Neural Network MRI Reconstruction for Low-Field Signal Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low-field MRI systems face challenges in producing high-quality images due to low signal-to-noise ratio and the need for longer scan times, as they often lack techniques for effectively utilizing fully sampled k-space data.
Innovation Solution
A method is developed to process fully sampled k-space MRI imaging data using a neural network trained on undersampled k-space data, allowing for the reconstruction of high-quality images by matching the input dimensions of the neural network with subsets of fully sampled data, thereby enabling the use of existing neural networks for both undersampled and fully sampled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fully sampled k-space data is acquired using low-field MRI systems, then image quality can be improved, but scan time increases
Solution Approach 1:
The neural network is pre-trained using undersampled k-space data during the training phase. During actual scanning, the network can process fully sampled data by simply removing the downsampling operation from the pipeline, leveraging the pre-learned features to maintain high image quality without requiring additional training or changing the acquisition protocol
Solution Approach 2:
The neural network architecture is designed to be universal and can process both undersampled and fully sampled k-space data. The same network model can reconstruct images from different sampling densities by adjusting the input data preparation, eliminating the need for separate networks for different acquisition scenarios
2Productivity
If parallel imaging techniques are used to accelerate MRI reconstruction, then scan time is reduced, but image quality deteriorates due to aliasing artifacts
Solution Approach 1:
The patent converts the harmful aliasing artifacts introduced by parallel imaging into beneficial training data for the neural network. During training, the network learns to recognize and correct aliasing patterns by processing undersampled data, enabling it to reconstruct high-quality images from accelerated acquisitions without requiring additional averaging
3Measurement precision
If additional averages are performed to increase SNR, then image quality improves, but scan time increases
Solution Approach 1:
The patent replaces the mechanical approach of performing multiple physical averages during data acquisition with a computational approach using neural networks. Instead of acquiring the same data multiple times to average out noise, the network learns to denoise and enhance SNR from single acquisitions by processing the underlying signal patterns and statistical characteristics during training
Data Source
AI summary
In a method for processing fully sampled k-space MRI imaging data associated with a tissue of interest within a FOV, a neural network may be trained using undersampled k-space MRI imaging data associated with the tissue of interest. At least one subset of the fully sampled k-space MRI imaging data may be obtained based on an input dimension of the trained neural network such that a dimension of each one of the at least one subset is the same as the input dimension. Each one of the at least one subset of the fully sampled k-space MRI imaging data may be processed by the trained neural network, respectively. Spatial domain MRI imaging data associated with the tissue of interest within the FOV may be accordingly determined based on corresponding output of the trained neural network.