ESTRO 2020 Abstract book

S1034 ESTRO 2020

image features concatenated with the clinical features. To the best of our knowledge, this is the first work that utilizes 3D convolutional neural networks to extract CT image features for outcome prediction on head and neck cancer patients. Results We evaluated the performances of the proposed pipeline for multiple outcome endpoint prediction at two years. In Table1, the achieved accuracy and area under the curve (AUC) are reported for both the training and test set. Prediction models with a good performance on the test set were obtained for the endpoints local recurrence (accuracy: 0.76; AUC: 0.69) and regional recurrence (accuracy: 0.84; AUC: 0.86) at two years. For distant metastasis: the model performance was reasonable (accuracy: 0.81; AUC: 0.60). The predictive performance of the model for disease free survival was limited (accuracy: 0.60; AUC: 0.54). Conclusion We developed and tested deep learning based models for prognostic outcome prediction of head and neck cancer patients with promising results by using image features identified by 3D CNNs in combination with clinical features. The prediction models for local and regional recurrence at two years showed good performances on the test set. PO-1765 Treatment of face and scalp with HelicalTomotherapy: feasibility, robustness and dosimetric accuracy L. Marrazzo 1 , G. Simontacchi 2 , C. Arilli 1 , S. Calusi 3 , M. Casati 1 , A. Compagnucci 1 , C. Talamonti 1,3 , L. Livi 2,3 , S. Pallotta 1,3 1 Azienda Ospedaliera Universitaria Careggi, Medical Radiation Physics, Firenze, Italy ; 2 Azienda Ospedaliera Universitaria Careggi, Radiation Oncology, Firenze, Italy ; 3 University of Florence, Department of Experimental and Clinical Biomedical Sciences “Mario Serio”, Florence, Italy Purpose or Objective Aim of this study is to report on the feasibility, robustness and dosimetric accuracy of total scalp and face irradiation (TSFI) with helical Tomotherapy. Patients affected by cutaneous lymphomas with an extensive involvement of scalp and face are eligible for this treatment. In spite of the rareness of this kind of lesions, an analysis is worthwhile, due to the challenges imposed by the shape and characteristics of the target. Material and Methods Patients underwent a CT scan (3mm slice thickness) using a thermoplastic head-shoulder mask. The clinical target volume (CTV) was delineated by including the patient's entire skin surface of scalp and face. The planning target volume (PTV) was obtained with an isotropic 3 mm expansion of the CTV, by clipping the structure at the patient surface. An effective template was generated for the treatment, in order to maximize target coverage and minimize dose to the internal organs, building on purpose help structures. Patients were treated with 14Gy/7fr. To test the dosimetric accuracy, the same template was applied to the head of an Alderson Rando phantom. The obtained plan was delivered and measured with gafchromic EBT3 films (placed on three axial slices). Due to the challenging dose distribution with steep dose gradients, plan robustness toward residual setup errors was evaluated (for the 3 patients treated) by recalculating clinical plans on the daily MVCT using the Plan Adaptive module of Tomotherapy planning station. Differences between planned and accumulated doses were recorded for some dosimetric parameters on CTV, PTV and organs at risk (OARs) and statistical significance was evaluated using paired two-sided Wilcoxon signed-rank test. Results

Conclusion This study shows the potential of the StyleGAN in creating realistic artificial subvolumes of CTs and T2w MRIs. These synthetic medical images could for instance be used to train deep learning models on other tasks. They are fully anonymous and thus potentially independent of patient consent. PO-1764 Prognostic outcome prediction for head and neck cancer patients using convolutional neural networks J. Guo 1,2 , T. Zhai 1,3 , R.J.H.M Steenbakkers 1 , S. Both 1 , J.A. Langendijk 1 , P.M.A. Van Ooijen 1,2 , N.M. Sijtsema 1 1 University of Groningen- University Medical Center Groningen UMCG, Department of Radiation Oncology, Groningen, The Netherlands ; 2 University of Groningen- University Medical Center Groningen UMCG, Machine Learning Lab- Data Science Center for Health DASH, Groningen, The Netherlands ; 3 Cancer Hospital of Shantou University College, Department of Radiation Oncology, Shantou, China Purpose or Objective In existing studies, radiomic and clinical features were used for prognostic outcome prediction. Deep learning, as a feature mining and data classification technique, enables the identification of even more representative and larger number of image features. The objective of this work was to develop and test deep learning (DL) based models for prognostic outcome prediction of head and neck cancer (HNC) patients by using image features identified by 3D convolutional neural networks (CNNs) in combination with clinical parameters. Material and Methods The dataset includes 444 HNC patients with contrast- enhanced CT scans. All patients were treated with curative (chemo)radiation between 2007 and 2015. The Gross Tumor Volume of the primary tumors were delineated by an experienced radiation oncologist. Clinical parameters (gender, age, TN-stage, clinical stage, treatment modality, WHO performance status and tumor site combined with HPV status) and multiple endpoints (local recurrence, regional recurrence, distant metastasis, disease free survival) were collected. Patients treated before June 2012 were included in the training set and the rest in the test set. We developed a DL based pipeline for the prediction of multiple outcome endpoints. As illustrated in Figure 1, the pipeline consists of two parts: the extraction of image features identified by a 3D CNN and the prediction of the outcomes based on the extracted image features combined with clinical features. We used a 3D CNN architecture based on ResNet for the extraction of image features from the CT volumetric data of the delineated primary tumors. Furthermore, an artificial neural network was used to optimize the prediction model based on the