Non-Local Means Filtering for Robust PET Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional non-local means (NLM) filtering methods for noisy PET images are not robust across different patients or organs, suffer from inaccurate similarity measurements due to high noise levels, and can introduce artifacts like artificial texture, especially when dealing with large intensity ranges and varying organ intensities.
Innovation Solution
The proposed solution improves NLM filtering by using similarity measurements based on image patches in a transform domain, incorporating channelized patch comparisons, and employing the Kullback-Leibler distance measure instead of Euclidean distance. Additionally, anatomical images are utilized to refine similarity weights and create mixed similarity weights, and the filtering process can be combined with local filtering methods for enhanced noise suppression and image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional non-local means filtering is applied to noisy PET images, then noise reduction is achieved, but robustness across different patients or organs deteriorates
Solution Approach 1:
The patent transforms the image data into a different domain (transform domain) where the filtering operation is performed. This parameter transformation allows the filter to operate on transformed coefficients rather than raw pixel values, making it adaptable to varying intensity levels across different patients and organs while maintaining noise reduction effectiveness.
Solution Approach 2:
The patent divides the image into patches and processes them independently through transformation and filtering operations. This segmentation approach allows local adaptation to different tissue types and intensity ranges while applying a consistent filtering methodology across the entire image, improving both robustness and noise reduction.
2Reliability
If conventional NLM filtering is applied to high noise level images, then noise reduction is achieved, but measurement precision deteriorates due to inaccurate similarity measurements
Solution Approach 1:
The patent replaces the conventional Euclidean distance-based similarity measurement with a transform-domain comparison approach. By operating in the transformed domain, the similarity measurement becomes more robust to noise, as the transformation separates signal from noise components, allowing for more accurate patch matching even in high noise conditions.
3Reliability
If conventional NLM filtering is applied to images with large intensity ranges, then noise reduction is achieved, but artifact generation increases due to artificial texture
Solution Approach 1:
The patent applies filtering in the transform domain rather than the spatial domain. This parameter change in the operating domain allows for more controlled noise suppression that preserves natural image textures. The transformation separates noise from structural information, enabling noise reduction without generating artificial texture artifacts that plague conventional spatial-domain filters.
Data Source
AI summary
An apparatus for performing a non-local means (NLM) filter is described. The pixel of the NLM-filtered image are weighted averages of pixels from a noisy image, where the weights are a measure of the similarity between patches of the noisy image. The similarity weights can be calculated using a Kullback-Leibler or a Euclidean distance measure. The similarity weights can be based on filtered patches of the noisy image. The similarity weights can be based on a similarity measure between patches of an anatomical image corresponding to the noisy image. The similarity weights can be calculated using a time series of noisy images to increase the statistical sample size of the patches. The similarity weights can be calculated using a weighted sum of channel similarity weights calculated between patches of noisy image that have been band-pass filtered. The NLM-filtered image can also be blended with a non-NLM-filtered image.