Multiple PET Reconstruction Assisted Non-local Mean Denoising of PET Images

Non-local mean (NLM) denoising is commonly used for noise suppression in natural as well as medical imaging. Basically, the NLM filter takes advantage of the redundant information present in the image in the form of repeated structures/patterns to identify the underlying signals. In medical imaging (particularly PET and SPECT imaging), different representations of the image data under study (target or original image) could be reconstructed via applying different reconstruction settings. These representative (auxiliary) images bear very similar patterns/structures to the original/target image with different signal-to-noise ratios (SNR) which are ideal for use in the NLM denoising approach. This study proposed the multiple-reconstruction NLM filtering approach (referred to as MR-NLM) for noise reduction in PET imaging, wherein the redundant information present in auxiliary PET images are employed to conduct the NLM denoising process. The MR-NLM method relies on 12 additional PET image reconstructions (apart from the target PET image) using the same iterative algorithm with different iterations and subset numbers. Thereafter, for each target voxel, patches of voxels are extracted at the same location from all auxiliary PET images to be fed into the NLM smoothing process. To evaluate the performance of the MR-NLM algorithm, post-reconstruction denoising approaches including the conventional NLM, bilateral, and Gaussian filters were implemented and compared using 25 18F-FDG clinical whole-body (WB) PET/CT studies. The clinical studies demonstrated superior performance of the MR-NLM approach which established a promising compromise between noise suppression and preservation of the underlying signal/structures in PET images leading to higher SNR compared to the conventional NLM approach (34.9±5.7 versus 32.4±5.5). Though MR-NLM exhibited promising performance, this method suffers from long processing time due to the requirement of multiple reconstructions of raw PET data.