MRI Reconstruction Model Using Deep Learning for Speed and Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current magnetic-resonance imaging techniques face limitations in speed and noise amplification due to hardware constraints and long reconstruction times, particularly with parallel imaging and compressed sensing methods.
Innovation Solution
A magnetic-resonance imaging method and system that utilizes an iterative algorithm to establish and train an initial imaging model, learning undetermined parameters, solving operators, and structural relationships, allowing for the generation of high-quality images from under-sampled K-space data using neural networks and sparse transform algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If parallel imaging is used to accelerate acquisition, then imaging speed is improved, but noise amplification occurs and acceleration multiple is limited due to hardware constraints
Solution Approach 1:
The patent introduces a deep learning-based reconstruction model as an intermediary between the under-sampled K-space data and the final image. This model learns the mapping relationship from incomplete data to high-quality images, enabling acceleration factors beyond traditional parallel imaging limits without proportional noise amplification. The learned reconstruction model acts as a mediator that recovers missing information intelligently.
Solution Approach 2:
The patent changes the reconstruction parameters from fixed iterative algorithms to learned parameters through deep neural networks. By training the model on paired datasets (under-sampled and fully-sampled images), the system learns optimal reconstruction parameters that adapt to different imaging conditions, achieving better quality-speed tradeoff than traditional methods.
2Speed
If compressed sensing is used to reduce K-space sampled points, then imaging speed is improved, but reconstruction time becomes very long due to iteration reconstruction
Solution Approach 1:
The patent replaces the mechanical iterative reconstruction process of compressed sensing with a deep learning-based direct reconstruction approach. Instead of repeatedly iterating through optimization algorithms, the trained neural network performs reconstruction in a single forward pass, dramatically reducing reconstruction time while maintaining image quality.
Solution Approach 2:
The patent performs preliminary training of the reconstruction model using paired datasets before actual imaging. This preliminary action pre-learns the reconstruction mapping, so that during actual use, only a single inference step is needed rather than iterative reconstruction, significantly reducing real-time reconstruction duration.
3Reliability
If traditional iterative algorithms are used for reconstruction, then image quality can be maintained, but the reconstruction process is complex and requires extensive parameter selection
Solution Approach 1:
The patent enables the reconstruction system to learn optimal parameters and structural relationships automatically from training data. The deep learning model self-adjusts its internal parameters through backpropagation during training, eliminating the need for manual parameter selection and complex iterative algorithm tuning. The system serves itself by learning from examples rather than requiring expert configuration.
Data Source
AI summary
Disclosed are a magnetic-resonance imaging method, apparatus and system, and a storage medium. The method includes acquiring an initial model of magnetic-resonance imaging and establishing an initial imaging model according to an iterative algorithm used for solving the initial model, where the iterative algorithm includes at least one of an undetermined parameter, an undetermined solving operator or an undetermined structural relationship; training the initial imaging model on the basis of sample data to generate a magnetic-resonance imaging model, where training of the initial imaging model is used for learning the at least one of the undetermined parameter, the undetermined solving operator or the undetermined structural relationship in the iterative algorithm; and acquiring under-sampled K-space data to be processed, inputting the under-sampled K-space data into the magnetic-resonance imaging model, and generating a magnetic-resonance image.