MRI k-space interpolation using CNN non-linear mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current magnetic resonance imaging (MRI) techniques using linear convolution kernels for reconstructing undersampled k-space data are limited by their ad-hoc nature and require additional scans for calibration, hindering real-life applicability and being prone to noise amplification.
Innovation Solution
A method employing a machine learning algorithm, specifically a convolutional neural network (CNN), is used to learn non-linear mapping functions from calibration data, allowing for the estimation of missing k-space data without relying on external databases or specific dataset designs, thereby improving noise resilience and reconstruction quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If linear convolution kernels are used for k-space interpolation, then the reconstruction process is simple and fast, but noise amplification occurs and additional calibration scans are required
Solution Approach 1:
The patent replaces the traditional linear mechanical convolution operation with a non-linear machine learning-based mapping function. This substitution allows the system to learn complex non-linear relationships from calibration data, improving noise resilience while maintaining reconstruction efficiency through the use of pre-trained models that can be applied rapidly to new data.
Solution Approach 2:
The patent changes the fundamental parameter of the interpolation approach from linear to non-linear by using machine learning models. This parameter change enables the system to adapt to varying noise conditions and data characteristics, improving reliability without sacrificing the computational efficiency needed for practical application.
2Measurement precision
If additional calibration scans are performed to generate convolution kernels, then reconstruction accuracy improves, but scan time increases
Solution Approach 1:
The patent applies preliminary action by performing the complex non-linear learning process during an initial calibration phase, and then using the pre-trained model for rapid reconstruction of subsequent scans. This allows high accuracy to be achieved without repeating the time-consuming calibration process for each new scan, thereby reducing overall scan time while maintaining reconstruction precision.
Solution Approach 2:
The machine learning model serves itself by learning from the calibration data and automatically adapting to the specific imaging conditions. Once trained, the model can independently perform high-accuracy reconstruction without requiring continuous human intervention or repeated calibration scans, thus reducing time loss while maintaining precision.
3Measurement precision
If non-linear approaches with virtual channels are used, then reconstruction quality improves, but the method becomes ad-hoc and dataset-specific
Solution Approach 1:
The patent creates a universal non-linear reconstruction framework using machine learning that can be applied across different datasets and imaging conditions. The model learns generalizable patterns from calibration data that transfer to various scenarios, making the high-quality reconstruction method broadly applicable rather than ad-hoc and dataset-specific.
Solution Approach 2:
The machine learning approach incorporates feedback mechanisms where the model continuously learns from calibration data and adjusts its parameters to optimize performance. This feedback loop enables the system to adapt to different datasets while maintaining high reconstruction quality, overcoming the limitation of ad-hoc methods that lack such adaptive learning capabilities.
Data Source
AI summary
Methods for reconstructing images from undersampled k-space data using a machine learning approach to learn non-linear mapping functions from acquired k-space lines to generate unacquired target points across multiple coils are described.