ESTRO 2020 Abstract book

S831 ESTRO 2020

re-calculated

eight

times.

Purpose or Objective Intensity modulated proton therapy (IMPT) uses multiple beams with steep in-field dose gradients and is therefore very sensitive for density and setup errors. Mathematical robust planning (RP) methods have been developed to account for these errors. However, RP methods take multiple scenario dose distributions into account during the optimization process, which makes RP generally slow as compared to margin-based planning methods. In this study, we developed and combined a Machine Learning Optimization (MLO) planning algorithm with a robust dose mimicking optimization algorithm to automatically create robust IMPT plans. We aimed to automatically generate robust IMPT plans with at least comparable target coverage robustness as clinically available robust IMPT plans for oropharyngeal cancer patients. Material and Methods In this study, robustly optimized IMPT plans of 65 HNC patients from our clinical and research archives were included. Dose distributions, contours, and CT image features of 60 patients were used to train a model to predict dose distributions for novel patients. Dose prediction was based on a random forest model with 96 trees at depth 10 including a conditional random field optimization. The target coverage during training were based on the primary and elective PTV. The predicted dose was converted to a deliverable plan using robust voxel- wise dose mimicking optimization, accounting for target robustness of 5 mm setup error and ±3% density uncertainty. The remaining five patients were used for validation of the MLO method. The beam configuration was copied from the clinical plans. Target robustness was assessed by a multi-scenario plan evaluation method comprising 16 dose recalculations with 8 positional isocenter shifts of 5 mm and a ±3% density uncertainty. The scenario dose distributions were combined into a voxel-wise minimum (vw-min) dose distribution and evaluated using the D 98% >94% criteria. In addition, we evaluated the spinal cord (D max <5400cGy), external D mean and several organs-at-risk (OARs) such as the parotid glands and oral cavity D mean . Results The fully automated robustly optimized MLO plans fulfilled the D98>94% criteria for primary CTV in all five patients, see figure 1. The D 98 of the MLO plans was on average (± SD) higher than in the clinical plans (6783 ± 58 cGy for the MLO plans and 6624±143 cGy for the clinical plans). On average, the vw-min D98 of the elective CTV of the MLO and clinical plans were 5139 ±139 cGy and 5160±75 cGy, respectively. The dose to spinal cord and external in the MLO plans were comparable to the clinical plans and fulfilled the clinical criteria. For the other OARs, the dose in the clinical plans was lower compared to the MLO plans, see figure 2.

Results The WEPL-based isodose levels were most similar to the full-fledged dose calculation for isodose levels of 15%, 25% and 35% with a median Dice value of 0.99 (range: 0.98- 1.00) across all scenarios and plans. The largest difference between the two plans occurred for an isodose level of 55%, due to different volume sizes of the isodoses. Plan 2 had a median Dice value of 0.94 (range: 0.72-0.98) and Plan 1 had a median Dice value of 0.76 (range: 0.55-0.93) (Figure 2). When an insert of bone were placed in front of each field, a median Dice value of 0.94 (range: 0.87-0.98) across both plans and all fields were calculated, while the median Dice value was 0.79 (range: 0.66-0.94), for an insert of air in front of each field.

Conclusion WEPL calculations compared to dose re-calculation using Eclipse TPS show promising results, especially for low isodose levels and small density changes. However further studies are needed to investigate how WEPL calculations can be used for higher isodose levels and more pronounced density changes. PO-1462 Automated robust planning for IMPT in oropharyngeal cancer patients using machine learning M. Huiskes 1 , R.G.J. Kierkels 1 , I.G. Van Bruggen 1 , M. Holmström 2 , H. Gruselius 2 , A. Fredriksson 2 , K. Berggren 2 , S. Both 1 , J.A. Langendijk 1 , F. Löfman 2 , E.W. Korevaar 1 1 University Medical Center Groningen, Radiation Oncology, Groningen, The Netherlands ; 2 RaySearch Laboratories, Machine Learning, Stockholm, Sweden