ESTRO 36 Abstract Book

S849 ESTRO 36 2017 _______________________________________________________________________________________________

therapy.

performance status, and total tumor dose. The BN model has an AUC of 0.67 (95% CI: 0.59–0.75) on the external validation set and an AUC of 0.65 on a 5-fold cross- validation on the training data. A model based on TNM stage performed with an AUC of 0.49 (95% CI: 0.39-0.59) on the validation set. Conclusion The distributed learning model outperformed the TNM stage based model for predicting 2-year survival in a cohort of NSCLC patients in an external validation set (AUC 0.67 vs. 0.49). This approach enables learning of prediction models from multiple hospitals while avoiding many boundaries associated with sharing data. We believe that Distributed learning is the future of Big data in health care. References [1] Dehing-Oberije C. et al. Int J Radiat Oncol Biol Phys 2008;70:1039–44. doi:10.1016/j.ijrobp.2007.07.2323. EP-1597 Focal dose escalation in prostate cancer with PSMA-PET/CT and MRI: planning study based on histology C. Zamboglou 1 , I. Sachpazidis 2 , K. Koubar 2 , V. Drendel 3 , M. Werner 3 , H.C. Rischke 1 , M. Langer 4 , F. Schiller 5 , T. Krauss 4 , R. Wiehle 2 , P.T. Meyer 5 , A.L. Grosu 1 , D. Baltas 2 1 Medical Center - University of Freiburg, Department of Radiation Oncology, Freiburg, Germany 2 Medical Center - University of Freiburg, Division of Medical Physics- Department of Radiation Oncology, Freiburg, Germany 3 Medical Center - University of Freiburg, Department of Pathology, Freiburg, Germany 4 Medical Center - University of Freiburg, Department of Radiology, Freiburg, Germany 5 Medical Center - University of Freiburg, Department of Nuclear Medicine, Freiburg, Germany Purpose or Objective First studies could show an increase in sensitivity when primary prostate cancer (PCa) was detected by addition of MRI and PSMA PET/CT information. On the other side the highest specificity was achieved when the intersection volume between MRI and PSMA PET/CT was considered. Aim of this study was to demonstrate the technical feasibility and to evaluate the tumor control probability (TCP) and normal tissue complication probability (NTCP) of IMRT dose painting using combined 68 Ga-HBED-CC PSMA- PET/CT and multiparametric MRI (mpMRI) information in patients with primary PCa. Material and Methods 7 patients (5 intermediate + 2 high risk) with biopsy- proven primary PCa underwent 68 Ga-HBED-CC-PSMA PET/CT and mpMRI followed by prostatectomy. Resected prostates were scanned by ex-vivo CT in a localizer and prepared for histopathology. PCa was delineated on histologic slices and matched to ex-vivo CT to obtain GTV- histo. Ex-vivo CT including GTV-histo and MRI data were matched to in-vivo CT(PET). Contours based on MRI (GTV- MRI, consensus volume by two experienced radiologist), PSMA PET (GTV-PET, semiautomatic using 30% of SUVmax within the prostate) or the combination of both (GTV- union/-intersection) were created. Three IMRT plans were generated for each patient: PLAN77, which consisted of whole-prostate radiation therapy to 77 Gy in 2.2 Gy per fraction; PLAN95, which consisted of whole-prostate RT to 77 Gy in 2.2 Gy per fraction, and a simultaneous integrated boost to the GTV-union (Plan95 union )/- intersection (Plan95 intersection ) to 95 Gy in 2.71 Gy per fraction. The feasibility of these plans was judged by their ability to reach prescription doses while adhering to the FLAME trial protocol. TCPs based on co-registered histology after prostatectomy (TCP-histo), and normal tissue complication probabilities (NTCP) for rectum and bladder were compared between the plans. Results

Conclusion We found significant correlations between the occurence of early toxicities and dose-volume parameters of associated organs at risk for patients with primary brain tumours or prostate cancer receiving proton therapy. A reduction of NTCP could be predicted for proton therapy based on comparative treatment planning. After validation, these results may be used to identify patients who are likely to benefit most from proton therapy, as suggested by the model-based approach [1]. [1] Langendijk JA et al. (2013) Radiother Oncol 107, 267 - 273. EP-1596 Developing and validating a survival prediction model for NSCLC patients using distributed learning A. Jochems 1 , T. Deist 1 , I. El-Naqa 2 , M. Kessler 2 , C. Mayo 2 , J. Reeves 2 , S. Jolly 2 , M. Matuszak 2 , R. Ten Haken 2 , J. Van Soes 1 , C. Oberije 1 , C. Faivre-Finn 3 , G. Price 3 , P. Lambin 1 , A. Dekker 1 1 MAASTRO Clinic, Radiotherapy, Maastricht, The Netherlands 2 University of Michigan, Radiation oncology, Ann-Arbor, USA 3 The University of Manchester, Manchester Academic Health Science Centre, Manchester, United Kingdom Purpose or Objective The golden standard for survival prediction in NSCLC patients, the TNM stage system, is of limited quality for patients receiving (chemo)radiotherapy[1]. In this work, we develop an up-to-date predictive model for survival prediction based on a large volume of patients using a big data distributed learning approach. Distributed learning is defined as learning from multiple patient datasets without these data leaving their respective hospitals. Furthermore, we compare performance of our model to a TNM stage based model. We demonstrate that the TNM stage system performs poorly on the validation cohorts, whereas our model performs significantly above the chance level. Material and Methods Clinical data from 1299 lung cancer patients, treated with curative intent with chemoradiation (CRT) or radiotherapy (RT) alone were collected and stored in 3 different cancer institutes. Two-year post-treatment survival was chosen as the endpoint. Data from two institutes (1152 patients at Institute 1 and 147 at Institute 2) was used to develop the model while data from the 3 rd institute (207 patients at Institute 3) was used for model validation. A Bayesian network model using clinical and dosimetric variables was adapted for distributed learning (watch the animation: link censored). The Institute 1 cohort data is publicly available at (link censored) and the developed models can be found at (link censored). Results A Bayesian network (BN) structure was determined based on expert advice and can be observed in figure 1. Variables included in the final model were TNM stage, age,