Neural Network MRI Reconstruction for Faster Subsampled Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing magnetic resonance imaging (MRI) techniques require longer acquisition times due to the need for full sampling of MR images, which can be addressed by using a neural network-based learning model to reconstruct MR images from subsampled data, thereby reducing image acquisition time.
Innovation Solution
An MRI apparatus and method utilizing a neural network to create a learning model that correlates subsampled MR images with fully sampled MR images, allowing for the reconstruction of high-quality images by removing aliasing artifacts through statistical machine learning, enabling efficient image acquisition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If full sampling of MR images is performed, then image quality is maintained, but acquisition time increases
Solution Approach 1:
The system performs preliminary actions by acquiring only a subset of k-space data (subsampled k-space) and then uses a neural network model trained on fully sampled images to predict and reconstruct the complete image. This preliminary subsampling reduces acquisition time while the neural network reconstruction restores image quality that would otherwise be lost.
Solution Approach 2:
The neural network learns to copy the characteristics of fully sampled MR images from training data onto subsampled input images. By training on pairs of subsampled and fully sampled images, the network creates a mapping that copies the complete image structure from the training data onto the reduced input data, enabling reconstruction without requiring full sampling.
2Productivity
If subsampled MR images are used, then acquisition time is reduced, but aliasing artifacts appear
Solution Approach 1:
The system converts the harmful aliasing artifacts introduced by subsampling into a beneficial training signal. During training, the neural network learns to recognize and eliminate these artifacts by processing pairs of subsampled images (with artifacts) and their corresponding fully sampled versions (without artifacts). The artifacts thus become part of the learning process that enables high-quality reconstruction from accelerated data.
Solution Approach 2:
The neural network employs feedback mechanisms during training where the difference between subsampled and fully sampled images is used to adjust network parameters. This feedback loop enables the network to progressively learn to remove aliasing artifacts and produce high-quality reconstructions from subsampled data, transforming the harmful artifacts into a learning opportunity.
3Loss of time
If neural network-based reconstruction is applied, then acquisition time is reduced, but computational complexity increases
Solution Approach 1:
The computationally intensive neural network model is trained in advance using pairs of subsampled and fully sampled MR images. This preliminary training phase occurs offline, allowing the network to store learned patterns and reconstruction rules in its parameters. During actual scanning, only lightweight inference operations are needed, significantly reducing real-time computational complexity while maintaining accelerated acquisition capabilities.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
A magnetic resonance imaging (MRI) apparatus includes a processor, and a memory storing a program including instructions that, when executed by the processor, cause the processor to acquire first data of a subsampled magnetic resonance (MR) image, acquire, based on a learning model using a neural network, first reconstructed data with respect to rows of pixels in a first phase encoding direction of the first data of the subsampled MR image, and obtain a reconstructed image corresponding to the subsampled MR image, using the first reconstructed data.