image pre-processing. TA features exhibited high

PV-0530 Parotid gland segmentation with deep learning using clinical vs. curated training data A. Hänsch 1 , T. Gass 2 , T. Morgas 3 , B. Haas 2 , H. Meine 1 , J. Klein 1 , H.K. Hahn 1 1 Fraunhofer MEVIS, Medical Image Computing, Bremen, Germany 2 Varian Medical Systems, Software Development, Baden, Switzerland 3 Varian Medical Systems, Product Management, Palo Alto, USA Purpose or Objective Modern radiotherapy planning requires careful delineation of organs. Done manually it is a very time- consuming task, hence fully automatic segmentation methods are desirable. Deep learning has shown to be promising for solving medical image segmentation tasks, see e. g. [1]. The availability of big amounts of training data with high quality expert annotations is often limited. We compare parotid gland segmentation results when training on a small set of curated data to training on a bigger set of more easily available routine-level

clinical annotations. Material and Methods

We train deep neural networks (DNN) on different CT datasets. Dataset A contains 50 training and 30 test cases