Comparing the Accuracy of the Bilateral Filter and Gaussian Filter for PET Image Post-Processing Through a Phantom Study
X Wu*, S Jani, M Dahlbom, D Low, J Lamb, University of California, Los Angeles, Los Angeles, CASU-E-I-84 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: To compare the accuracy of an edge-preserving filter to the standard Gaussian filter for PET image post-processing through a phantom study.
Methods: A phantom was used, consisting of four different spheres (diameter 1-4 cm, filled with a solution of 11-C) that were mounted inside an acrylic cylinder (filled with a solution of 18-FDG) to generate a time-varying signal to background ratio (SBR). To mimic lung and liver lesions, SBRs were chosen to range from 2 to 30 and 2 to 5, respectively. The edge-preserving filter used was the bilateral filter, which weights pixels based on both spatial distance and intensity difference. Reconstructed 3D-PET images were separately smoothed with the bilateral filter and Gaussian filter. A commercially available gradient-based technique was used to segment images and measure sphere volumes. Results were evaluated by the ratio of measured to true volume (RMT) of spheres. Paired two-tailed t-tests were applied to test for statistical significance.
Results: In the lung case, RMT differences were not statistically significant across all sphere sizes. In the liver case, statistically significant differences were obtained for the 3 cm and 4 cm spheres, with average RMTs of 1.14 (bilateral) and 1.11 (Gaussian). However, these differences did not amount to a clinically significant level. For the 2 cm sphere, the differences were not statistically significant; for the 1 cm sphere, RMT differences were statistically, and potentially clinically, significant: 0.66 (bilateral) vs. 1.31 (Gaussian). However, it should be noted that volume segmentation of the 1 cm sphere was not fully reproducible on a case-by-case basis.
Conclusion: This work indicates that the bilateral filter has a comparable accuracy to the Gaussian filter in volume measurements of 3D-PET images. The commercially available gradient-based segmentation algorithm is reasonably correct for spheres greater than 2 cm, but further studies are needed for smaller spheres.
Funding Support, Disclosures, and Conflict of Interest: NIH R01 CA096679