Hybrid K-Space MRI Reconstruction With Noise-Adaptive Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based image reconstruction methods in MRI face challenges in generalizing to varying noise levels and conditions, leading to over-smoothing or poor denoising results, and existing techniques lack effective methods for dynamically adapting regularization parameters to noise levels.
Innovation Solution
A method that utilizes a hybrid reconstruction pipeline combining a sensitivity encoded reconstruction and a neural network, with a regularization parameter selected based on estimated noise levels, to generate a de-noised MR image by dynamically controlling the influence of the neural network output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a deep learning model is trained on a single noise level, then the model achieves good reconstruction quality for that specific noise level, but the model fails to generalize to other noise levels, producing overly smooth or aliased images
Solution Approach 1:
The patent implements dynamic adaptability by introducing a noise level estimation mechanism that operates at runtime to determine the appropriate regularization parameter based on the actual noise characteristics of the input data. This allows the system to adapt to varying noise levels without requiring separate trained models for each condition, resolving the contradiction between optimization for specific conditions and generalization across conditions.
Solution Approach 2:
The patent changes the regularization parameter dynamically based on estimated noise levels. Instead of using a fixed regularization parameter trained for a specific noise level, the system estimates the noise level of the input data and adjusts the regularization parameter accordingly, enabling the model to generalize across different noise conditions while maintaining optimal reconstruction quality.
2Adaptability or versatility
If multiple deep learning models are trained to handle different noise levels, then the system achieves adaptability to varying noise conditions, but the training time and computing resource requirements increase significantly
Solution Approach 1:
The patent creates a single universal deep learning model that can handle multiple noise levels through runtime noise estimation and dynamic parameter adjustment. This multi-functional approach eliminates the need to train and maintain multiple separate models for different noise conditions, significantly reducing training time and computing resource requirements while maintaining adaptability across varying noise levels.
Solution Approach 2:
The system performs self-adaptation by automatically estimating the noise level of the input data and adjusting the regularization parameter without requiring external intervention or pre-trained models for each noise level. This self-service mechanism enables the single model to serve multiple noise conditions efficiently.
3Object-affected harmful factors
If the regularization parameter is increased to reduce noise, then the noise level decreases, but the image becomes overly smooth and loses important details
Solution Approach 1:
The patent dynamically adjusts the regularization parameter based on the estimated noise level of the input data. By changing the regularization parameter adaptively rather than using a fixed high value, the system achieves effective noise reduction while preserving image sharpness and important details, resolving the contradiction between noise reduction and detail preservation.
Data Source
AI summary
The present disclosure is generally directed to systems and methods for generating de-noised MR images that are reconstructed from a hybridization of two separate image reconstruction pipelines, at least one of which includes the use of a neural network. Further, the amount of influence that the neural network reconstruction has on the hybrid reconstructed image is controlled via a regularization parameter that is selected based on an estimated noise level associated with the initial image acquisition, which can be calculated from pre-scan data.