Transformer Network for Undersampled MRI Artifact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing MRI reconstruction methods struggle with artifact reduction and image quality when dealing with undersampled radial k-space data, particularly due to the time-consuming gridding process and difficulty in learning to compensate for gridding errors in non-Cartesian trajectories.
Innovation Solution
The implementation of a transformer network that directly predicts unacquired radial k-space spokes, rearranging radial k-space data into sequential time-series data, and utilizing attention mechanisms to learn dependencies between arbitrary locations of radial spokes, thereby avoiding the need for repeated gridding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If gridding is performed on non-Cartesian k-space data before Fourier transformation, then the data can be processed using standard Fourier transforms, but the gridding process is time-consuming and must be repeated with each iteration of network training and testing
Solution Approach 1:
The patent performs gridding of non-Cartesian k-space data onto a Cartesian basis before network training, so that the gridding operation is done only once rather than repeatedly during each training and testing iteration. This preliminary preparation eliminates the time-consuming repeated gridding computations while maintaining compatibility with standard Fourier transform operations.
2Loss of time
If k-space data is gridded before network training, then repeated gridding is avoided, but networks have difficulty learning to compensate for gridding errors
Solution Approach 1:
The patent implements a feedback mechanism where the network is trained to explicitly learn and compensate for gridding errors. The training process uses ground truth fully-sampled k-space data to provide feedback signals, enabling the network to learn the inverse of gridding operations and correct artifacts introduced by the gridding process, thereby maintaining high image fidelity.
3Loss of time
If undersampling is used to accelerate MRI acquisition, then acquisition time is reduced, but artifacts and degraded image quality are produced
Solution Approach 1:
The patent introduces a deep learning-based intermediary reconstruction network that acts as a mediator between the undersampled k-space data and the final image reconstruction. This network learns to infer missing k-space information and remove artifacts by leveraging patterns learned from fully-sampled training data, thereby recovering image quality that would otherwise be degraded by undersampling.
4Ease of operation
If existing transformer techniques crop and reshape image patches to sequential vectors, then the images can be processed by transformer networks, but the global streaking artifacts are cut off and the network becomes less effective at removing them
Solution Approach 1:
The patent processes k-space data rather than image space data, working in the frequency domain where radial spokes have a different structural organization. By operating on the radial spoke structure in k-space and using attention mechanisms that respect the radial geometry, the network can capture global dependencies across the entire field of view without cropping, thereby effectively removing streaking artifacts while maintaining compatibility with transformer architecture.
Data Source
AI summary
A system and methods for recreating images from undersampled magnetic resonance imaging (MRI) image data to remove artifacts from the undersampling. Using trained transformer networks, undersampled k-space data can be acquired and the networks can predict additional projected radial MRI k-space data from the undersampled radial MRI k-space data. Images with fewer artifacts can then be generated.