Machine-Learned MRI Reconstruction for Patient Motion Compensation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing MRI reconstruction techniques struggle with motion artifacts due to patient movement during extended measurement times, leading to inaccurate reconstructions, particularly in free-breathing liver and heart acquisitions.
Innovation Solution
An iterative optimization method using a machine-learned algorithm that incorporates a regularization operation and data-consistency operation, where multiple prior images from different motion states and time points are concatenated and warped to a reference state, combined with spatial, temporal, and motion-state convolutions to mitigate motion artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If measurement time duration is extended to improve image quality, then image quality is improved, but motion artifacts increase due to patient movement
Solution Approach 1:
The system performs preliminary actions by warping multiple prior images from different motion states to a reference motion state before reconstruction, and by using trained neural network models that have been pre-trained to recognize and compensate for motion patterns, thereby preparing the data in advance to mitigate motion artifacts
Solution Approach 2:
The system converts the harmful effect of motion artifacts into a beneficial outcome by using the motion information contained in multiple prior images to improve reconstruction accuracy. The motion states that initially cause artifacts are transformed into useful information through warping and neural network processing, enabling motion-compensated reconstruction that actually improves image quality
2Manufacturing precision
If iterative optimization with multiple prior images is used to reduce motion artifacts, then motion artifact reduction is improved, but computational complexity increases
Solution Approach 1:
The system replaces complex mechanical iterative optimization processes with a trained neural network model that has learned the reconstruction patterns. Instead of performing heavy computational iterations during reconstruction, the pre-trained network performs inference, substituting mechanical computation with learned patterns that achieve similar or better results with less computational burden
Solution Approach 2:
The neural network model is pre-trained in advance using extensive training data that includes various motion patterns. This preliminary training phase performs the heavy computational work upfront, so that during actual reconstruction, the system only needs to apply the pre-learned model, significantly reducing real-time computational complexity
3Productivity
If undersampling trajectory is used to accelerate data acquisition, then productivity is improved, but reconstruction accuracy deteriorates
Solution Approach 1:
The system replaces traditional mechanical reconstruction algorithms with a trained neural network that has learned to accurately reconstruct images from undersampled data. The network substitutes conventional iterative methods with learned patterns, achieving high reconstruction accuracy even with aggressive undersampling by recognizing anatomical structures and patterns that traditional algorithms miss
Data Source
AI summary
The disclosure relates to MRI reconstruction of multiple MRI measurement datasets acquired throughout a measurement time duration. Patient motion can occur during the measurement time duration. Warping operators, sometimes also referred to as motion field, are incorporated into an iterative optimization of the MRI reconstruction.