Program Information
Adaptive Non-Local Means Image Denoising with Local Similarity Characterization
J Chappelow*, P Jordan , J Shea , E Chao , B Harstad , C Maurer , Accuray Incorporated, Madison, WI
Presentations
WE-RAM3-GePD-I-1 (Wednesday, August 2, 2017) 10:30 AM - 11:00 AM Room: Imaging ePoster Lounge
Purpose: To design a volumetric denoising method that preserves spatial resolution without the need for case-specific parameter selection. A standard non-local means (NLM) filter is challenging to apply on a broad range of imaging modalities and anatomies, as selection of acceptable parameters is not straightforward. A locally adaptive algorithm that practically eliminates the need for parameter selection, and provides improved quality of the filtered images with no loss in spatial resolution, is developed.
Methods: The standard NLM algorithm combines similar patches or small regions around neighboring pixels via a weighted average, and requires careful selection of a noise level parameter and search range. In our adaptive NLM algorithm, the maximum number of neighbors to combine, N, is selected on a per-pixel basis by a process of local similarity characterization using a pixel’s immediate neighbors. This contrasts with usual approaches where N is fixed. Local similarity characterization at a pixel involves using the similarity of the k-th most similar pixel in a small neighborhood as a threshold for inclusion in the weighted average, thus dynamically setting N. This spatially-varying value reflects several intrinsic local qualities, including noise and structural complexity.
Results: Tests on clinical and phantom data demonstrate broadly applicable parameters, unlike standard NLM, and superior performance to anisotropic diffusion (AD) in terms of noise reduction and edge preservation. Single-blind evaluation of TomoTherapy MVCT quality was performed with adaptive NLM and AD. Adaptive NLM was favored over AD by 75% of viewers on pelvic and abdominal images. Gains in detectability of low-contrast targets were equivalent to imaging with 2x dose. With no change in filter parameters, spatial resolution was improved, while maintaining noise reduction, by using a high frequency-boosting filter during MVCT reconstruction.
Conclusion: Our algorithm provides easily-configured denoising without edge degradation, resulting in improved spatial resolution and low-contrast object detectability.
Contact Email: