Post-processingof Monte Carlo calculated dose distributions

Sammanfattning: This Master Thesis focuses on denoising of Monte Carlo calculated dose distributions of radiosurgery treatment plans. The objective of this project is to implement a Denoising Autoencoder (DAE) and investigate its denoising performance when it has been trained on Monte Carlo calculated dose distributions generated with lower number of photon showers. The DAE is trained in a supervised setting to learn the mapping between corrupted observations and clean ones. The questions this thesis aims to answer are: (i) Can a DAE be used to denoise Monte Carlo calculated dose distributions, and thus predict the dose prior to a full simulation? Additionally, (ii) does incorporating prior knowledge of shot position increase the denoising performance? The results in this investigation have shown that the network successfully predicts the dose for low number of photon showers. In very heavy noise inputs the network denoising was in general successful, and the network could fill in missing data. The results indicated that the DAE could reduce the level of noise with an amount comparable with simulations that were done with 102 times more samples.