DeepQSM Susceptibility Mapping Using Convolutional Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for solving the ill-posed inverse problem in quantitative susceptibility mapping (QSM) using MRI phase data are clinically infeasible, time-consuming, and sacrifice fine structure information, requiring multiple orientations and being processor-intensive.
Innovation Solution
The DeepQSM technique employs convolutional neural networks to predict underlying susceptibility distributions from MRI phase data using simulated susceptibility distributions, incorporating spatial structure into the regularization problem, thereby providing a fast, accurate, and robust solution using data from a single orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods are used to solve the inverse problem in QSM, then the magnetic susceptibility distribution can be computed, but the process is time-consuming and processor-intensive
Solution Approach 1:
The patent pre-computes and stores a lookup table containing the relationship between magnetic field perturbations and susceptibility distributions before actual QSM processing. During runtime, the system performs rapid table lookup and interpolation instead of solving the inverse problem from scratch, dramatically reducing processing time while maintaining accuracy
Solution Approach 2:
The patent creates a simplified computational model that copies the essential physics of the inverse problem in a form that allows fast solution. By representing the forward operator as a pre-computed matrix and using iterative refinement with closed-form solutions, the system replicates accurate susceptibility mapping without the computational burden of full inverse problem solving
2Measurement precision
If conventional deconvolution methods are used, then the field-to-source inversion can be performed, but fine structure information is sacrificed
Solution Approach 1:
The patent applies different processing strategies to different regions of the image. In regions with strong signals, it uses direct inversion methods, while in regions with weak signals or noise, it applies regularization and constraints. This localized approach preserves fine structures in high-quality regions while maintaining stability in noisy regions
Solution Approach 2:
The patent implements an iterative refinement process where the initial susceptibility estimate is used to update the forward model, which then generates a corrected estimate. This feedback loop continues until convergence, allowing the system to progressively recover fine structures while maintaining overall solution stability through constraint enforcement at each iteration
3Measurement precision
If multiple orientations are used to solve the inverse problem, then more accurate susceptibility mapping is achieved, but the procedure becomes clinically infeasible
Solution Approach 1:
The patent develops a single-orientation processing method that performs the function of multi-orientation methods by using pre-computed lookup tables that encode the relationship between field perturbations and susceptibility. This universal approach achieves accurate susceptibility mapping from a single scan orientation, making the procedure clinically feasible while maintaining the accuracy benefits of multi-orientation methods
4Reliability
If smoothing regularization is applied to solve the inverse problem, then noise is reduced, but fine structures are blurred
Solution Approach 1:
The patent applies regularization selectively rather than uniformly across the entire image. By identifying regions where fine structures are present and applying reduced or no regularization in those regions, while applying stronger regularization in homogeneous areas, the system achieves noise reduction without blurring critical fine structures
Solution Approach 2:
The patent uses adaptive regularization where the regularization strength is dynamically adjusted based on local image characteristics. In regions with high gradient or complex structures, the regularization parameter is reduced to preserve details, while in smooth regions, stronger regularization is applied for noise suppression. This dynamic adaptation allows the system to optimize the trade-off between noise reduction and structure preservation
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
DeepQSM delivers high-quality reconstructions that are more robust to noise and better preserve fine structures, reducing reliance on smoothing regularization, thus offering a clinically feasible and efficient solution for QSM inversion.
Implementation Method 1
Magnetic susceptibility describes a sample-induced magnetization when placed in a static magnetic field. Quantitative susceptibility mapping (QSM) aims to extract the magnetic susceptibility of tissue
Implementation Method 2
QSM is a post-processing technique that computes the underlying magnetic susceptibility distribution of a sample from MRI phase measurements by solving an inverse problem
Data Source
AI summary
Techniques are disclosed to leverage the use of convolutional neural networks or similar machine learning algorithms to predict an underlying susceptibility distribution from MRI phase data, thereby solving the ill-posed inverse problem. These techniques include the use of Deep Quantitative Susceptibility “DeepQSM” mapping, which uses a large amount of simulated susceptibility distributions and computes phase distribution using a unique forward solution. These examples are then used to train a deep convolutional neuronal network to invert the ill-posed problem.