MRI Reconstruction via Deep Subspace Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning-based MRI image reconstruction methods face memory limitations, restricting their ability to handle high-dimensional data due to high memory and computational requirements, which prevents them from efficiently processing higher-dimensional MRI techniques.
Innovation Solution
Deep subspace learning reconstruction (DSLR) method that converts high-dimensional sensor data into a compressed representation, using simpler neural networks for reconstruction in the compressed domain, integrating MRI physics-based modeling for data consistency, and decompressing at the end for image visualization, allowing for memory-efficient training and inference across various dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep neural networks are used for MRI image reconstruction, then image quality and reconstruction speed are improved, but memory requirements increase significantly
Solution Approach 1:
The patent segments the high-dimensional reconstruction problem into lower-dimensional subproblems by decomposing the data into a compressed representation with reduced dimensionality. This allows the neural network to process and reconstruct data in a simplified space, reducing memory requirements while maintaining reconstruction quality and speed benefits.
Solution Approach 2:
The patent transforms the high-dimensional data into a compressed representation with lower dimensionality, effectively moving the problem to a different dimensional space. This dimensionality reduction enables the use of simpler neural networks with fewer parameters, thereby reducing memory requirements while preserving the ability to reconstruct high-quality images.
2Measurement precision
If deep neural networks are trained on high-dimensional data, then reconstruction accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent segments the computational task by decomposing the high-dimensional data into a compressed representation. This segmentation allows the neural network to operate on lower-dimensional data, reducing the computational complexity of each operation while maintaining overall reconstruction accuracy through the compressed representation that preserves essential information.
Solution Approach 2:
The patent changes the parameter space by transforming data from high-dimensional original space to low-dimensional compressed space. This parameter transformation reduces the number of parameters the neural network needs to process, thereby reducing computational complexity while maintaining reconstruction accuracy through the preserved data structure in compressed space.
3Quantity of substance
If conventional iterative reconstruction methods are used, then memory requirements are reduced, but reconstruction time increases
Solution Approach 1:
The patent performs preliminary compression of the data into a compressed representation before feeding it to the neural network. This preliminary action reduces the data dimensionality and memory requirements upfront, allowing the subsequent neural network reconstruction to operate more efficiently with less memory while maintaining fast reconstruction speeds.
Data Source
AI summary
A method for MR imaging includes acquiring with an MR imaging apparatus undersampled k-space imaging data having one or more temporal dimensions and two or more spatial dimensions; transforming the undersampled k-space imaging data to image space data using zero-filled or sliding window reconstruction and sensitivity maps; decomposing the image space data into a compressed representation comprising a product of spatial and temporal parts, where the spatial part comprises spatial basis functions and the temporal part comprises temporal basis functions; processing the spatial basis functions and temporal basis functions to produce reconstructed spatial basis functions and reconstructed temporal basis functions, wherein the processing iteratively applies conjugate gradient and convolutional neural network updates using 2D or 3D spatial and 1D temporal networks; and decompressing the reconstructed spatial basis functions and reconstructed temporal basis functions to produce a reconstructed MRI image having one or more temporal dimensions and two or more spatial dimensions.