CT Noise Reduction via Edge-Preserving and Morphological Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
C-arm computed tomography datasets suffer from high noise levels, which limit soft tissue contrast and spatial resolution, making it difficult to achieve high-contrast images without compromising bone structures or spatial resolution.
Innovation Solution
A combination of edge-preserving and morphological non-linear filters is applied in a post-processing step to the reconstructed computed tomography dataset, using a guided filter for high-contrast edges and a bitonic filter for other areas, with a weighting mechanism to balance noise reduction and edge preservation, ensuring optimal soft tissue contrast and spatial resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If a smoothing reconstruction kernel is used to reduce noise, then noise level decreases and soft tissue contrast improves, but spatial resolution is reduced and bone structures become blurred
Solution Approach 1:
The patent segments the image processing into multiple stages: first applying a smoothing reconstruction kernel to reduce noise, then applying a non-linear edge-preserving filter to restore edges and fine structures. This multi-stage segmentation allows each processing step to optimize for its specific function without compromising the other.
Solution Approach 2:
The non-linear filter applies different processing characteristics to different regions of the image: in homogeneous regions it applies stronger smoothing to reduce noise, while at edges and boundaries it preserves or enhances contrast to maintain spatial resolution. This local adaptation resolves the contradiction between noise reduction and edge preservation.
2Object-affected harmful factors
If slice thickness is increased to reduce noise, then noise level decreases, but spatial resolution and detail visibility are reduced
Solution Approach 1:
Instead of using mechanical means (increasing slice thickness) to reduce noise, the patent substitutes a computational approach: applying non-linear filters with edge-preserving properties that can reduce noise while maintaining fine structural details through intelligent pixel value adjustment based on local image characteristics.
3Object-affected harmful factors
If non-linear filters developed for photographs are applied to computed tomography datasets, then noise reduction occurs, but image quality deteriorates due to amplified artifacts and loss of high-contrast details
Solution Approach 1:
The patent modifies the parameters and characteristics of non-linear filters specifically for computed tomography applications. The filter design incorporates understanding of CT image characteristics (high dynamic range, bone-soft tissue contrast) to adjust filtering strength, edge detection thresholds, and noise models, thereby achieving noise reduction without the artifacts that plague photographic filters applied to medical imaging.
Data Source
AI summary
A method for noise reduction in a three-dimensional computed tomography dataset, which is reconstructed from two-dimensional projection images recorded with an x-ray device using different recording geometries, is provided. In a post-processing section following the reconstruction of the computed tomography dataset, to obtain a first intermediate dataset, a first, edge-preserving filter is applied to the reconstructed computed tomography dataset. To obtain a second intermediate dataset, a second, morphological filter is applied to the reconstructed computed tomography dataset. A first weighting dataset weighting edges more strongly is established from a subtraction dataset of the first intermediate dataset and the second intermediate dataset. A noise-reduced result dataset is established as a weighted sum of the first intermediate dataset and the second intermediate dataset. The first intermediate dataset is weighted with the first weighting dataset, and the second intermediate dataset is weighted with one minus the first weighting dataset.