MRI Reconstruction Using Multi-Prior Learning for Undersampled Scans
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing MRI reconstruction methods face challenges in producing high-quality images from under-sampled data, leading to aliasing artifacts and poor signal-to-noise ratio, particularly at high acceleration rates, due to hardware constraints and noise amplification, and require time-consuming empirical parameter tuning.
Innovation Solution
A consistency-aware multi-prior network that integrates parallel imaging and compressed sensing within a deep learning framework, leveraging image, k-space, and calibration priors through a collaborative learning approach, using neural networks to iteratively enhance and refine MRI reconstructions by exploring data redundancy across adjacent slices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If under-sampled k-space data is used for MRI reconstruction, then scan time is reduced, but image quality deteriorates with aliasing artifacts and poor signal-to-noise ratio
Solution Approach 1:
The patent divides the reconstruction task into two separate neural networks: one operating in image space and another in k-space domain. Each network specializes in handling specific aspects of the reconstruction problem, allowing the system to maintain high image quality from under-sampled data by addressing different spatial and frequency domain characteristics separately.
Solution Approach 2:
The patent performs transformations between image space and k-space domain, operating in both spatial and frequency dimensions. By converting between domains and applying modifications in both spaces, the system recovers lost information from under-sampled data and reduces aliasing artifacts that would otherwise degrade image quality.
2Productivity
If high acceleration rates are applied, then scan time is reduced, but noise amplification increases
Solution Approach 1:
The patent implements iterative reconstruction where the output of one neural network is fed back as input to the other network. The image-space network and k-space network exchange information through multiple iterations, allowing the system to progressively refine the reconstruction and suppress noise amplification that occurs at high acceleration rates.
Solution Approach 2:
The patent introduces an intermediary frequency fusion operation that combines results from both image-space and k-space processing. This fusion mechanism acts as a mediator that integrates the strengths of both approaches while mitigating their individual weaknesses, particularly in reducing noise amplification during high acceleration reconstruction.
3Measurement precision
If empirical parameter tuning is performed, then reconstruction accuracy can be improved, but processing time increases
Solution Approach 1:
The patent employs neural networks that are pre-trained to automatically learn optimal reconstruction parameters and patterns from training data. During actual reconstruction, the networks perform self-service by directly processing the under-sampled data without requiring manual empirical parameter tuning, thus maintaining high reconstruction accuracy while significantly reducing processing time.
Solution Approach 2:
The patent transforms the reconstruction problem from requiring manual parameter adjustment to using learned parameters embedded in the neural network weights. By changing from empirical parameter tuning to data-driven parameter learning, the system achieves high reconstruction accuracy automatically without the time cost of iterative manual optimization.
Data Source
AI summary
Techniques for performing iterative MRI image reconstruction by learning complementary multi-prior knowledge from images, k-space data, and calibration data are disclosed. In one method, k-space data is obtained from an MRI scan. Image-space modifications are performed on the k-space data using a first neural network trained to operate on data in image space. The k-space data is converted from the frequency domain to a spatial domain to produce input image-space data. Using the first neural network, output image-space data is generated, which is then converted from the spatial domain to the frequency domain. K-space modifications are performed on the k-space data using a second neural network trained to operate on data in k-space. ACS are encoded using a third neural network to guide the second neural network in learning consistency-aware k-space correlations. The k-space data is converted from the frequency domain to the spatial domain to obtain a reconstructed image.