Multi-slice MRI Reconstruction via Transformer Attention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Magnetic resonance imaging (MRI) faces challenges with low imaging speed, low image quality, and low diagnosis accuracy due to the ill-conditioned nature of parallel imaging data, which leads to noise amplification and residual artifacts, hindering clinical diagnosis and machine utilization.
Innovation Solution
A multi-slice MRI method utilizing a deep learning reconstruction model with a Transformer structure, incorporating learnable positional and imaging parameter embeddings, which preprocesses input data to improve image reconstruction by leveraging spatial and prior imaging information, and combines gradient-based data consistency updating for end-to-end iterative training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If parallel imaging technology is used to increase imaging speed, then imaging speed is improved, but image quality deteriorates due to noise amplification and residual artifacts
Solution Approach 1:
The patent introduces a deep learning reconstruction model as an intermediary between the parallel imaging data acquisition and the final image reconstruction. This model processes the undersampled k-space data and reconstructed images through multiple layers of neural networks, acting as a mediator that transforms the noisy parallel imaging data into high-quality images while maintaining the accelerated imaging speed
Solution Approach 2:
The patent changes the reconstruction parameters by using learnable parameters in the deep learning model instead of fixed reconstruction parameters. The model learns optimal reconstruction parameters from training data, allowing it to adaptively adjust to different imaging conditions and acceleration factors, thereby maintaining image quality across various parallel imaging scenarios
2Loss of time
If acceleration multiple is increased to reduce scanning time, then scanning time is reduced, but noise amplification and residual artifacts increase
Solution Approach 1:
The patent implements feedback mechanisms through the deep learning reconstruction model that processes both the undersampled k-space data and initial reconstructed images. The model uses loss functions that compare reconstructed images with reference images, providing feedback that guides the optimization of reconstruction parameters and reduces noise amplification and artifacts even at high acceleration multiples
Solution Approach 2:
The patent performs preliminary actions by pre-training the deep learning reconstruction model on large datasets before actual imaging. This preliminary training allows the model to learn effective reconstruction patterns and noise characteristics, enabling it to handle high acceleration multiples with reduced artifacts and noise when deployed in actual imaging scenarios
Data Source
AI summary
The invention provides a multi-slice magnetic resonance imaging method and device based on long-distance attention model reconstruction. The method includes that: a deep learning reconstruction model is constructed; data preprocessing is performed on multiple slices of simultaneously acquired signals, and multiple slices of magnetic resonance images or K-space data is used as data input; learnable positional embedding and imaging parameter embedding are acquired; the preprocessed input data, the positional embedding and the imaging parameter embedding are input into the deep learning reconstruction model; and the deep learning reconstruction model outputs a result of the magnetic resonance reconstruction image. The invention further provides a device for implementing the method. The invention may improve the quality of the magnetic resonance image, improve the diagnosis accuracy of a doctor, increase the imaging speed, and improve the utilization rate of a magnetic resonance machine.