Abstract Book

S264

ESTRO 37

features were extracted from each sub-volume. Different machine-learning algorithms were used to build radiomic models for the prediction of loco-regional tumour control (LRC). The prognostic performance was measured by the concordance index (C-Index) . Finally, patients were stratified into groups of low and high risk of recurrence using the median risk value. Differences in LRC were evaluated by log-rank tests. Results The validation C-Index averaged over all learning algorithms and feature selection methods using the GTV e revealed a high prognostic performance for LRC (C-Index: 0.63±0.03 (mean±std)). The boundary sub-volumes GTV 5mm and GTV 10mm showed a slightly improved accuracy (C-Index: 0.64±0.03 and 0.64±0.02, respectively), while models based on the corresponding core volumes had a lower accuracy (C-Index: 0.59±0.03 and 0.60±0.03, respectively, (Fig. 2A)). Also the risk groups could be better separated using the GTV 5mm ( p <0.001), compared to the GTV e ( p =0.005) and the corresponding core volume ( p =0.16, (Fig. 2B)). The extension of the GTV 5mm sub- region by 2mm led to a similar prognostic performance (C-Index: 0.65±0.03). Conclusion In our investigation, radiomic models based on the boundary of the tumour showed a higher prognostic performance for LRC compared to models based on the tumour core. This indicates that the tumour boundary may contain more prognostic information than other parts of the tumour. The identification of tumour sub- volumes associated with treatment outcome may further improve the performance of radiomic risk models.

OC-0509 Prediction of radiation induced mucositis of H&N cancer patients based on a large patient cohort C.R. Hansen 1,2 , A. Bertelsen 1 , R. Zukauskaite 2,3 , L. Johnsen 1 , U. Bernchou 1,2 , D.I. Thwaites 4 , J.G. Eriksen 2,3 , J. Johansen 3 , C. Brink 1,2 1 Odense University Hospital, Laboratory of Radiation Physics, Odense, Denmark 2 University of Southern Denmark, Institute of Clinical Research, Odense, Denmark 3 Odense University Hospital, Department of Oncology, Odense, Denmark 4 School of Physics- University of Sydney, Institute of Medical Physics, Sydney, Australia Purpose or Objective Radiation-induced mucositis is a serious acute side effect, which can jeopardize treatment compliance and influence patient weight during treatment. The aim of this study was to develop a model to predict the risk of severe mucositis for inclusion in the overall treatment optimization during treatment planning. Material and Methods 535 patients from one institution receiving curative RT for H&N cancer were included. Doses were 66, 68 or 76 Gy in 33, 34 or 56fx. Except for simple glottic tumours, all patients were treated with IMRT/VMAT. Mucosal reactions were scored weekly during RT, as well as 2 and 8 weeks after RT using: 0 none, 1 erythema, 2 patchy mucositis, 3 confluent mucositis, 4 ulceration. The highest observed score was used as endpoint and dichotomised in stage 0-2 and 3+. DVH of the extended oral cavity (Brower et al.) was extracted from the TPS. Principal component analysis was used to uncouple the highly correlated dose metrics (V5,…,V75). Predictors available for the logistic model were the first 5 principal dose components (PC1-5), gender (M/F), weekly low dose chemotherapy (+/-), Nimorazole (+/-), treatment acceleration (5 or 6/10 fx/w), age (+/- 70 y), smoking status (never/former/current), site (larynx/hypopharynx, oral cavity/oropharynx/parotid, or nasopharynx/nasal cavity/unknown primary), and extended oral cavity volume. Parameter selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) within the statistical package R. The LASSO tuning parameter was chosen using 10-fold cross