Semi-Supervised GAN for MRF Dictionary Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for generating dictionaries for Magnetic Resonance Fingerprinting (MRF) are time-consuming, especially when complex physics are involved, requiring days to weeks for calculations.
Innovation Solution
A semi-supervised learning system, specifically a Generative Adversarial Network (GAN), is used to simulate signal evolutions, significantly reducing the time needed to generate an MRF dictionary by training on a set of MRF data and control variables, allowing for on-the-fly generation of dictionaries with tissue properties of interest.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional physics simulations are used to generate MRF dictionaries, then accuracy is improved, but time required increases significantly
Solution Approach 1:
The patent uses a Generative Adversarial Network (GAN) to create synthetic signal evolutions that copy the statistical properties and characteristics of real MRF signals. The generator network learns to produce dictionary entries that are indistinguishable from those generated by complex physics simulations, thereby achieving high accuracy without the computational burden of traditional methods.
Solution Approach 2:
The patent replaces the mechanical computation process of physics-based simulations with a neural network-based generative model. Instead of solving differential equations and simulating spin dynamics step-by-step, the GAN learns the underlying patterns and generates signal evolutions directly, substituting computational mechanics with learned patterns.
2Adaptability or versatility
If complex physics simulations are included in dictionary calculation, then completeness of tissue characterization is improved, but calculation time increases to days or weeks
Solution Approach 1:
The patent performs preliminary training of the GAN model on a dataset of physics-simulated signals to learn the relationships between tissue parameters and signal evolutions. Once trained, the model can rapidly generate complete tissue characterizations for new parameters without performing full physics simulations, enabling versatile tissue characterization at high speed.
Solution Approach 2:
The patent changes the approach from computing each dictionary entry by solving physics equations with specific parameters to using a trained neural network that can rapidly evaluate any parameter combination. The GAN learns the parameter-to-signal mapping and can generate signal evolutions for any tissue parameters within the training range instantaneously.
Data Source
AI summary
A method for creating a dictionary for a magnetic resonance fingerprinting (MRF) reconstruction includes training a semi-supervised learning system based on at least a set of MRF data and a set of control variables and generating a plurality of signal evolutions using the trained semi-supervised learning system. The method also includes generating an MRF dictionary using the plurality of signal evolutions generated using the trained semi-supervised learning system and storing the MRF dictionary in a memory. In one embodiment, the semi-supervised learning system is a MRF generative adversarial network (GAN).