Spectral CT Material Decomposition Using Spatial Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing material decomposition algorithms in spectral computed tomography face inaccuracies due to beam hardening artifacts and statistical noise, especially when trying to decompose more than two materials, as they struggle to separate materials without additional energy thresholds or bins, which increases noise and errors.
Innovation Solution
A machine-learned model is trained to decompose spectral CT data, utilizing information from surrounding locations to improve accuracy, allowing for the decomposition of three or more materials by using convolutional neural networks or recurrent neural networks with long-short term memory, and dictionary embedding, reducing the impact of beam hardening and statistical noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If additional energy bins or thresholds are used to decompose three or more materials, then material decomposition capability is improved, but statistical noise increases and beam hardening artifacts worsen
Solution Approach 1:
The patent transitions from analyzing only spectral information (energy dimension) to incorporating spatial context from surrounding locations. By adding the spatial dimension to the decomposition process through machine learning models that consider neighboring voxels, the system can accurately decompose three or more materials without requiring additional energy bins, thus avoiding increased noise and beam hardening artifacts.
2Device complexity
If individual location decomposition is performed based on local measurements only, then processing simplicity is maintained, but decomposition accuracy deteriorates due to beam hardening and statistical noise
Solution Approach 1:
The patent combines information from multiple locations (surrounding voxels) to perform decomposition at a given location. By merging spatial context with spectral information through machine learning models, the system overcomes the limitations of individual location analysis while maintaining computational feasibility through efficient neural network architectures.
3Adaptability or versatility
If energy bins are narrowed to distinguish more materials, then material differentiation capability is improved, but beam hardening artifacts increase
Solution Approach 1:
The patent replaces the traditional physical approach of using narrower energy bins for material differentiation with a computational approach. Machine learning models analyze spectral shapes and patterns across a broader energy range, substituting the mechanical constraint of narrow binning with intelligent algorithms that can distinguish materials based on their unique spectral signatures without exacerbating beam hardening effects.
Data Source
Figure 1~2
Figure 3~4
AI summary
For material decomposition in medical imaging, a machine-learned model is trained to decompose. For example, spectral CT data for a plurality of locations is input, and the machine-learned model outputs the material composition. Using information from surrounding locations for the decomposition by the machine-learned model for a given location may allow for more accurate material decomposition and/or three or more material decomposition.