Abstract
1666
Objectives: In this work we evaluate and compare wavelet-based denoising approaches for whole body FDG PET imaging. Apart from assessing the level of denoising, one of the main objectives of the study was to evaluate whether wavelet-based denoising is able to preserve local quantitative information.
Methods: Six different techniques (based on soft thresholding and Stein's Unbiased Risk Estimate) were tested and compared with Gaussian filtering as a traditional denoising approach. The main principle of these approaches relies on the selection of an adaptive threshold specific to the various levels of resolution present in an image. These different resolution layers were obtained using a wavelet transform which in our case was the Isotropic Undecimated Wavelet Transform implemented using the "à trous" algorithm. Test images consisted of a GATE simulation of a cylinder containing spheres filled with FDG and a series of fourteen clinical oncology whole-body images.
Results: The Gaussian filtering reduces noise drastically (up to -80%) but at the expense of both significant intensity loss (up to -30% in lesions) and contrast attenuation (up to -45%). In contrast, noise suppression can be achieved using wavelet-based methods (up to -76%) without alteration of contrast or mean intensity (<-5%for both parameters).
Conclusions: As expected, Gaussian filtering despite noise suppression leads also to undesired quantitative and qualitative changes On the other hand, wavelet-based approaches lead to significant denoising without altering resolution or quantitative accuracy.
- Society of Nuclear Medicine, Inc.