ESTRO 36 Abstract Book

S233 ESTRO 36 2017 _______________________________________________________________________________________________

OC-0444 Pareto-optimal plans as ground truth to validate a commercial knowledge-based DVH- prediction system E. Cagni 1 , A. Botti 1 , Y. Wang 2 , M. Iori 1 , S.F. Petit 2 , B.J. Heijmen 2 1 Arcispedale S. Maria Nuova - IRCCS, Medical Physics Unit, Reggio Emilia, Italy 2 Erasmus MC Cancer Institute, Department of Radiation Oncology, Rotterdam, The Netherlands Purpose or Objective The purpose of the current study was two fold. First, to evaluate the DVH prediction accuracy of RapidPlan (Varian Medical Systems, Palo Alto) using a large database of Pareto optimal treatment plans. These were consistently generated with automated prioritized multi-criterial treatment plan optimization, independent of RapidPlan, and therefore can be considered as an unbiased, ground truth of achievable plan quality. Second, to determine the importance of the size/variability of the plan database on the accuracy of the RapidPlan DVH predictions. Using the automatically generated plans, the prediction accuracy of RapidPlan could be investigated without an impact of unavoidable plan quality variations related to manual planning. Material and Methods A previously published database of 115 Pareto-optimal prostate VMAT plans, consistently generated with automatic prioritized planning, was used in this study [Wang et al, PMB, 2016]. Separate Rapidplan prediction models were generated for training groups consisting of 20, 30, 45, 55, or 114 randomly selected plans, the latter using a leave-one-out technique. In a second experiment, Model-20 was also built for 4 other groups of randomly selected training patients. Prediction accuracy of all models was assessed using a fixed, independent validation group of 60 plans and comparing predicted dose parameters (rectum Dmean, V65, and V75, anus Dmean, and bladder Dmean) with the achieved values of the Pareto optimal plans. Results For Model-114, the absolute (relative) prediction errors (mean±SD), for rectum Dmean, V65, and V75 were 1.8±1.4Gy (6.9±5.6%), 1.0±0.9% (8.9±13.4%), and 1.6±1.4% (36.3±62.2%), respectively. For anus and bladder Dmean, these errors were 2.2±1.7Gy (18.3±21.3%) and 1.8±1.3Gy (4.9±4.2%), respectively. For 63.3% of the validation plans, Model-114 predicted a lower rectum V65 than could actually be achieved. Because of the prioritized optimization used for generating the input Pareto-optimal plans, this can only be realized by underdosing the target. In 36.7% of plans, the predicted V65 was higher than obtained in the input plan, possibly losing an opportunity for lower rectum dose. Table 1 demonstrates equal prediction accuracies for model-114 and the smaller models (only first Model-20 included). Table 2 compares Model-114 with all 5 investigated Models-20, showing significant differences in performance of the Models-20.

Conclusion Rapidplan DVH prediction is a useful tool to guide treatment plan generation, although significant prediction inaccuracies may occur, even when the training database consists of as many as 114 treatment plans. Since all plans to train and validate RapidPlan DVH prediction were Pareto optimal, prediction errors for a database of manually optimized treatment plans are likely larger than presented here. For models based on a small number of plans (N=20), prediction performance depends strongly on the selected training patients, therefore larger models are recommended. OC-0445 Probabilistic optimization of the dose coverage – applied to treatment planning of cervical cancer D. Tilly 1,2 , A. Holm 2 , E. Grusell 1 , A. Ahnesjö 1 1 Uppsala University Hospital, Department of Immunology- Pathology and Genetics, Uppsala, Sweden 2 Elekta, R&D, Stockholm, Sweden Purpose or Objective Probabilistic optimization is an alternative to margins for handling geometrical uncertainties in treatment planning of radiotherapy where the uncertainties are explicitly incorporated into the plan optimization through sampling of treatment scenarios and thereby better exploit patient specific geometry. In this work, a probabilistic method is presented based on statistical measures of dose coverage, similar to the basis for margin based planning. The idea is that the dose planner requests a dose coverage to a specified probability, which the algorithm then fulfils. Material and Methods The van Herk margin recipe is designed to deliver sufficient target dose coverage in 90% of the treatments. The probability is however rarely specified. We generalize this prescription approach to include the probability explicitly through the concept of Percentile Dose Coverage (PDC), i.e. the dose coverage that is at least fulfilled to a specified probability. The PDC used in this work for target minimum dose criterion is the probability for the dose-volume criteria D 98% to be fulfilled with 90% probability which we denote as D 98%,90% . For optimization, we make use of the Expected Percentile Dose Coverage (EPDC), defined as the average dose coverage below a given PDC. The EPDC is, in contrast to PDC, a convex measure which allows for standard optimization techniques to be used for finding an optimal treatment plan. We propose an iterative method where a treatment optimization is performed at each iteration and