Purpose: To quantitively assess treatment planning robustness for intensity modulated proton therapy (IMPT) plans at our clinic using a machine learning (ML) model that predicts proton spot delivery uncertainties. During the regular robust optimization process the machine’s performance is generally not considered. This project seeks to assess the robustness of the plan when spot delivery errors are taken into account.
Methods: A left sided, breath hold accelerated partial breast irradiation (APBI) patient was selected for analysis. A machine learning model took as input the following: target spot position, MU, energy, and snout extension. The model then provided each spot’s predicted position and delivered MU as output. The dose distribution with the ML predicted spots was calculated in Raystation 11A and compared with the original calculation. Plan robustness was analyzed by perturbing patient CT of 5 mm in X, Y and Z directions and patient mass density by ±3.5%, resulting in 16 scenarios. The dose was recalculated for all perturbed scenarios.
Results: In the nominal scenario predicted by ML model’s spot positions, the dose to 99% of the CTV was 99.6% of the prescription dose, or a slight underdose of the CTV. For all perturbed scenarios at least 90% of the CTV received at least 100% of the prescription dose, and for all perturbed scenarios at least 95% of the CTV received at least 95% of the prescription dose. The left lung showed the largest variation in dose based on the robustness scenario, although in all scenarios lung dose was well within clinically acceptable dose constraints.
Conclusion: By including predictions of machine performance in robust analysis, we are confident our robustness settings in this APBI case are ensuring adequate coverage of the CTV. Future work by our group will explore CTV coverage and OAR dosages in more advanced case studies.
TH- External Beam- Particle/high LET therapy: Proton therapy – treatment planning/virtual clinical studies