GAN Artifact Reduction for Undersampled MRI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI techniques that allow patients to breathe freely during scans, such as self-gated 3D stack-of-radial MRI, require long acquisition times due to the need for numerous radial spokes, making them impractical for clinical use, and existing acceleration methods like parallel imaging and compressed sensing are computationally intensive or limited by hardware.
Innovation Solution
A computer-implemented method using a Generative Adversarial Network (GAN) to remove artifacts from undersampled MRI data by training a generator model with adversarial loss, L2 loss, and structural similarity index measure loss, allowing for faster acquisition times without compromising image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If free-breathing MRI techniques are used to allow patients to breathe during scans, then patient comfort and compliance are improved, but acquisition time increases significantly
Solution Approach 1:
The system performs preliminary actions by acquiring respiratory signal data during the MRI scan to enable subsequent motion correction and artifact reduction processing, allowing free-breathing acquisition without compromising image quality
Solution Approach 2:
The system changes the sampling parameters by using non-Cartesian radial trajectories with variable density sampling, allowing flexible adjustment of sampling rates to balance acquisition time and image quality while accommodating free-breathing conditions
2Manufacturing precision
If the number of radial spokes is increased to reduce respiratory motion artifacts, then image quality is improved, but acquisition time increases
Solution Approach 1:
The system applies partial action by acquiring fewer radial spokes than traditionally required, then uses deep learning-based artifact reduction to compensate for the reduced sampling, achieving acceptable image quality with shorter acquisition time
Solution Approach 2:
The system introduces an intermediary processing step using deep neural networks that acts as a mediator between the undersampled k-space data and the final image reconstruction, removing streaking artifacts without requiring full sampling
3Loss of time
If parallel imaging is used to accelerate acquisition, then acquisition time is reduced, but hardware arrangement limits the acceleration factor
Solution Approach 1:
The system replaces the mechanical hardware-based parallel imaging approach with a software-based deep learning system that processes undersampled data, eliminating hardware constraints on acceleration factors and enabling higher flexibility in sampling strategies
4Productivity
If compressed sensing is used to accelerate acquisition, then higher acceleration rates are achieved, but computational intensity and parameter tuning complexity increase
Solution Approach 1:
The system uses a trained deep neural network model that can be deployed once and then rapidly processes multiple scans with minimal computational overhead, replacing the iterative reconstruction process of compressed sensing with a more efficient forward-pass inference process
Data Source
AI summary
Systems and methods for generative adversarial networks (GANs) to remove artifacts from undersampled magnetic resonance (MR) images are described. The process of training the GAN can include providing undersampled 3D MR images to the generator model, providing the generated example and a real example to the discriminator model, applying adversarial loss, L2 loss, and structural similarity index measure loss to the generator model based on a classification output by the discriminator model, and repeating until the generator model has been trained to remove the artifacts from the undersampled 3D MR images. At runtime, the trained generator model of the GAN can be generate artifact-free images or parameter maps from undersampled MRI data of a patient.