Multi-Stage Neural Network for MR Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional multistage neural networks for MR image reconstruction face challenges such as increased computation time with the number of parameters, risk of overfitting, and the need for extensive training data and re-training when modifying the number of stages or loss functions.
Innovation Solution
The approach involves independently training each stage of the neural network using a specific training method that allows for different loss functions and optimization methods for each stage, reducing the overall computational time and data requirements, and enabling the addition of stages without re-training existing ones.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If the number of stages in a multistage neural network is increased to improve reconstruction quality, then image reconstruction quality is improved, but computation time and the number of parameters increase non-linearly
Solution Approach 1:
The patent divides the training process into separate stages, where each stage is trained independently using a sequential training approach. This segmentation allows each stage to be trained with fewer parameters initially, avoiding the non-linear computation time increase that would occur if all stages were trained simultaneously with all parameters from the beginning.
2Manufacturing precision
If the number of stages is increased to improve reconstruction quality, then image reconstruction quality is improved, but the risk of overfitting increases and more training examples are needed
Solution Approach 1:
The patent applies preliminary action by training earlier stages before training later stages. Each stage is trained sequentially, with previous stages already trained and fixed when training the next stage. This preliminary training approach allows each stage to learn progressively without requiring excessive training examples that would be needed if all stages were trained simultaneously, thereby reducing the risk of overfitting while still achieving high reconstruction quality.
3Manufacturing precision
If conventional training methods are used to learn optimal network parameters for all stages, then reconstruction quality can be improved, but the entire network must be re-trained if the number of stages is changed
Solution Approach 1:
The patent segments the network training into independent stage-wise training modules. Each stage can be trained independently and then combined with other stages to form networks with different numbers of stages. This segmentation provides adaptability because stages can be added, removed, or modified without requiring complete re-training of the entire network, while still maintaining reconstruction quality through the sequential training approach.
4Manufacturing precision
If conventional training methods are used, then all stages can be optimized, but different loss functions cannot be easily applied to different stages
Solution Approach 1:
The patent applies local quality by allowing different loss functions to be used for different stages of the network. Each stage can have its own customized loss function tailored to its specific reconstruction task, while the sequential training framework enables independent optimization of each stage's parameters. This provides versatility in selecting appropriate loss functions for different stages without compromising the overall optimization of network parameters.
Data Source
AI summary
A system and method include generation of modified imaging data and modified acquisition parameters of a training example based on the target imaging data and target acquisition parameters of each of a plurality of target examples, and generation, for each training example, of an initial reconstructed image based on the modified imaging data and modified acquisition parameters of the training example. A first network stage of a multi-stage network is trained based on the modified imaging data, modified acquisition parameters and initial reconstructed image of each training example, a first output image is generated for each training example by inputting the modified imaging data, modified acquisition parameters and initial reconstructed image of the training example to the trained first network stage, and a second network stage of the multi-stage network is trained based on the modified imaging data, modified acquisition parameters and first output image of each training example.