ESTRO 2020 Abstract book

S1033 ESTRO 2020

to expect an efficient protocol and a rigorous interpretation of the results. PO-1763 Generation of artificial realistic MR and CT images using StyleGAN L. Fetty 1 , M. Bylund 2 , G. Heilemann 1 , D. Georg 1 , T. Nyholm 2 , P. Kuess 1 , T. Löfstedt 2 1 Medical University of Vienna, Department of Radiation Oncology, Vienna, Austria ; 2 Umeå University, Department of Radiation Sciences, Umeå, Sweden Purpose or Objective Deep learning got very popular and relevant over the last few years for medical imaging and image processing. Deep learning methods must be trained on large sets of data in order to achieve high performance, but this can be challenging in medical applications. It may be even difficult to obtain data, because of ethical considerations and hospital policies, which have made techniques such as data augmentation popular. If a research group is not able to collect enough data, images can be stochastically manipulated to artificially increase the training data set size. One possibility to achieve an increase in the data set size is to synthesis images by generative models, such as generative adversarial networks (GAN). So far this type of network has had the drawback of only being able to generate low-resolution images, which is only useful in limited cases. A recent GAN method (StyleGAN), developed by Karras et al. (CVPR 2019), generates high- resolution outputs of high quality, which could help research groups to train models even with low amounts of real data. In our study, we applied the StyleGAN to generate artificial Computed Tomography (CTs) and T2- weighted Magnetic Resonance Imaging (T2w MRIs) and explore the self-modeled latent space for one important feature—the longitudinal slice position. Material and Methods For the training process, pelvic patient data from 100 patients with T2w MR and CT images were collected. The trained StyleGAN model generates synthetic CT or T2w MR images with a resolution of 512x512. We evaluated manipulating the latent vectors in the style space of the StyleGAN by identifying the direction encoding the longitudinal slice position, and performed affine transformations using the found direction. To achieve this, an encoder was trained to generate latent style vectors from the input medical images, and a logistic regression model was used to predict the training data slice position (high or low) from the corresponding latent style vectors. Results The StyleGAN model produced realistic CT and T2w MR images with an FID score of about 12. The model encoded the slice positions along a linear direction, making it possible to generate whole 3D volumes, as illustrated in Figure 1. Using the learned logistic regression model, the latent vectors were manually adapted to match user specific needs, such as generating the next slice in either direction. The error in the generated slice (compared to the ground truth) increased with the distance to the starting slice index, which as seen in Figure 1. Starting from a central slice, the errors for MR and CT were below 0.3 RMSE and 50 HU MAE, respectively.

Conclusion The proposed protocol provides a feasible, practical and efficient multi-center cross-platform MR-sim daily QA solution. PO-1762 Dummy Run for bone SBRT in french multicentric study S. Thureau 1 , L. Lebret 1 , F. Jean Christophe 2 , R. Modzelewski 3 , N. Bonnet 4 , V. Marchesi 5 , A. Lisbona 6 1 Center Henri Becquerel, Radiotherapy, Rouen, France ; 2 Institut de Cancerologie de Lorraine - Alexis Vautrin, Radiotherapy, Nancy, France ; 3 Ceter Henri Becquerel, Medical Informatic, Rouen, France ; 4 Unicancer, Unitrad, Paris, France ; 5 Institut de Cancérologie de Lorraine - Alexis Vautrin, Radiotherapy and Physic Medical, Nancy, France ; 6 Institut de Cancérologie de l'Ouest - Gauducheau, Radiotherapy and Physic Medical, Nantes, France Purpose or Objective Oligometastasis concept was proposed for patients with 1 to 5 metastases with intermediate prognosis. Randomized phase 3 trial STEREO-OS (NCT03143322) assess interest for stereotactic radiotherapy in addition with systemic treatment for oligometastatics patients. The effectiveness of treament is related to the quality of radiotherapy. In this context, quality controle with a dummy-run (DR) was established before the activation of participating centers. Material and Methods 2 prescription schedules were accepted, 35Gy/5 fractions/3 fractions per week or 27Gy/3 fractions/3 fractions per week. A double DR was performed with an analysis of recaluation and contouring and an analysis of planification (realized from each own center planification software). The two dummy-run were realized independently. The contouring dummy-run consisted on a vertebral metastasis with provision of a CT-scan and an RMN T2. They should contouring GTV, CTV, PTV and spinal cord. The material was then returned on Aquishare plateform software for evaluation of correlation index and dosimetric constraints analysis. Results For contouring DR, Kappa index (KI) for GTV was highly correlated (mean=0,76 [0,7-1]), KI for CTV was almost perfect (mean=0,89 [0,83-1]) such as KI for PTV (mean=0,89 [0,83-1]). For GTV, 13 centres had an acceptable KI whereas for CTV and PTV all centers had an excellent KI. Concerning planification DR, of the 19 centers, 17 were able to hold dose constraints, 2 had inacceptable deviations requiring respectively 1 and 2 corrective planifications. Conclusion This DR shown a high correlation for target volumes between centers. This has shown a good knowledge of American recommendations published in 2012 and used internationally. Despite the use of different planification softwares, 89% of the centers had the ability to follow dose constraints of protocol. This quality assurance is essential