ESTRO 2021 Abstract Book

S217

ESTRO 2021

compared to NLM.

Method / Metric

MSE

SSIM (%) GIPR 5%/5mm (%) GIPR 3%/3mm (%) GIPR 2%/2mm (%)

Slice 1e10 particles 3.13e-4 87.5

99.2±0.49

81.6±7.04

31.3±8.73

NLM on 1e10 particles 2.04e-4 94.4

98.2±1.92

80.5±11.2

35.9±13.2

Stacked ConvLSTM

9.73e-5 97.3

97.7±3.43

83.2±12.9

42.7±15.2

Conclusion The results highlight that recurrent neural networks show promise to accelerate dose distribution MC simulations. This project has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No. 880314. We thank GENCI for the resources provided in the Grands Challenges HPC on Joliot-Curie. OC-0309 Non-coplanar beam angle optimization in IMRT using a total-beam-space reference plan B. Schipaanboord 1 , B. Heijmen 1 , S. Breedveld 1 1 Erasmus MC Cancer Institute, Radiotherapy, Rotterdam, The Netherlands Purpose or Objective Beam angle optimization (BAO) is a discrete, combinatorial optimization problem that is difficult to solve to optimality. A novel, fully automated planning approach for BAO in IMRT is proposed. For each patient, plan generation is steered by an ‘ideal’ IMRT reference plan that includes all possible beam directions. The novel approach was validated for CyberKnife. Materials and Methods For each individual patient, the Pareto-optimal ideal reference plan (‘Ideal’) was generated with our in-house multi-criterial optimizer, using fluence map optimization including all available 91 non-coplanar candidate beams. Then, our in-house 3D dose-based segmentation algorithm was used to generate a deliverable plan with minimized dosimetric differences with the reference plan, for a pre-set number of final beams. Favorable MLC segments were sequentially identified and added to the plan. Initially, segments could be assigned to all 91 candidate beams, but when the pre-set final number of beams was reached, segments could only be added to the already selected beams. During plan generations, all delivery restrictions of the CyberKnife were taken into account, and the applied dose calculation engine was the same as used in the clinical TPS. Therefore, generated plans could in principle be clinically delivered. Validation was performed for 20 prostate SBRT patients, treated with 4 x 9.5 Gy. Per patient, 9 treatment plans were generated with varying pre-set maximum numbers of beams: from 10 to 50 in steps of 5 (‘Novel’). No manual fine-tuning of the automatically generated Novel plans was applied, and all clinical constraints were respected. Treatment plans created manually with the clinical TPS were used as a reference (‘Clinical’).