Spectral X-Ray Material Decomposition With Dual-Model Frequency Split
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based material decomposition methods in spectral X-ray imaging suffer from bias-variance trade-offs, leading to inaccurate results due to high noise and bias, which conventional statistical estimation algorithms like MLE cannot adequately address.
Innovation Solution
Employ two AI models with differing bias-variance characteristics, one with low bias and high variance, and the other with high bias and low variance, and apply low-pass and high-pass filtering respectively, followed by a linear combination of their outputs to achieve low bias and low noise in material decomposition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single AI model is used for material decomposition, then computational efficiency is improved, but accuracy deteriorates due to the bias-variance trade-off
Solution Approach 1:
The patent divides the single AI model into two separate AI models with different configurations. The first model is optimized for low bias (e.g., deeper network, different loss function weighting) while the second model is optimized for low variance (e.g., regularized, different architecture). This segmentation allows each model to excel at different aspects of the decomposition problem, and their results are combined to achieve both accuracy and efficiency.
Solution Approach 2:
The patent creates a composite decomposition result by combining the outputs of two different AI models. Rather than relying on a single model's output, the system integrates the predictions from both models (e.g., through averaging or weighted combination), creating a composite solution that leverages the strengths of each individual model to overcome their respective weaknesses.
2Measurement precision
If MLE algorithm is used for material decomposition, then accuracy is improved, but computational resource consumption increases
Solution Approach 1:
The patent replaces the iterative optimization mechanism of MLE with a direct neural network inference approach. Instead of performing repeated numerical optimization steps that consume significant computational resources, the system uses pre-trained AI models that can provide decomposition results through forward propagation, significantly reducing computational energy while maintaining accuracy through the dual-model strategy.
3Measurement precision
If AI model with low bias is used, then systematic accuracy is improved, but noise in results increases due to higher variance
Solution Approach 1:
The patent segments the decomposition task into two specialized models: one focused on minimizing bias (systematic error) and another on minimizing variance (noise). The first model uses configurations that reduce systematic errors (e.g., appropriate loss functions, training strategies) while the second model uses configurations that reduce output variability (e.g., regularization, ensemble techniques). By combining their outputs, the system achieves both low bias and low variance.
Solution Approach 2:
The patent introduces a combination mechanism as an intermediary between the two AI models and the final result. This intermediary step (e.g., averaging, weighted fusion) acts as a mediator that balances the high-variance low-bias model output with the low-variance high-bias model output, producing a final result that has both properties moderated to acceptable levels.
Data Source
AI summary
A method for material decomposition of an object based on spectral X-ray scan data for the object and based on application of a frequency split approach. The method comprises using two AI models in parallel to perform the material decomposition analysis based on input spectral X-ray data, wherein the models are configured such that one exhibits higher bias and lower variance (lower noise) than the other. The input spectral X-ray data is fed to both models. The output material composition data from the low bias model is low-pass filtered and the output material composition data from the low variance model is high pass filtered. The outputs from the two models are linearly combined, either before the filtering or after. The resulting combined material decomposition data has both lower bias and lower noise compared to the output generated if just one AI model were to be used.