MRI Reconstruction Using Convolution Kernels for Multi-Contrast Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Magnetic resonance imaging (MRI) technologies face challenges in efficiently reconstructing images for multiple contrasts, leading to prolonged scanning times and compromised image quality due to the need for repeated scans and repeated collection of high-frequency information.
Innovation Solution
The method involves undersampling and fullsampling techniques in MRI to collect k-space data, training convolution kernels based on central k-space data, and combining these to determine uncollected data, thereby reconstructing images for multiple contrasts efficiently and improving image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If repeated scans are performed to obtain images for multiple contrasts, then image quality and tissue structure information are improved, but scanning time is prolonged
Solution Approach 1:
The patent segments k-space into central region and peripheral region, collecting data from both regions separately. The central k-space data is collected through repeated scans to ensure high quality, while the peripheral k-space data is collected through undersampling to reduce scanning time. This segmentation allows different sampling strategies to be applied to different parts of the data collection process.
Solution Approach 2:
The patent applies partial sampling (undersampling) to the peripheral k-space region, collecting only a subset of the required data points. This partial action reduces the total scanning time while the missing data is reconstructed using convolution kernels trained on the fully sampled central region, achieving acceptable image quality without complete data collection.
2Measurement precision
If repeated collection of high-frequency information is performed for multiple contrasts, then image clarity and lesion detection accuracy are improved, but scanning efficiency deteriorates
Solution Approach 1:
The patent divides k-space into central and peripheral regions, where the peripheral region contains high-frequency information. By undersampling only the peripheral region and using convolution-based reconstruction, the patent reduces the repeated collection of high-frequency data while maintaining lesion detection capability through the trained convolution kernels that learn from fully sampled central region data.
Solution Approach 2:
The patent uses convolution kernels trained on fully sampled central k-space data to generate and fill in the undersampled peripheral k-space data. This copying approach allows the system to reconstruct high-frequency information without physically collecting it through repeated scans, thereby improving scanning efficiency while preserving image clarity.
3Measurement precision
If full sampling is performed for all k-space data for multiple contrasts, then image reconstruction accuracy is improved, but scanning time and data collection burden increase
Solution Approach 1:
The patent segments the k-space sampling process into two parts: full sampling of the central region and undersampling of the peripheral region. This segmentation allows the system to collect complete data for accurate convolution kernel training in the central region while reducing total data collection time through selective undersampling in the peripheral region.
Solution Approach 2:
The patent performs preliminary full sampling of the central k-space region to train convolution kernels before undertaking the main data collection process. This preliminary action creates a foundation of accurate reference data that enables subsequent undersampled peripheral region collection to achieve acceptable reconstruction accuracy without requiring complete full sampling of all k-space data.
Data Source
AI summary
Methods and devices for reconstructing magnetic resonance images for contrasts are provided. In an example, the method includes: for each channel for each contrast, collecting k-space data of a subject in the channel by scanning the subject in an undersampling manner, collecting central k-space data by scanning a k-space central region of the subject in k-a fullsampling manner, training a convolution kernel of respective phase encoding lines in the channel based on the central k-space data of the contrasts, and obtaining entire k-space data in the channel based on the convolution kernel of respective phase encoding lines in the channel and collected k-space data in the channels, and obtaining a respective magnetic resonance image for each of the contrasts by performing image reconstruction on the entire k-space data in each channel for the contrast.