MRI Under-sampling Using Neural Network k-space Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for under-sampling in magnetic resonance imaging (MRI) struggle to produce high-quality images efficiently, often erasing minute spatial changes and failing to effectively remove artifacts, which is problematic for clinical applications like early-stage cancer detection.
Innovation Solution
An under-sampling apparatus using machine learning that differentiates between a fully sampled center region and an under-sampled peripheral region of the k-space image, employing a pre-learned neural network for image reconstruction, including inverse and Fourier transforms to correct and generate high-resolution MR images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If compressed sensing method is used for under-sampled MRI reconstruction, then imaging time is reduced, but image quality deteriorates due to erasure of minute spatial changes
Solution Approach 1:
The patent divides the k-space into multiple regions (first region, second region, third region) with different sampling densities. The first region (center) uses full sampling, the second region uses under-sampling, and the third region uses full sampling again. This segmentation allows the system to reduce overall imaging time through under-sampling while preserving image quality by maintaining full sampling in critical regions and using deep learning to recover information from under-sampled regions.
Solution Approach 2:
The patent employs a deep learning network that is pre-trained using full-sampled MRI data. The network learns the mapping between under-sampled and fully sampled k-space data in advance, enabling it to reconstruct high-quality images from under-sampled data during actual imaging. This preliminary training action allows the system to achieve both time reduction and quality preservation during the actual scanning process.
2Loss of time
If deep learning network is used for image reconstruction, then imaging time is reduced, but artifacts in reconstructed image increase
Solution Approach 1:
The patent applies different sampling strategies to different regions of the k-space. The first region (center) and third region use full sampling to provide high-quality reference data, while the second region uses under-sampling where artifacts are less critical. This local differentiation allows the deep learning network to reconstruct images with reduced artifacts by learning from the high-quality full-sampled regions and applying that knowledge to the under-sampled regions.
Solution Approach 2:
The patent uses a multi-region k-space sampling approach where the fully sampled first and third regions provide feedback information to the deep learning network. The network learns from these high-quality reference regions and uses the learned patterns to correct and reduce artifacts in the under-sampled second region, creating a feedback mechanism that improves overall image quality while maintaining time reduction benefits.
3Measurement precision
If full sampling is performed in k-space, then image quality is maintained, but imaging time increases
Solution Approach 1:
The patent segments the k-space into three distinct regions with different sampling approaches. The first region (center portion) and third region use full sampling to preserve critical image information and provide training data for the deep learning network. The second region uses under-sampling to reduce the total number of measurements. This segmentation strategy maintains image quality by fully sampling critical regions while reducing overall imaging time through under-sampling in less critical regions.
Solution Approach 2:
The patent applies full sampling (excessive action) to specific critical regions of the k-space (first and third regions) while using under-sampling (partial action) in the second region. This partial application of full sampling ensures that the most important image information is captured with high fidelity, while the deep learning network compensates for the reduced sampling in other regions, achieving a balance between image quality and imaging time.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach enables the acquisition of high-quality MR images while significantly reducing imaging time, preserving critical spatial details necessary for clinical diagnostics.
Implementation Method 1
a first image converter that converts the first k-space image to a first MR image
Implementation Method 2
a second image converter that converts the second MR image to the second k-space image
Data Source
AI summary
An under-sampling apparatus for MR image reconstruction by using machine learning and a method thereof, an MR image reconstruction device by using machine learning and a method thereof, and a recoding medium thereof are disclosed. The disclosed under-smapling apparatus includes: a setting portion that sets a region corresponding to a center of the k-space image as a first region and remaining regions as a second region; and an under-sampling portion that full-samples the first region and under-samples the second region, wherein in the under-sampling performed in the second region, lines are selected at regular intervals and then only the selected line is full-sampled. According to the under-sampling apparatus, a high-resolution MR image can be acquired while reducing imaing time.