Denoising dose distributions from Monte Carlo simulations of proton therapy treatments

Radiation therapy is extensively used for cancer treatments. To compute radiation dose distributions in proton therapy, Monte Carlo dose calculation is known to provide better accuracy than analytical algorithms typically available in the treat- ment planning system. However, due to the stochastic nature of Monte Carlo methods, the resulting dose maps suffer from unavoidable statistical noise, which may affect clinical metrics such as the dose-volume histogram (DVH). A procedure to quantify this statistical uncertainty is described and discussed in this work. Noise can be reduced by increasing the number of simulated particles, which also increases the computation time. In order to keep the calculation time practical, several noise filtering techniques are studied here, including Gaussian and median filters, wavelet threshold denoising (WTD) and the non-local means (NLM) filter. Improvements regarding performance and speed of the NLM filter are proposed. The denoising performance of all considered methods are compared using various image quality criteria and dosimetric measures. Based on two test cases (i.e., lung and prostate tumor), it was found that the evaluated methods provided acceleration of MC results towards stastically more accurate results. The improved NLM filter produces the best results compared to the other methods.