Deep learning prediction of scenario doses for direct plan robustness evaluations in IMPT for head-and-neck

Abstract

Objective. Intensity modulated proton therapy (IMPT) is susceptible to uncertainties in patient setup and proton range. Robust optimization is employed in IMPT treatment planning to ensure sufficient coverage of the clinical target volume (CTV) in predefined scenarios, albeit at a price of increased planning times. We investigated a deep learning (DL) strategy for dose predictions in individual error scenarios in head and neck cancer IMPT treatment planning, enabling direct evaluation of plan robustness. The model is able to differentiate between scenarios by using embeddings of the scenario index.Approach. To accommodate resolution disparities in planning CT-scans and accommodate the setup error scenarios, we introduced scenario-specific isocentric distance maps as inputs to the DL models. For 392 H&N cancer patients, high-quality 9-scenario ground truth (GT) robust plans were generated with wish-list driven fully automated multi-criteria optimization. The scenario index is converted to one-hot-vector that is used to derive the scenarios embeddings through the training of the DL model, aiding the model to predict a scenario specific dose distribution.Main results. The model achieved within 1%-point of agreement with the GT the predictedV95%of the voxelwise minimum dose for CTV Low and CTV High for 96% and 75% respectively of the test patients. Considering all robustness scenarios, median differences were 0.035%-point for CTV HighV95%, 0.11%-point for CTV LowV95%, 0.29 GyE for parotidsDmean, 0.7 GyE for submandibular glandsDmeanand 0.9 GyE for oral cavityDmean. Prediction of full 3D dose distributions for all scenarios took around 14 s.Significance. Predicting individual scenarios for robust proton therapy using DL dose prediction is feasible, enabling direct robustness evaluation of the predicted scenario doses.