Non-local mean denoising using multiple PET reconstructions.

Area of Science:

• Medical Imaging
• Image Processing
• Nuclear Medicine

Background:

• Non-local mean (NLM) filtering is a standard technique for denoising natural and medical images, relying on image texture redundancy.
• Positron Emission Tomography (PET) and SPECT imaging generate data that can be reconstructed with varying parameters, yielding images with similar structures but different noise characteristics.

Purpose of the Study:

• To introduce and evaluate a novel Multiple-Reconstruction Non-local Mean (MR-NLM) filtering approach for PET image denoising.
• To leverage redundant information from multiple PET reconstructions to improve noise suppression and signal preservation.

Main Methods:

• The MR-NLM approach utilized twelve auxiliary PET images reconstructed with varying iterations and subsets, alongside the target image.
• For each voxel, patches from auxiliary images at the same location were used for denoising, bypassing the exhaustive search of conventional NLM.
• Performance was evaluated against conventional NLM, Gaussian, and bilateral filtering using a Jaszczak phantom and 25 clinical PET/CT studies.

Main Results:

• MR-NLM filtering significantly improved signal-to-noise ratio (SNR) in phantom studies (28.8) compared to Gaussian (27.9) and conventional NLM (27.9).
• Quantification bias was reduced with MR-NLM (31.1%) and NLM (32.0%) compared to Gaussian filtering (35.4%).
• Clinical studies showed reduced bias in malignant lesions using MR-NLM (-2.2 ± 1.2%) compared to NLM (-3.5 ± 1.3%) and Gaussian (-12.3 ± 2.3%).

Conclusions:

• MR-NLM filtering demonstrates superior performance in noise reduction and signal preservation for PET images, leading to higher SNRs.
• The method offers improved quantitative accuracy, particularly in clinical applications.
• Further research is needed to address the computational overhead associated with multiple PET reconstructions.