Deep Learning MRI Reconstruction for Undersampled K-Space
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional MRI reconstruction methods face challenges with aliasing artifacts and noise amplification due to undersampling, particularly in accelerated MRI techniques, which are not adequately addressed by existing parallel imaging methods like SENSE and GRAPPA.
Innovation Solution
A deep learning regularized reconstruction method using k-space convolution kernels and neural networks with multiple convolutional layers iteratively enhances image quality by preserving image features and minimizing noise, particularly in high acceleration factors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If k-space data is fully sampled to satisfy the Nyquist-Shannon sampling theorem, then aliasing is avoided and accurate signal reconstruction is achieved, but scan time becomes excessively long
Solution Approach 1:
The patent applies preliminary action by acquiring a fully-sampled calibration region (e.g., central k-space lines) before the main undersampled imaging acquisition. This calibration data is used to compute sensitivity maps and convolution kernels that enable accurate reconstruction from the subsequently acquired undersampled data, thus preparing the necessary information in advance to overcome the limitations of undersampling
Solution Approach 2:
The patent introduces intermediary elements including sensitivity maps derived from the calibration region and convolution kernels computed from the calibration data. These intermediaries serve as mediators between the undersampled k-space data and the final reconstructed image, enabling accurate signal recovery without requiring full k-space sampling
2Loss of time
If k-space data is under-sampled to reduce scan time, then scan time is reduced, but aliasing artifacts and reconstruction errors are introduced
Solution Approach 1:
The patent performs preliminary computation of sensitivity maps and convolution kernels using fully-sampled calibration data before the main imaging acquisition. This preliminary preparation enables the system to accurately reconstruct images from undersampled data by having the necessary correction information ready in advance
Solution Approach 2:
The patent implements feedback mechanisms through iterative reconstruction algorithms that repeatedly apply convolution operators and sensitivity map corrections. The algorithm continuously refines the image estimate by comparing reconstructed data with actual measurements and adjusting the solution accordingly, progressively reducing aliasing artifacts and reconstruction errors
3Productivity
If SENSE-based methods are used for parallel imaging reconstruction, then image reconstruction from undersampled data is enabled, but accuracy depends on precise coil sensitivity maps that are difficult to obtain
Solution Approach 1:
The patent computes sensitivity maps and convolution operators in advance using fully-sampled calibration region data. This preliminary computation ensures that accurate sensitivity information is available before the main imaging acquisition, eliminating the need for precise real-time sensitivity map estimation during reconstruction
Solution Approach 2:
The patent makes the system self-service by using the acquired calibration region data itself to compute the necessary sensitivity maps and convolution kernels. The calibration data serves dual purposes: it is both part of the imaging data and the source material for generating the reconstruction operators, eliminating the need for separate sensitivity mapping acquisitions
4Reliability
If GRAPPA is used for parallel imaging reconstruction, then data consistency in k-space is ensured, but computational complexity increases significantly for non-uniform sampling patterns
Solution Approach 1:
The patent pre-computes convolution operators and sensitivity maps from fully-sampled calibration data before the main imaging acquisition. This preliminary computation transforms the complex real-time reconstruction problem into a simpler application of pre-computed operators during image reconstruction, significantly reducing computational complexity
Solution Approach 2:
The patent uses a partial calibration region (e.g., central k-space lines) rather than requiring full k-space sampling for calibration. This partial action provides sufficient information to compute accurate sensitivity maps and convolution operators while dramatically reducing the computational burden compared to full calibration acquisitions
5Measurement precision
If iterative SENSE-based methods are used for reconstruction, then reconstruction accuracy is improved, but the methods remain affected by inaccuracies in sensitivity maps and FOV truncation
Solution Approach 1:
The patent computes sensitivity maps and convolution operators from fully-sampled calibration region data before the main imaging acquisition. This preliminary computation ensures that accurate sensitivity information is available in advance, eliminating the dependency on real-time sensitivity map estimation and making the reconstruction robust to sensitivity map inaccuracies
Solution Approach 2:
The patent introduces convolution operators as intermediary elements that directly operate on k-space data to produce corrected images. These operators, derived from calibration data, serve as mediators that compensate for undersampling effects without requiring continuous sensitivity map updates, thereby reducing sensitivity to sensitivity map errors and FOV truncation
Data Source
AI summary
Embodiments of the present disclosure relate to systems and methods for reconstructing magnetic resonance imaging (MRI) images from under sampled k-space data. The method includes acquiring multi-coil under sampled k-space data with a fully-sampled calibration region, determining a k-space convolution kernel, and producing an initial multi-coil MRI image estimate via inverse Fourier transform. The estimate is regularized using a trained deep learning regularizer, transformed to synthetic fully-sampled k-space data, and convolved with the kernel to produce data-consistent synthetic k-space data. This is combined with the original under sampled data to produce an estimated fully-sampled k-space data, which is then inverse transformed to update the multi-coil MRI image estimate. The process iteratively improves image quality, reduces noise, and preserves features, enhancing MRI reconstruction from under sampled data.