ESTRO 2020 Abstract book

S1022 ESTRO 2020

different organs (Nikolov et al., 2018). This model will be further trained to also segment the ventricles and the atria. Material and Methods Our dataset consisted of 186 CT scans acquired at Radboudumc, Nijmegen, with ground truth manual heart delineations made by an expert radiation oncologist following the atlas by Feng et al. (2011). Training, validation and test sets, composed of 132, 19 and 35 patients, respectively, were drawn at random, ensuring similar proportions of contrast CTs and CT scanners in all groups. The implemented residual 3D U-net architecture consisted of 7 residual contracting blocks (encoder), 7 residual upsampling blocks (decoder) and one fully connected block between them, as described in Nikolov et al. (2018). The inputs to the network had a size of 512 x 512 x 21, columns, rows and slices. The volumes were cropped around the heart with cropped Hounsfield Units outside the -200 to 300 range, following the variation range in the training set. Results We achieved 0.936 ± 0.05 mean Dice in our test set and 8.43 ± 7.28 mm mean Hausdorff distance at the 95th percentile (Figure 1). The segmentations of the inferior half of the heart showed less agreement to the ground truth relative to the superior half. The difficulty of distinguishing the heart and the liver in the transition between both organs may account for part of this discrepancy relative to the segmentation of the rest of the heart. Conclusion With additional improvements, the used residual U-net model may achieve near expert level accuracy in large datasets for research purposes such as modeling of radiotherapy induced toxicity. It may also be used in clinical practice to provide segmentations that require minimum adjustments, saving considerable time during the radiotherapy planning procedure. PO-1748 Carbon-ion boost followed by photon IMRT for PCa: dosimetric and geometric evaluations, AIRC- IG14300 G. Marvaso 1 , S.G. Gugliandolo 1 , S. Comi 1 , M. Pepa 1 , S. Russo 2 , B. Vischioni 2 , F. Valvo 2 , T. Giandini 3 , B. Avuzzi 4 , R. Valdagni 4 , D. Ciardo 1 , B.A. Jereczek-Fossa 1 , F. Cattani 5 , R. Orecchia 6 1 IEO- European Institute of Oncology IRCCS, Radiotherapy, Milan, Italy ; 2 National Center for Oncological Hadrontherapy CNAO, Clinical Department, Pavia, Italy ; 3 Fondazione IRCCS Istituto Nazionale dei Tumori, Unit of Medical Physics, Milano, Italy ; 4 Fondazione IRCCS Istituto Nazionale dei Tumori, Department of Radiation Oncology, Milano, Italy ; 5 IEO- European Institute of Oncology IRCCS, Medical Physics Unit, Milan, Italy ; 6 IEO- European Institute of Oncology IRCCS, Scientific Directorate, Milan, Italy Purpose or Objective The aim of this study is to evaluate dosimetric uncertainties of a mixed beam approach for patients with high-risk prostate cancer (PCa). The treatment consists of a carbon ion radiotherapy (CIRT) boost followed by whole- pelvis intensity modulated RT (IMRT). Dosimetric

Conclusion A deep learning neural network was successfully trained on 20 patients with expert delineation of BB structures, and subsequently tested on a further 5 patients. Similarity metrics were used to quantitatively assess the performance of the model, with an average dice similarity coefficient of 0.89 and an average Jaccard index of 0.80 across all test patients. The model has great potential to reduce the time taken for segmentation of OAR and improve the workflow for RT planning and adaptive strategies. Further work will involve the use of data augmentation, and 3D networks to increase accuracy in the underperforming areas of segmentation. Additional training datasets will also be employed to improve the model to a standard for clinical implementation. PO-1747 Segmentation of the heart using a Residual U- net model M. Fernandes 1 , J. Teuwen 2 , R. Wijsman 3 , B. Stam 4 , N. Moriakov 2 , J. Bussink 1 , R. Monshouwer 1 1 UMC St Radboud Nijmegen, Radiation Oncology, Nijmegen, The Netherlands ; 2 UMC St Radboud Nijmegen, Radiology, Nijmegen, The Netherlands ; 3 UMC Groningen, Oncology, Groningen, The Netherlands ; 4 Netherlands Cancer Institute, Radiotherapy, Amsterdam, The Netherlands Purpose or Objective Preclinical and clinical studies have shown that exposure of the heart to incidental radiation dose during radiotherapy for lung cancer is associated with an increased toxicity. However, robust dose-effect relationships of the heart have not yet been established due to the use of small cohorts. For more detailed toxicity prediction, it is essential to model the dependency of toxicity on specific substructures of the heart. However, manual delineation is a very time-consuming activity, especially since substructures of the heart are not routinely segmented for clinical cohorts. Automated whole heart segmentation (WHS) using deep learning (DL) can be used to segment heart substructures in large patient cohorts. For such models, the current best reported WHS generalized Dice score is approximately 0.9 and Hausdorff Distance 25 mm in CT scans, using variations of the original U-net (Zhuang et al., 2019). While promising, these results show that such DL models still need improvement to be used in clinical practice or radiation toxicity research. In this work we describe heart segmentation using a recently proposed residual 3D U-net architecture which previously achieved human expert level segmentation accuracy in