Deep Learning Network for CT Image Quality via Physical Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CT image reconstruction methods face a trade-off between computational speed and image quality, with analytical methods being fast but providing lower quality, and iterative reconstruction methods offering better quality but at the cost of increased time and computation.
Innovation Solution
A deep learning neural network is trained using physical-model information to filter initial CT images reconstructed with analytical methods, aiming to achieve high image quality comparable to iterative reconstruction methods while maintaining computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If iterative reconstruction (IR) methods are used, then image quality is improved, but computation time and computational resources increase significantly
Solution Approach 1:
The patent performs preliminary action by training a deep learning neural network offline using pairs of analytical and iterative reconstruction images. The trained network captures the quality improvement of iterative methods in advance, enabling fast inference during actual CT reconstruction without performing computationally intensive iterative reconstruction in real-time.
Solution Approach 2:
The patent creates a copy of the iterative reconstruction quality characteristics by training the neural network to learn the mapping from analytical to iterative reconstruction images. The network replicates the quality improvement effects of iterative methods through learned patterns rather than actual iterative computation, providing a computationally efficient approximation.
2Productivity
If analytical reconstruction methods are used, then computation speed is maintained, but image quality deteriorates
Solution Approach 1:
The patent introduces a deep learning neural network as an intermediary between analytical reconstruction and the final high-quality image. The network takes analytical reconstruction output as input and transforms it into an image with quality comparable to iterative reconstruction, mediating between the speed advantage of analytical methods and the quality advantage of iterative methods.
Solution Approach 2:
The patent changes the parameter space by training the neural network on pairs of analytical and iterative reconstruction images with varying parameters (noise levels, anatomical structures, pathologies). This enables the network to adaptively improve image quality across different scanning conditions while maintaining the computational efficiency of analytical methods.
Data Source
AI summary
A method and apparatus is provided that uses a deep learning (DL) network to improve the image quality of computed tomography (CT) images, which were reconstructed using an analytical reconstruction method. The DL network is trained to use physical-model information in addition to the analytical reconstructed images to generate the improved images. The physical-model information can be generated, e.g., by estimating a gradient of the objective function (or just the data-fidelity term) of a model-based iterative reconstruction (MBIR) method (e.g., by performing one or more iterations of the MBIR method). The MBIR method can incorporate physical models for X-ray scatter, detector resolution/noise/non-linearities, beam-hardening, attenuation, and/or the system geometry. The DL network can be trained using input data comprising images reconstructed using the analytical reconstruction method and target data comprising images reconstructed using the MBIR method.