Physics-Constrained Neural Network Training for Emission Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial neural networks are trained without incorporating a priori knowledge or physical constraints, leading to an unsuitable solution space for certain inputs, particularly in applications like emission imaging where accurate attenuation correction is needed without additional radiation exposure.
Innovation Solution
The training of neural networks is constrained by known physical phenomena, using a physical constraint evaluator to minimize errors based on both output and physical characteristics, combining losses to iteratively modify network weights and structure, ensuring outputs conform to physical reality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional neural network training is used without physical constraints, then the network can be trained quickly with simple loss minimization, but the solution space becomes unsuitable for certain inputs and applications
Solution Approach 1:
The patent changes the training parameters by introducing physics-based constraints and custom loss functions that incorporate domain-specific knowledge. This transforms the standard loss minimization approach into a constrained optimization problem that guides the network toward physically meaningful solutions, improving reliability for specific applications like emission imaging.
Solution Approach 2:
The training process is segmented into multiple components: a standard loss function for general accuracy and separate physics-based constraint terms for domain-specific correctness. This segmentation allows the network to learn both general patterns and physical consistency independently, resolving the contradiction between simplicity and reliability.
2Measurement precision
If additional CT scans are performed to obtain accurate attenuation maps, then image accuracy improves, but radiation exposure increases
Solution Approach 1:
The patent introduces a neural network as an intermediary that learns to map emission data directly to attenuation maps without requiring additional CT scans. The network is trained with physics-based constraints to ensure the generated attenuation maps are physically consistent, thereby achieving accurate attenuation correction while avoiding the harmful radiation exposure of additional CT imaging.
Solution Approach 2:
The patent replaces the mechanical imaging system (CT scanner) with a computational system (neural network). Instead of physically scanning the patient with ionizing radiation, the network computationally generates attenuation maps from existing emission data, substituting a harmful physical process with a safe computational one while maintaining measurement precision.
Data Source
AI summary
A system and method includes input of a plurality of sets of training data to a neural network to generate a plurality of sets of output data, determination of a first loss based on the plurality of sets of output data and on the plurality of sets of ground truth data, determination if a second loss based on the plurality of sets of output data and one or more physics-based constraints, and modification of the neural network based on the first loss and the second loss.