Abstract Book

S342

ESTRO 37

organisation/scientific-council/task-forces/european- particle-therapy-network). EPTN has several work packages (WPs), including but not limited to imaging, treatment planning, radio-biology and health economics. One WP is dedicated to clinical trials and it is foreseen that it will actively collaborate with the EORTC. More specifically, a fellowship at EORTC has been available since 2018 to liaise EPTN trials and EORTC. Of note, it is anticipated that in most patients, particle therapy will be delivered to prevent radiation-induced side effects and/or induction of secondary tumours. For the validation of these types of applications, there is a growing awareness that equating evidence-based medicine with randomized contr4olled trials is an excessive simplification and that other methodologies, such as the model-based approach, are available and need further exploitation. The model-based approach is an alternative evidence-based methodology designed to generate evidence to inform rational selection of patients who would most likely derive clinically relevant benefit from particle therapy in terms of prevention of potential adiation-induced side effects. The rationale behind the model-based approach is that particle therapy will only lead to improved clinical outcome due to less toxicity in patients, when two essential requirements are both met: (1) normal tissue sparing can actually be obtained with particles (Δdose), and (2) Δdose will actually result in clinically significant lower complication risk (or else lower normal tissue complication probability (ΔNTCP)). This model-approach could be applied in the field of particle therapy using the EORTC trial tools and EORTC/ESTRO-EPTN networks. SP-0645 Automated treatment planning at NKI: benefits and challenges E. Damen 1 , R. Harmsen 1 , G. Wortel 1 , A. Olszewska 1 , G. Van der Veen 1 , G. Retel 1 , P. Pronk 1 , D. Eekhout 1 , A. Duijn 1 , E. Lamers 1 , A. Tijhuis 1 , T. Janssen 1 , T. Wiersma 1 , R. De Graaf 1 , J. Trinks 1 1 Netherlands Cancer Institute, Radiotherapy department, Amsterdam, The Netherlands In the next decade, radiotherapy will see a trend towards daily plan adaptation, stimulated by improved imaging capabilities on the treatment machine, computational power and software for fast re-planning. Current treatment planning systems are not fit for this job, since they require a considerable amount of user interaction and suffer from unnecessary overhead. Automation of the entire treatment planning process is therefore a prerequisite for daily plan adaptation. At NKI we started a project to develop fully automated treatment planning and re-planning. The purpose of this work is to describe the results of the project so far and discuss the challenges we encountered. Material and Methods Treatment planning automation requires three main components: 1. A software framework that performs and tracks all tasks and is an intermediate between planning system and other applications; 2. Automatic creation of contours for organs at risk; 3. Automatic plan creation and dose optimization. At NKI we created FAST: a Framework for Automatic Segmentation and Treatment planning. Through strict separation of input data (all information needed to create a plan) from actual control of the treatment Abstract text Purpose/Objective: Symposium: Automatic planning – the challenges for success

planning system, FAST is scalable from one treatment site to the other, and easily expandable in functionality. FAST is unique in creating truly single-click treatment planning: it is started from our delineation software (Mirada, Mirada Medical) and results in a fully optimized and documented plan using the Pinnacle treatment planning system (Philips Medical Systems). Results: FAST has been in use in our clinic since 2015 for prostate treatment planning and has been expanded since then. Currently about 30% of all treatment plans (see Table 1) are automatically generated using FAST. The main challenges we encountered during development were: Quality of auto-contouring: in order to be useful in the automation process, the success rate of auto-contouring should be ≥ 80% for each organ individually. The current version of the Pinnacle auto-contouring module SPICE only reaches such high success rates for organs with high contrast on CT (e.g. lung). Auto-contouring of low contrast organs (e.g. parotid glands) result in a high failure rate. To possibly overcome this problem, we are currently investigating the quality of atlas-based segmentation in Mirada. Quality of the dose optimization engine: The Pinnacle Auto-Planning module mimics a human planner by creating help structures and continuously changing optimization objectives during the optimization process to reach the best plan for each patient. Finding the input parameters that result in the best plan for the majority of patients within a particular tumor group is rather labor-intensive. For Head&Neck it took a physicist several weeks of educated trial-and-error to find an optimal solution. When an optimal solution was found, blind tests among specialized planning RTTs and Radiation- Oncologists showed a clear preference for automatically generated plans over manually created clinical plans in 60% (prostate), 100% (rectum), and 80% (H&N) of the cases. Clinical plan acceptance rates: The rate at which automatically generated clinical plans are accepted without any change is around 60-90%, depending on the treatment site. For all sites, this rate is lower than in the blind tests described above. This is mainly due to the fact that it is very hard to judge the quality of a treatment plan which you did not optimize yourself. Consequently, planning RTTs tend to try to improve the automatically created plan although in many cases this does not lead to clinically significant improvements. To overcome this, we are currently investigating the use of a ‘knowledge- based’ plan QA approach in which we employ machine learning to rate plans based on prior plans for the same tumor site. Possible loss of skills and knowledge: Loss of planning skills may be prevented by excluding some planning procedures from automation. In our clinic, some ART procedures and planning for retreatment are still being planned manually. In addition, we started planning challenges, in which an automatically created plan is compared to manual plans created by several RTTs for the same patient. The results are discussed with all RTTs, resulting in a lively debate on plan quality and ways to improve the plan. The main benefits of FAST are a significant reducion in hands-on time for planning RTTs (Table 1), increased plan quality and increased uniformity of planning methods and planning data. Conclusions Complete automation of the treatment planning process is feasible and results in considerable reduction in hands- on time, consistent and high-quality plans, as well as data standardization. Improvements are needed in auto- contouring and dose optimization to fully exploit the potential of automation. In addition, new methods like