Neural Network Magnetic Susceptibility Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining magnetic susceptibility distributions in magnetic resonance imaging (MRI) face challenges in accurately separating internal and external magnetic susceptibility sources, leading to incorrect quantifications and errors in background field subtraction and dipole inversion.
Innovation Solution
A training method using deep convolutional neural networks to jointly train two artificial neural networks, where the first network removes the influence of external magnetic susceptibility sources and the second network performs magnetic dipole inversion, treating background field subtraction and dipole inversion as interrelated problems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional QSM methods are used to separate internal and external magnetic susceptibility sources, then the magnetic susceptibility distribution can be determined, but the separation accuracy is insufficient leading to incorrect quantifications
Solution Approach 1:
The patent segments the magnetic susceptibility sources into internal and external components, using separate neural networks to process each type. The first neural network specifically handles external sources while the second handles internal sources, allowing specialized processing that improves separation accuracy and reliability compared to traditional unified QSM methods.
Solution Approach 2:
The patent introduces neural networks as intermediary processing layers between the raw magnetic resonance data and the final susceptibility distribution. These neural networks act as mediators that learn to distinguish and separate internal from external susceptibility sources, improving the reliability of background field subtraction through learned patterns rather than traditional algorithms.
2Productivity
If deep convolutional neural networks are used to solve the QSM inverse problem, then processing speed and robustness are improved, but the complexity of the solution method increases
Solution Approach 1:
The patent divides the complex QSM inverse problem into two separate neural network tasks: one network for removing external susceptibility influences and another for dipole inversion of internal sources. This segmentation reduces the complexity of each individual network while maintaining overall processing speed and robustness through specialized processing.
Solution Approach 2:
The patent replaces traditional mechanical QSM processing algorithms with neural network-based processing. The neural networks learn optimal processing strategies from training data, substituting conventional iterative mathematical methods with learned patterns that achieve faster processing speeds and improved robustness against variations in data quality.
3Ease of operation
If background field subtraction and dipole inversion are treated as separate problems, then each can be solved independently, but errors propagate and reduce overall accuracy
Solution Approach 1:
The patent implements a feedback mechanism where the output of the first neural network (external susceptibility removal) serves as the input to the second neural network (dipole inversion). This sequential processing with feedback ensures that errors do not propagate independently but are corrected in the subsequent processing stage, improving overall quantification accuracy while maintaining operational simplicity.
Solution Approach 2:
The patent merges the background field subtraction and dipole inversion processes into a unified neural network pipeline. By combining these traditionally separate operations into an integrated system where the first network's output directly feeds the second network, the patent eliminates error propagation issues while preserving the computational advantages of staged processing.
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
This approach provides a more accurate, robust, and fast solution to the ill-posed QSM inverse problem, overcoming limitations in clinical applications related to processing speed and robustness, and achieving improved reliability and accuracy in magnetic susceptibility mapping.
Implementation Method 1
applying a trained first artificial neural network, which is trained to remove an influence of one or more external magnetic susceptibility sources, to the first phase image
Implementation Method 2
applying a trained second artificial neural network, which is trained for carrying out a magnetic dipole inversion, to an output of the first artificial neural network
Implementation Method 3
Magnetic susceptibility describes a sample induced magnetization when placed in a static magnetic field. The measurement of tissue magnetic susceptibility using magnetic resonance imaging, MRI
Data Source
AI summary
A training method for training neural networks to determine a magnetic susceptibility distribution of a sample may include: storing a simulated magnetic susceptibility map of the sample, generating a modified magnetic susceptibility map by combining an influence of one or more external magnetic susceptibility sources with the simulated magnetic susceptibility map and storing the modified magnetic susceptibility maps. The method may include generating a first training image by applying a quantitative susceptibility mapping model the modified magnetic susceptibility map and storing the first training image, applying the first neural network to the first image and a second neural network to an output of the first neural network and changing network parameters of the first and the second neural network depending on a deviation of an output of the second artificial neural network from the simulated magnetic susceptibility map.