ESTRO 37 Abstract book

S158

ESTRO 37

parotid glands increased with decreasing planning priority from 30 ± 11 Gy (P0) (mean ± SD) to 38 ± 13 Gy (P14), which benefitted the oral cavity, larynx and MCI (Fig 1 top). Decreasing the planning priority of the parotids led to a large increase in prediction error (SD) of the corresponding KB models, from only 1.6 Gy in the absence of inter-OAR dependency (P0) to 5 Gy for low priority (P14) (Fig 1 bottom). Especially the oral cavity benefitted from lowering the parotid priority, expressed by both a decrease in Dmean from 60 ± 7 Gy (P0) to 48 ± 11 Gy (P14) (Fig 2 top), and a decrease in prediction error (SD) of the corresponding KB models from 5 Gy to 2 Gy for the (Fig 2 bottom).

Conclusion RapidPlan for protons was able to generate KBPs with adequate OAR sparing. This study shows that fast, automated IMPT planning is feasible and that a single model can be used to create plans for other centers. It also confirms that planning is prone to variation. Hot spots outside PTVs should be rectifiable by a subsequent optimization. A limitation of the study is that only one standard beam angle set-up was used for KBPs, while clinical TPs had patient-specific beam arrangements. Further refinements to the KBP model and the optimization algorithm are needed, however completely automated planning is within reach. OC-0305 Knowledge-based models for automated planning are strongly affected by inter-organ dependency Y. Wang 1 , B.J.M. Heijmen 1 , S.F. Petit 1 1 Erasmus MC Cancer Institute, Radiation Oncology, Rotterdam, The Netherlands Purpose or Objective Knowledge-based (KB) dose prediction models have gained in popularity as a method for automated treatment planning. KB models predict OAR doses based on their geometry and location relative to the PTV, using a plan database of prior patients. Most published models consider each OAR separately. But for complex treatment sites, such as head and neck (HN), the achievable OAR doses may strongly depend on their priority compared to competing OARs. In these cases, KB models yield prediction errors and therefore suboptimal plans. In this study, we systematically investigated the effect of inter- OAR dependency on the prediction accuracy of KB models In total 108 oropharyngeal cancer patients were included in the study. Our fully automated, multi-criterial treatment planning system (not knowledge based) was used to generate plans that are Pareto-optimal and have consistent prioritization between sparing different OARs. Therefore the plans can be considered as golden standard to train and evaluate KB models. For each patient, 15 VMAT treatment plans were generated (1620 in total). For plan P0 the left parotid gland had the highest priority and was therefore not influenced by inter-OAR dependency. For each of the remaining 14 plans per patient (labeled P1, P2,…,P14) the planning priority of sparing the parotid glands vs. the other OARs was systematically lowered. Next, for each of the 15 sets of plans a KB model was trained on 54 patients and evaluated on the other 54. The KB models were based on the overlap volume histogram and principal component analysis. The effect of inter-OAR dependency was determined by comparing the prediction errors of the KB models as function of the priority of sparing the parotid glands. Results For all 1620 plans, the PTV coverage met the clinical constraints. As expected, the achieved D mean of the for HN cancer patients. Material and Methods

Conclusion The prediction errors of KB models are low in the absence of inter-OAR dependency (SD < 2 Gy), but strongly increase up to a SD of 5 Gy as the competition between sparing different OARs increases. The large influence of