Deep Learning MRI Reconstruction from Partial Fourier Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional MRI image reconstruction methods, such as POCS, suffer from decreased image signal-to-noise ratio and suboptimal performance when dealing with partial Fourier data, especially in the presence of rapid image phase variations.
Innovation Solution
The development of deep learning-based systems and methods for robust partial Fourier reconstruction, which involve training convolutional neural networks (CNNs) with high-quality MRI data and applying these models to reconstruct high-quality images from incomplete partial Fourier data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If partial Fourier data acquisition is performed to accelerate scan, then scan time is reduced, but image signal-to-noise ratio decreases
Solution Approach 1:
The patent replaces conventional iterative mathematical reconstruction methods (POCS) with deep learning-based neural networks. The neural network is trained on fully-sampled k-space data and complex MRI images, then applied to reconstruct images from partial Fourier data, achieving both accelerated scanning and high-quality reconstruction with preserved phase information.
2Productivity
If partial Fourier fraction is reduced to further accelerate scan, then scanning efficiency increases, but reconstruction performance deteriorates especially with rapid phase variations
Solution Approach 1:
The patent performs preliminary training of the neural network using large datasets of fully-sampled k-space data and corresponding complex MRI images before deployment. This pre-training phase enables the network to learn optimal reconstruction patterns, allowing it to successfully reconstruct images from highly undersampled partial Fourier data even with rapid phase variations.
Solution Approach 2:
The patent changes the approach from iterative mathematical parameter adjustment (POCS) to learned parameter transformation through neural networks. The network learns to map partial Fourier k-space data to complete image representations, adapting to various sampling fractions and phase conditions through training on diverse data.
Data Source
AI summary
Disclosed are deep learning based methods for magnetic resonance imaging (MRI) image reconstruction from partial Fourier-space (i.e., k-space) data, involving: obtaining high-quality complex MRI image data or fully-sampled k-space data as training data; training reconstruction models to predict high-quality complex MRI image data or complete k-space data from incomplete or partial k-space data; and applying trained models to reconstruct high-quality complex MRI image data or complete k-space data from partial k-space data.