Magnetic Particle Imaging Neural Reconstruction With Forward-Model Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing magnetic particle imaging (MPI) reconstruction methods suffer from significant noise interference and require large amounts of training data, which is scarce due to the high cost of equipment, leading to poor reconstruction quality and limited neural network training.
Innovation Solution
A magnetic particle imaging reconstruction method using a neural network constrained by a forward model, incorporating a total variation regularization term and adjusted regularization parameters, is employed to improve reconstruction quality, utilizing a system matrix and single data set for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If system matrix reconstruction method is used, then reconstruction accuracy is improved, but noise interference increases and artifacts remain
Solution Approach 1:
The patent introduces a neural network as an intermediary component between the system matrix reconstruction method and the final reconstruction result. The neural network is trained to recognize and filter noise patterns while preserving accurate structural information, thereby mediating between the high accuracy of system matrix methods and the noise reduction need.
Solution Approach 2:
The patent employs parameter changes by adjusting regularization parameters in the reconstruction algorithm and optimizing neural network parameters through training. By dynamically adjusting these parameters, the system achieves optimal balance between reconstruction accuracy and noise suppression for different imaging conditions.
2Measurement precision
If traditional neural network training is used, then reconstruction quality can be improved, but large amounts of training data are required which are scarce
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network on simulated magnetic particle imaging data before deploying it with limited real measurement data. This pre-training phase establishes initial weight configurations that capture general reconstruction patterns, enabling effective fine-tuning with scarce real data.
Solution Approach 2:
The patent uses copying by generating synthetic training data through simulation that replicates the characteristics of real magnetic particle imaging data. These simulated copies serve as training samples, allowing the neural network to learn reconstruction patterns without requiring large quantities of expensive real measurement data.
Data Source
AI summary
In a magnetic particle imaging reconstruction method based on a neural network constrained by a forward model, a system matrix is obtained through calibration, voltage data generated by a specimen is measured, collected data are transformed into a frequency-domain by Fourier transform, frequency features of data are screened by a signal-to-noise ratio threshold, a reconstruction network is built by a Pytorch to realize mapping from one-dimensional voltage data to a multi-dimensional magnetic particle concentration distribution, the system matrix is used as the forward model of magnetic particle imaging, and simulated voltage data is generated according to a reconstructed multi-dimensional magnetic particle concentration distribution, a difference between the simulated voltage data and input voltage data is calculated as a loss function for network parameter updating, a total variation regularization term is added to the loss function, and training parameters and regularization parameters are adjusted to achieve an optimal reconstruction effect.