MRI Coil Sensitivity Estimation Using Cascaded Deep Learning Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI coil sensitivity estimation methods, such as eigenvalue approaches like ESPIRiT, are computationally complex, limited by the need for fully sampled auto-calibration signals, and cannot be integrated with deep learning reconstruction models, restricting acceleration and image quality in parallel imaging and compressed sensing techniques.
Innovation Solution
A system utilizing cascades of regularization networks and deepsets coil sensitivity map networks for end-to-end training, enabling the estimation and refinement of coil sensitivity maps from subsampled auto-calibration signals, which are permutation invariant or equivariant, allowing for further MRI acceleration while preserving image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If eigenvalue approach (ESPIRiT) is used for coil sensitivity estimation, then coil sensitivity maps can be obtained, but computational complexity increases significantly especially for 3D problems
Solution Approach 1:
The patent replaces the traditional eigenvalue decomposition method (mathematical/algorithmic approach) with a deep learning neural network model. The neural network is trained to directly predict coil sensitivity maps from k-space data, substituting the computationally intensive eigenvalue approach with a faster inference-based method that maintains accuracy while reducing computational burden during actual MRI reconstruction
Solution Approach 2:
The patent performs preliminary training of the neural network model using eigenvalue-based coil sensitivity maps as ground truth. By pre-training the model with accurate reference data, the network learns to replicate the eigenvalue approach's accuracy while enabling rapid inference during actual scanning, thus resolving the contradiction between accuracy and computational complexity
2Measurement precision
If fully sampled ACS is used for coil sensitivity estimation, then accurate sensitivity maps can be obtained, but maximum acceleration is limited
Solution Approach 1:
The patent uses a small fully sampled ACS region to generate reference coil sensitivity maps, then applies these sensitivity maps to reconstruct images from highly undersampled k-space data. The neural network learns to generalize from the limited fully sampled information to accurately estimate sensitivities for the entire k-space, enabling high acceleration while maintaining accuracy
Solution Approach 2:
The patent changes the parameter of ACS sampling from fully sampled to highly undersampled by applying random masks. The neural network is trained to handle this undersampled input and still produce accurate coil sensitivity estimates, thus enabling faster acquisition without sacrificing the quality of sensitivity map estimation
3Measurement precision
If traditional CSM estimation approach is used, then coil sensitivity maps can be estimated, but integration with deep learning reconstruction models is not possible
Solution Approach 1:
The patent merges the coil sensitivity estimation function directly into the deep learning reconstruction model by incorporating a neural network module that outputs coil sensitivity maps as part of the end-to-end reconstruction pipeline. This integration allows the CSM estimation and image reconstruction to be jointly optimized, enabling both accurate sensitivity estimation and seamless compatibility with deep learning-based accelerated MRI methods
Solution Approach 2:
The patent designs a universal neural network framework that can perform multiple functions: estimating coil sensitivity maps, reconstructing images from undersampled data, and adapting to different acceleration factors. The model serves as a multi-functional component that replaces both traditional CSM estimation methods and reconstruction algorithms, providing versatility across different MRI acceleration scenarios
Data Source
AI summary
A system is provided for MRI coil sensitivity estimation and reconstruction At least two cascades of regularization networks are serially connected such that the output of a cascade is used as input of a following cascade, at least two deepsets coil sensitivity map networks are serially connected such that the output of a deepsets coil sensitivity map network is used as input of a following deepsets coil sensitivity map network (CR), and wherein the outputs of the deepsets coil sensitivity map networks are also used as inputs for the cascades.