Contour-guided denoising of radiological dose profiles.

Abstract

Effective denoising of the output from radiation transport simulations can dramatically reduce the simulation time required for convergence. The fundamental goal in denoising is to preserve features of the profile while eliminating statistical noise. This is nontrivial, because the size of the variations due to small features is often no larger than the size of the variations due to noise. Traditional denoising algorithms have been based on replacing the deposited dose with a weighted average from nearby pixels. In such systems, if the averaging region is chosen too small, the noise is not well eliminated. If, on the other hand, the averaging region is chosen too large, small features are inadequately preserved. The algorithm described herein attempts to utilize global information about the profile in order to guide the denoising process. Specifically, the program extracts a set of isodose lines from the noisy data, which provide a high-level representation of the medium-scale characteristics of the profile. The isodose lines are extracted directly from the noisy (unconverged) dose profile and are, consequently, noisy themselves. The noise in the isodose lines is then surpressed by a wavelet threshold filter. Then the denoised isodose lines are modified to retain the positional relationships of the original isodose lines. Finally, a new data set is computed by interpolation between the smoothed isodose lines.