Denoising of medical images corrupted by Poisson noise

Medical images are often noisy owing to the physical mechanisms of the acquisition process. The great majority of the denoising algorithms assume additive white Gaussian noise. However, some of the most popular medical image modalities are degraded by some type of non-Gaussian noise. Among these types, we refer the Poisson noise, which is particularly suitable for modeling the counting processes associated to many imaging modalities such as PET, SPECT, and fluorescent confocal microscopy imaging. The aim of this work is to compare the effectiveness of several denoising algorithms in the presence of Poisson noise. We consider algorithms specifically designed for Poisson noise (wavelets, Platelets, and minimum descritpion length) and algorithms designed for Gaussian noise (edge preserving bilateral filtering, total variation, and non-local means). These algorithms are applied to piecewise smooth simulated and real data. Somehow unexpectedly, we conclude that total variation, designed for Gaussian noise, outperforms more elaborated state-of-the-art methods specifically designed for Poisson noise.