Locally adaptive denoising of Monte Carlo dose distributions via hybrid median filtering

A fundamental prerequisite of computer aided radiotherapy treatment is the accurate estimation of the dose distributions so as to deliver a high homogeneous dose volume to the tumor without causing unnecessary side effects for the patient. The Monte Carlo (MC) method is considered as the most effective dose distribution computational technique. However, it is too slow and contaminated with noisy degradations that could affect the dose contour visibility and the estimates of dosimetric parameters. In this work, we propose a feature-adaptive median hybrid filter for the denoising of MC dose distributions. Median filtering has been shown to outperform the moving average (mean) in removal of impulsive noise (outliers) and preservation of edges, but it fails to provide the same degree of smoothness in homogeneous regions. We combine linear filters with the median operation to produce hybrid median filters. The filter output can be obtained as a weighted sum of the linear filter and the median operation depending on the properties of the local neighborhood. We evaluated the technique on different datasets, a challenging 2-D synthetic dataset of different geometric shapes at different scales with added noise and blurring, and 2-D/3-D water phantoms. The proposed filter, judged by mean square error, performed well in comparison with currently existing techniques. Denoising of full 3-D real treatment plan datasets has shown similar promise.