Spatially-guided non-local mean filter for denoising of clinical whole-body PET images

A novel spatially-guided non-local mean (SG-NLM) filter for denoising whole-body PET images is proposed. In this approach, PET images are first clustered into a number of regions with homogeneous uptake and a Bayesian framework exploited to estimate the noise level for automatic parameter setting. Then, instead of defining a search window to find similar patches (as in conventional NLM), only patches belonging to the same regions are collected and compared to facilitate the patch searching process. The SG-NLM filter was compared to conventional post-reconstruction Gaussian and NLM filters. Experimental measurements using the Jaszczak phantom and clinical whole-body PET/CT studies were used to evaluate the performance of the proposed approach. In the Jaszczak phantom, the signal to noise ratio (SNR) improved from 26.5 when using Gaussian smoothing to 29.9 and 30.7 (averaged over all six spheres in the Jaszczak phantom) using the NLM and SG-NLM filters, respectively. The clinical studies further demonstrated the superior performance of SG-NLM method by leading to higher SNR while yielding a quantification change in malignant lesions of -2.4% compared to -11.7% and -2.9% achieved using Gaussi-an and conventional NLM filters, respectively. The SG-NLM proved to improve the contrast, SNR and quantitative accuracy compared to Gaussian and NLM approaches, and can be utilized as an alternative post-reconstruction filter in clinical whole-body PET/CT imaging.