Machine Learning Image Reconstruction Reducing Computational Overhead
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional tomographic reconstruction techniques face challenges in balancing computational efficiency, patient dose, scanning speed, image quality, and artifacts, with optimization-based methods like MBIR offering improved noise reduction and resolution but at a high computational cost.
Innovation Solution
A machine learning approach is employed to train algorithms using image pairs generated from different reconstruction methods, allowing a neural network to emulate the characteristics of computationally intensive algorithms like MBIR, reducing computational overhead while maintaining image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If optimization-based reconstruction methods like MBIR are used, then image quality and noise reduction are improved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent performs a single-pass direct reconstruction to generate an initial image estimate before applying iterative optimization. This preliminary action provides a starting point that reduces the number of iterative cycles needed, thereby maintaining image quality while decreasing computational cost and processing time
Solution Approach 2:
The patent uses the direct reconstruction image as a copy or approximation of the final desired image, and then applies limited iterative refinement to this copy rather than performing full iterative reconstruction from scratch. This approach achieves acceptable image quality with significantly reduced computational overhead
2Reliability
If full MBIR iterative optimization is applied, then noise reduction and artifact suppression are enhanced, but hardware requirements and computational resources increase
Solution Approach 1:
The patent applies a limited number of iterative optimization cycles rather than full MBIR optimization. This partial action provides sufficient noise reduction and artifact suppression for clinical purposes while avoiding the excessive computational resources and hardware complexity required for complete iterative optimization
3Productivity
If direct reconstruction methods are used, then processing speed is maintained, but noise suppression and artifact reduction capabilities are insufficient
Solution Approach 1:
The patent performs direct reconstruction as a preliminary step to quickly generate an initial image estimate, preserving processing speed. This initial reconstruction is then refined with limited iterative optimization to enhance noise suppression and artifact reduction, achieving improved image quality without sacrificing overall processing efficiency
Data Source
AI summary
The present approach relates to the training of a machine learning algorithm for image generation and use of such a trained algorithm for image generation. Training the machine learning algorithm may involve using multiple images produced from a single set of tomographic projection or image data (such as a simple reconstruction and a computationally intensive reconstruction), where one image is the target image that exhibits the desired characteristics for the final result. The trained machine learning algorithm may be used to generate a final image corresponding to a computationally intensive algorithm from an input image generated using a less computationally intensive algorithm.