A plan verification platform for online adaptive proton therapy using deep learning-based Monte–Carlo denoising

Abstract

Background PurposeThis study focused on developing a fast Monte Carlo (MC) plan verification platform via a deep learning (DL)-based denoising approach. It can maintain the MC dose calculation accuracy while significantly reducing the computation time. We also investigated its potential applications for online adaptive proton therapy (APT). MethodsFirst, we modeled an MC platform for proton therapy using the beam data library (BDL) required for treatment planning systems and then tested it with measured data. To accelerate the dose calculation, a dl-based denoising model with deep ResNet-deconvolution networks was developed. It was trained on the MC dose distribution of tumor sites obtained from 52 patients. The input MC dose distribution was with 1 × 106 simulated protons and the reference was 1 × 108. Fivefold cross-validation was performed. ResultsComparing the MC model with measured data, the range agreement (point-to-point difference) was better than 0.85 mm, and the lateral dose profile difference was below 2.41 %. For the denoising approach, we found a significant improvement in the dose volume histogram (DVH) for predicted images compared with input images. The root mean squared error (RMSE) for predicted versus reference images was 3.94 times lower than that of the input versus reference images. Moreover, for the gamma passing rate (3 mm/3%), the predicted versus reference images have an average of 99 %, much higher than the 82 % of the input versus reference images. The MC model successfully denoised the test dose map (high noise) to approach the reference (low noise). The elapsed time can be reduced to < 60 s (simulation time [high noise] + predicted time), much lower than the simulation time of a low noise dose map (e.g., >100 min of simulation of 1 + E8 particles). ConclusionsWe propose an analogous end-to-end fast plan verification platform using the combination of MC and DL methods. The platform yields dose calculation accuracy similar to MC codes while significantly reducing the elapsed time and can be used for online APT as an alternative to online plan verification.