Iterative MR Image Reconstruction with Parameter-Decoupled Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing MR image reconstruction methods require extensive training for each application, increasing computational time and effort due to application-specific parameters, and existing machine learning models are not effectively decoupled from these parameters.
Innovation Solution
A method where a trained machine learning model (MLM) is formally separated from application-specific parameters, allowing it to be used across various applications by optimizing a loss function iteratively and enhancing images using a predefined MLM for MR image reconstruction, independent of specific parameters like motion compensation or acquisition processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a dedicated machine learning model is trained for each application with application-specific parameters, then the reconstruction quality for that specific application is optimized, but the overall training effort and computational time increase significantly
Solution Approach 1:
The patent applies universality by developing a single machine learning model that can handle multiple different applications and acquisition protocols without requiring separate training for each. The model is trained on a diverse dataset encompassing various applications (e.g., structural imaging, functional imaging, different k-space sampling patterns), enabling it to generalize and perform well across all these different scenarios using the same trained parameters, thus eliminating the need for multiple dedicated models
Solution Approach 2:
The patent utilizes parameter changes by dynamically adjusting reconstruction parameters such as regularization weights, iteration counts, and loss function components based on the specific application and acquisition protocol, while keeping the core machine learning model architecture and trained parameters fixed. This allows optimization for each application without retraining the model
2Reliability
If application-specific parameters are incorporated into the machine learning model training, then the model performance for that application improves, but the model complexity and data requirements increase
Solution Approach 1:
The patent applies segmentation by separating the reconstruction process into two independent parts: a universal machine learning model that handles the complex pattern recognition and image enhancement tasks, and application-specific parameters that handle the physics-based reconstruction details. This segmentation allows the model to remain simple and general-purpose while application-specific parameters provide the necessary customization without increasing model complexity
Solution Approach 2:
The patent uses an intermediary approach where the machine learning model serves as a mediator between the raw k-space data and the final reconstructed image. The model learns to handle various application-specific characteristics during training by processing diverse data, thereby mediating the differences between applications without requiring each application to have its own dedicated model
3Speed
If multiple receiver coils and k-space subsampling techniques are used for parallel imaging, then the imaging speed and acceleration are improved, but the reconstruction complexity increases requiring sophisticated techniques
Solution Approach 1:
The patent replaces conventional mechanical reconstruction techniques (such as iterative SENSE or GRAPPA methods that require solving linear systems or performing matrix operations) with a machine learning-based approach. The trained neural network directly maps undersampled k-space data from multiple coils to the final image, substituting complex mathematical reconstruction algorithms with a learned transformation that handles parallel imaging artifacts automatically
Data Source
AI summary
For MR image reconstruction, MR measurement data representing an imaged object is obtained and, for each iteration of at least two iterations, a prior MR image for the respective iteration is received, an optimized MR image is generated by optimizing a predefined first loss function, which depends on the MR measurement data and on the prior MR image, and an enhanced MR image is generated by applying a trained machine learning model, MLM, for image enhancement to the optimized MR image. The prior MR image of the respective iteration corresponds to the enhanced MR image of a preceding iteration, unless the respective iteration corresponds to an initial iteration of the at least two iterations, and the prior MR image of the initial iteration corresponds to a predefined initial image.