Learning-Based MRF Reconstruction Eliminates Dictionary Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current magnetic resonance fingerprinting (MRF) reconstruction processes are inefficient due to reliance on dictionary matching, which is time-consuming and prone to quantization errors, and struggles with aliasing artifacts and noise, especially in iterative reconstructions with finer gridded parameter spaces and complex models.
Innovation Solution
A learning-based MRF reconstruction method using machine-learning algorithms to estimate tissue parameters directly from MR signal evolutions without relying on dictionaries or dictionary matching, allowing for faster reconstruction and continuous parameter output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dictionary matching is used for tissue parameter estimation, then the reconstruction process can be implemented, but the reconstruction time becomes excessively long and quantization errors occur
Solution Approach 1:
The patent replaces the mechanical dictionary matching process with a machine learning-based direct estimation system. Instead of searching through pre-computed dictionaries (mechanical search process), the system uses trained neural networks to directly predict tissue parameters from MR signal evolutions, substituting the sequential matching mechanism with a parallel computational approach that eliminates both the time-consuming search and the quantization inherent in dictionary discretization
Solution Approach 2:
The patent performs preliminary training of machine learning models using synthetic data generated from Bloch equation simulations before actual reconstruction. This preliminary action creates pre-trained estimators that can directly map MR signals to tissue parameters without requiring runtime dictionary matching, thus resolving the time-accuracy tradeoff by preparing the estimation system in advance
2Measurement precision
If iterative reconstruction with finer gridded parameter space is used, then parameter estimation accuracy improves, but computational complexity and processing time increase significantly
Solution Approach 1:
The patent substitutes the iterative optimization process with machine learning-based direct estimation. Instead of performing repeated iterations to converge on parameter values (complex mechanical process), the system uses pre-trained neural networks that directly output parameter estimates in a single pass, dramatically reducing computational complexity while maintaining the accuracy benefits of fine-gridded parameter spaces through the training data coverage
Solution Approach 2:
The patent uses synthetic data generated from Bloch equation simulations as training copies of real MR data. These synthetic training examples cover the full parameter space with fine gridding, allowing the machine learning model to learn accurate mappings without requiring iterative reconstruction during actual processing, thus capturing the accuracy of fine gridding without the computational burden
3Measurement precision
If dictionary matching is used, then tissue parameters can be estimated, but quantization errors occur due to finite dictionary entries
Solution Approach 1:
The patent replaces the discrete dictionary matching mechanism with continuous machine learning-based estimation. Instead of forcing measurements into discrete dictionary bins (which causes quantization), the system uses neural networks to directly predict continuous parameter values, eliminating the quantization error inherent in finite dictionary approaches while maintaining measurement accuracy through the model's learned mappings
4Speed
If undersampling is used to reduce scan time, then aliasing artifacts occur that require robust matching, but dictionary matching becomes less reliable
Solution Approach 1:
The patent substitutes the matching process with direct machine learning estimation that is inherently more robust to undersampling artifacts. Instead of relying on pattern matching that degrades with aliasing and noise, the system uses neural networks trained on diverse synthetic data that includes variations accounting for undersampling effects, enabling reliable parameter estimation even when scan time is reduced through undersampling
Data Source
AI summary
A learning-based magnetic resonance fingerprinting (MRF) reconstruction method for reconstructing an MR image of a tissue space in an MR scan subject for a particular MR sequence is disclosed. The method involves using a machine-learning algorithm that has been trained to generate a set of tissue parameters from acquired MR signal evolution without using a dictionary or dictionary matching.