Motion-Resolved MRI Reconstruction Without k-Space Data Consistency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 4D MRI reconstruction techniques are computationally expensive and time-consuming due to the need for iterative reconstruction algorithms that enforce data consistency in k-space, which is particularly slow when dealing with large amounts of data from multiple motion cycles.
Innovation Solution
A convolutional neural network is used to reconstruct 4D images by exploiting spatial, temporal, and coil correlations without enforcing explicit data consistency in k-space, allowing for continuous k-space data accumulation and motion signal detection, followed by image classification and reconstruction using a residual U-net network and spatiotemporal patch embedding network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If iterative reconstruction algorithms are used to enforce data consistency in k-space, then manufacturing precision of image reconstruction is improved, but productivity is worsened due to slow computation
Solution Approach 1:
The patent replaces the iterative mechanical reconstruction process with a deep learning-based system. A neural network is trained on pairs of aliased images and corresponding high-quality reference images, learning to directly generate high-quality images from aliased inputs without iterative k-space consistency enforcement, thereby achieving fast reconstruction with high quality
Solution Approach 2:
The patent performs preliminary training of the neural network offline using pre-collected training data consisting of aliased images and reference images. This preliminary action allows the network to learn the mapping from aliased to high-quality images in advance, enabling fast real-time reconstruction without requiring iterative processing during actual reconstruction
2Productivity
If the number of k-space points is reduced for fast acquisition, then productivity is improved, but manufacturing precision is worsened due to aliasing artifacts
Solution Approach 1:
The patent converts the harmful aliasing artifacts caused by reduced k-space sampling into beneficial training data. The neural network is trained specifically on aliased images as inputs, learning to recognize and correct the characteristic patterns of aliasing, thereby transforming the previously harmful undersampling into a manageable input condition that the network can effectively process
Solution Approach 2:
The neural network acts as an intermediary between the aliased image data and the final high-quality output. Instead of directly processing the aliased images or requiring iterative correction, the trained network serves as a mediator that automatically learns and applies the appropriate corrections, separating the acquisition phase (where aliasing is acceptable) from the reconstruction phase (where quality is restored)
Data Source
AI summary
Systems and methods for fast reconstruction of motion-resolved magnetic resonance images using space-time-coil convolutional networks are disclosed. The system can receive a plurality of k-space data sets. The system can detect a motion signal therefrom. The system can classify the k-space data sets according to states of the motion signals. The system can resolve the k-space data set to Euclidean space images. The system can resolve the Euclidean space images to a combined Euclidian space image. For example, the system can use a convolutional network that exploits spatial, temporal and coil correlations without k-space data consistency to minimize computation time.