Quantitative Comparison of Anisotropic Diffusion, Non-Local Means and Gaussian Post-Filtering Effects on FDG-PET Lesions

Objectives While PET is a clinically valuable tool for cancer staging and monitoring response to treatment, its performance for lesion detection and quantification is limited due to high data noise. Post-reconstruction smoothing is routinely used to suppress noise and improve signal-to-noise ratios (SNR).The aim of this study is to qualitatively and quantitatively compare the performances of nonlinear Anisotropic Diffusion (AD) and Non-Local Means (NLMs) filters with those of linear Gaussian filters across a wide range of testing conditions.

Methods Nine cancer patients covering all BMI categories were selected with IRB approval and patient’s consent.FDG-PET data were acquired in a Toshiba TOF-PET/CT Large Bore scanner and reconstructed using fully 3D list-mode OSEM with model-based physical corrections in the system model, including image-domain PSF modeling. PET data corresponding to lower injection doses and shorter imaging times were also mimicked by uniform sub-sampling of the original data. For Gaussian filters, different filter strength levels were used. For AD filters, a wide range of parameters controlling the edge-preservation level were evaluated. For NLM filters, patches of anatomical information from CT (NLM-CT) were used to determine similarities between neighboring PET voxels. Resulting NLM-CT filtered images were blended with unfiltered images to match liver SNRs to those in Gaussian filtered images. Nineteen lesions across a range of anatomical locations such as lung, liver and breast were segmented and quantitatively analyzed.

Results Lesion contrast vs. liver coefficient of variation (COV) curves were plotted for each lesion. At matched liver COV levels, appropriately blended NLM-CT post-filtering resulted in the highest lesion contrast.

Conclusions Both AD and NLM filters outperformed Gaussian filters over a wide range of parameter selection and testing conditions. NLM-CT filtering produced highest SNR levels among the investigated post-filters.