Neural Network SPECT Reconstruction Physics Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current SPECT image reconstruction techniques, such as iterative reconstruction and deep learning, are computationally intensive and do not adequately incorporate the physics of image formation, leading to inefficiencies in processing and accuracy.
Innovation Solution
Configuring neural networks based on the physics of image formation and optimizing them using test image data to reduce computational requirements and improve reconstruction accuracy, allowing for faster gradient descent reconstruction and native operation on GPUs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative reconstruction is used to model image formation as a linear operator, then reconstruction accuracy is improved, but computational time and memory requirements increase significantly
Solution Approach 1:
The patent segments the image formation process into discrete physical operations (rotation, attenuation, flood correction, etc.) that can be represented as separate neural network layers. This segmentation allows the system to process each physical effect independently rather than computing the entire projection operator, significantly reducing computational time while maintaining reconstruction accuracy.
Solution Approach 2:
The patent replaces the traditional mechanical iterative reconstruction system with a neural network-based system. By substituting the linear operator approach with a trained neural network model, the system achieves the same reconstruction accuracy without the computational burden of iterative calculations, effectively replacing a computationally intensive mechanical process with a more efficient information-processing approach.
2Measurement precision
If iterative reconstruction is used to model image formation as a linear operator, then reconstruction accuracy is improved, but memory requirements increase due to storing probability lists
Solution Approach 1:
The patent extracts and removes the need to store large probability lists by representing image formation effects as a compact sequence of neural network layers. Each layer encodes a specific physical operation, allowing the system to achieve the same reconstruction accuracy without storing the extensive probability data required by traditional iterative methods.
Solution Approach 2:
The patent changes the parameter representation from storing detailed probability lists to using compact neural network weight parameters. This parameter transformation reduces memory requirements while preserving the essential information needed for accurate reconstruction, as the neural network weights efficiently encode the same physical transformations.
3Productivity
If deep learning is used for SPECT reconstruction with generic neural networks, then processing speed is improved, but reconstruction accuracy deteriorates due to lack of physics-based modeling
Solution Approach 1:
The patent applies local quality by making each neural network layer specialized for a specific physical effect (e.g., one layer for rotation, another for attenuation). This localized specialization ensures that each part of the network accurately models its corresponding physical process, improving overall reconstruction accuracy while maintaining the computational efficiency of the deep learning approach.
Solution Approach 2:
The patent performs preliminary action by training the neural network with physics-based constraints and test image data before actual reconstruction. This pre-training with physical knowledge embedded in the network architecture ensures that the model understands the underlying physics of image formation, improving accuracy while maintaining fast processing speeds during actual use.
Data Source
AI summary
Systems and methods for image reconstruction based on modeling image formation as one or more neural networks. In accordance with one aspect, one or more neural networks are configured based on physics of image formation (202). The one or more neural networks are optimized using acquired test image data (204). An output image may then be reconstructed by applying current image data as input to the one or more optimized neural networks (208).