Suitability of bilateral filtering for edge-preserving noise reduction in PET
1 PET Centre, Institute of Radiopharmacy, Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany
2 Department of Nuclear Medicine, University Hospital Carl Gustav Carus, Technische Universität Dresden, Dresden, Germany
EJNMMI Research 2011, 1:23 doi:10.1186/2191-219X-1-23Published: 5 October 2011
To achieve an acceptable signal-to-noise ratio (SNR) in PET images, smoothing filters (SF) are usually employed during or after image reconstruction preventing utilisation of the full intrinsic resolution of the respective scanner. Quite generally Gaussian-shaped moving average filters (MAF) are used for this purpose. A potential alternative to MAF is the group of so-called bilateral filters (BF) which provide a combination of noise reduction and edge preservation thus minimising resolution deterioration of the images. We have investigated the performance of this filter type with respect to improvement of SNR, influence on spatial resolution and for derivation of SUVmax values in target structures of varying size.
Data of ten patients with head and neck cancer were evaluated. The patients had been investigated by routine whole body scans (ECAT EXACT HR+, Siemens, Erlangen). Tomographic images were reconstructed (OSEM 6i/16s) using a Gaussian filter (full width half maximum (FWHM): Γ0 = 4 mm). Image data were then post-processed with a Gaussian MAF (FWHM: ΓM = 7 mm) and a Gaussian BF (spatial domain: ΓS = 9 mm, intensity domain: ΓI = 2.5 SUV), respectively. Images were assessed regarding SNR as well as spatial resolution. Thirty-four lesions (volumes of about 1-100 mL) were analysed with respect to their SUVmax values in the original as well as in the MAF and BF filtered images.
With the chosen filter parameters both filters improved SNR approximately by a factor of two in comparison to the original data. Spatial resolution was significantly better in the BF-filtered images in comparison to MAF (MAF: 9.5 mm, BF: 6.8 mm). In MAF-filtered data, the SUVmax was lower by 24.1 ± 9.9% compared to the original data and showed a strong size dependency. In the BF-filtered data, the SUVmax was lower by 4.6 ± 3.7% and no size effects were observed.
Bilateral filtering allows to increase the SNR of PET image data while preserving spatial resolution and preventing smoothing-induced underestimation of SUVmax values in small lesions. Bilateral filtering seems a promising and superior alternative to standard smoothing filters.