MRI K-Space Augmentation for Motion-Corrected Image Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Motion artifacts in magnetic resonance imaging (MRI) due to subject movement during data acquisition lead to image blurring and artifacts, which existing motion correction techniques like PROPELLER, navigator-based, and self-navigation methods struggle to fully address, especially in reconstructing images with varying tissue contrast modalities.
Innovation Solution
An image reconstruction method that utilizes a magnetic resonance modality conversion neural network to generate synthetic k-space data, augmenting sparse regions in motion-corrected k-space data with synthetic data of the desired tissue contrast modality, enhancing image quality through techniques like compressed sensing and optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If motion correction techniques like PROPELLER or navigator-based methods are used, then motion artifacts are reduced, but image quality in sparse k-space regions remains insufficient
Solution Approach 1:
A neural network is introduced as an intermediary to generate synthetic k-space data that fills sparse regions. The neural network takes fully sampled k-space data as input and produces synthetic k-space data with matching anatomical structures, which then complements the motion-corrected but sparse k-space data to reconstruct high-quality images.
Solution Approach 2:
The method merges motion-corrected sparse k-space data with synthetic k-space data generated by the neural network. By combining these two data sources in the k-space domain and reconstructing together, the approach achieves both motion artifact reduction and high image quality in sparse regions.
2Ease of operation
If traditional Cartesian sampling is used, then data acquisition is straightforward, but subject movement during acquisition causes blurring and artifacts
Solution Approach 1:
The k-space data is segmented into fully sampled regions and sparse regions. The fully sampled regions are used to train and generate synthetic data, while the sparse regions are filled using the neural network's output. This segmentation allows the system to leverage both traditional and AI-based approaches.
Solution Approach 2:
The neural network learns to copy anatomical structures from fully sampled k-space data and reproduces them in the synthetic k-space data. This copying mechanism preserves anatomical accuracy while enabling fast acquisition with reduced susceptibility to motion artifacts.
3Productivity
If k-space data is acquired quickly to reduce scan time, then productivity increases, but sparse k-space regions result in poor image quality
Solution Approach 1:
The traditional mechanical approach of acquiring complete k-space data is replaced with an AI-based system. The neural network substitutes for the physical measurement process in sparse regions, generating synthetic data that would otherwise require time-consuming acquisition, thus maintaining high image quality while enabling faster scanning.
Solution Approach 2:
The sampling density parameter is changed across different k-space regions. Fully sampled regions maintain high density for accurate neural network training and reference, while sparse regions use reduced sampling that is compensated by synthetic data generation, achieving overall faster acquisition without quality loss.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Disclosed herein is an image reconstruction method of reconstructing an augmented magnetic resonance image (310). The method comprises receiving (500) motion corrected k-space data (102) having a first tissue contrast modality; identifying (502) sparse k-space regions (104) within the motion corrected k-space data; receiving (504) an alternative magnetic resonance image (300) with a second tissue contrast modality. The method further comprises generating (506) a synthetic magnetic resonance image (302) which has the first tissue contrast modality in response to inputting the alternative magnetic resonance image into a magnetic resonance modality conversion neural network (312) and generating (508) synthetic k-space data from the synthetic magnetic resonance image. An augmented magnetic resonance image is reconstructed (510) by using the motion corrected k-space data such that the sparse k-space regions of the motion-corrected k-space data are augmented with the synthetic k-space data. Systems are presented that are configured to perform the method.