Homomorphic determination of noise variance and denoising using a non-local means filter for assessing the accuracy of automated segmentation

Automated segmentation algorithms have been used more and more frequently for research and clinical purposes. There are available software packages that can determine volumes of brain structures and lesions. While signal to noise ratio in volumetric images is one of the determinants in the accuracy of such software, the effect of noise to the output results is usually not well described. We have studied effects of increased noise variance and denoising in evaluating the performance of automated segmentation tool using a synthetic phantom, one human dataset with artificially added noise, and 46 subjects scanned twice.