Neural Network Enhanced Tomographic Reconstruction for Low Dose Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current imaging technologies in computed tomography, cone beam computed tomography, and tomosynthesis face challenges in achieving high-quality images with low dose X-ray irradiation, as they often result in artifacts and trade-offs between acquisition speed and resolution, leading to suboptimal image quality due to limitations in pixel resolution and readout speed.
Innovation Solution
A method combining iterative reconstruction with a trained neural network to enhance image quality, utilizing a Convolutional Neural Network (CNN) trained on high and low quality tomographic image data to improve noise content, resolution, and reduce artifacts, leveraging prior information and data redundancy to overcome resolution limitations and artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional iterative reconstruction with regularization is used, then image quality can be improved, but the process is cumbersome to tune and may delete real image content
Solution Approach 1:
The patent replaces the manual regularization tuning process with an automated deep learning system. The neural network learns optimal reconstruction parameters and image enhancement strategies from training data, eliminating the need for manual regularization parameter adjustment while preserving real image content through learned priors from the training dataset.
2Measurement precision
If pixel resolution is increased to improve image quality, then readout speed decreases due to the trade-off in flat panel detector capabilities
Solution Approach 1:
The patent performs preliminary actions by training the neural network offline on high-resolution image data. During actual operation, the pre-trained network rapidly processes lower-resolution input images, generating high-resolution output without requiring the detector to physically resolve fine details, thus maintaining fast readout speeds while achieving high effective resolution.
Solution Approach 2:
The neural network creates a computational copy or super-resolution version of the input image that exceeds the physical resolution limits of the detector. The network learns the mapping from low-resolution to high-resolution images, effectively copying fine detail information that was not captured by the detector but can be inferred from training data.
3Productivity
If acquisition speed is increased to reduce motion artifacts, then resolution is reduced due to the trade-off in imaging systems
Solution Approach 1:
The patent substitutes the physical imaging system's resolution limits with a computational enhancement system. The neural network compensates for the lower resolution caused by faster acquisition by learning from training data how to reconstruct high-resolution images from lower-resolution inputs, effectively decoupling acquisition speed from final image resolution.
4Object-affected harmful factors
If low dose X-ray irradiation is used, then image quality deteriorates with increased artifacts and noise
Solution Approach 1:
The patent converts the harmful effect of low-dose imaging (increased noise and artifacts) into a beneficial learning opportunity. The neural network is trained on pairs of low-dose and high-dose images, learning to recognize and correct noise patterns and artifacts. During operation, it applies these learned corrections to low-dose images, recovering image quality that would normally require higher doses.
Solution Approach 2:
The patent replaces the physical need for high X-ray doses to achieve good image quality with a computational correction system. The neural network learns the statistical characteristics of noise and artifacts in low-dose images and applies learned transformations to remove these degradations, substituting computational processing for physical dose increase.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The method effectively enhances image quality by improving resolution and reducing artifacts, achieving superior performance compared to traditional interpolation methods and other non-deep learning-based approaches, particularly in medical imaging where diagnostic accuracy is critical.
Implementation Method 1
A method combining iterative reconstruction with a trained neural network to enhance image quality, utilizing a Convolutional Neural Network (CNN) trained on high and low quality tomographic image data to improve noise content, resolution, and reduce artifacts
Data Source
AI summary
Tomographic images acquired by iterative reconstruction of low quality projection images, are enhanced by the steps of correcting at an iteration step the result of the previous iteration step by means of a back-projection of the result of a comparison of a projection image and the forward projection of the result of the previous iteration step whereby this result is enhanced by subjecting it to a trained neural network.