Conditional MRI Reconstruction Network With Metadata-Tuned CNN Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model-based deep-learning algorithms for MRI reconstruction are limited by their dependence on specific measurement schemes, requiring separate training for each acquisition setting and lacking flexibility to adapt to variations in signal-to-noise ratio, field strength, and image content, leading to performance degradation and inefficiencies in data storage and deployment.
Innovation Solution
A conditional unrolled architecture, termed Meta-MoDL, uses a multilayer perceptron (MLP) to modulate CNN feature weights and regularization parameters based on acquisition metadata, enabling a single network to adapt to different settings, reducing the need for multiple networks and improving generalizability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model-based deep-learning algorithms are trained for specific measurement schemes, then reconstruction performance is improved, but device complexity increases due to the need for multiple networks
Solution Approach 1:
The patent applies universality by designing a single neural network architecture that can handle multiple measurement schemes through conditional unrolling. The network uses acquisition-specific conditional vectors to adapt its behavior, allowing one network to perform multiple functions that would traditionally require separate specialized networks for different MRI acquisition settings.
Solution Approach 2:
The patent implements dynamics by making the network parameters adaptive through conditional unrolling. The network dynamically adjusts its reconstruction behavior based on acquisition-specific conditional vectors, allowing the same network structure to optimize performance across different measurement schemes without requiring static, separate networks for each scheme.
2Reliability
If separate networks are trained for each acquisition setting, then reconstruction performance is optimized, but loss of time increases due to training data requirements and model switching
Solution Approach 1:
The patent trains a single universal network that can handle multiple acquisition settings through conditional unrolling, eliminating the need to train and switch between multiple separate networks. This universal approach reduces training time by consolidating data requirements into one comprehensive training process and eliminates runtime model switching overhead.
3Device complexity
If a single universal network is trained for all applications, then device complexity is reduced, but reconstruction performance degrades due to parameter adaptation requirements
Solution Approach 1:
The patent resolves the performance degradation issue by implementing dynamic adaptation through conditional unrolling. The single network adjusts its parameters and behavior dynamically based on acquisition-specific conditional vectors, allowing it to optimize performance for different measurement schemes while maintaining a unified network structure.
Solution Approach 2:
The patent applies parameter changes by using acquisition-specific conditional vectors to modulate network parameters during inference. This allows the single universal network to adapt its effective parameters based on the specific acquisition settings, maintaining high performance across different applications without requiring multiple fixed networks.
4Reliability
If regularization parameters are adapted for each acquisition setting, then reconstruction quality is improved, but ease of operation decreases due to parameter tuning complexity
Solution Approach 1:
The patent implements self-service by having the network automatically adapt regularization parameters based on acquisition-specific conditional vectors during inference. This eliminates the need for manual parameter tuning and complex user intervention, allowing the system to self-adjust to different acquisition settings while maintaining high reconstruction quality.
Data Source
Figure 1A~1C
Figure 2
Figure 3
AI summary
An apparatus for reconstructing or filtering medical image data is provided. The apparatus includes processing circuitry to receive a first medical image data and meta-parameters related to the first medical image data; apply the received meta-parameters to inputs of a first trained machine-learning (ML) network, e.g., a multilayer perceptron, to obtain, from outputs of the first trained ML network, tuning parameters of a second ML network (e.g., a convolutional neural network) different from the first ML network; apply the received first medical image data to inputs of the second ML network, as tuned by the obtained tuning parameters output from the first ML network, to obtain, from outputs of the second ML network, second medical image data; and output the second medical image data. In one embodiment, the first medical image data is magnetic-resonance k-space data and the second medical data is a magnetic-resonance image.