Deep Learning MRI Reconstruction With K-Space High-Frequency Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based MRI reconstructions often suffer from over-smoothing, leading to a loss of high-frequency details and textures, and existing generative adversarial network-based methods are prone to artificial structure hallucinations and training difficulties.
Innovation Solution
A post-processing technique that refines DL-based MRI reconstructions by estimating a k-space null-space convolutional kernel from auto-calibration lines, using convex optimization to improve k-space domain projections, and optionally incorporating virtual coil augmentation for enhanced refinement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep learning methods are used for MRI reconstruction, then reconstruction speed and overall image quality are improved, but high-frequency details and textures are lost due to over-smoothing
Solution Approach 1:
The invention separates the reconstruction process into two distinct stages: first, a deep learning-based unrolled neural network performs rapid reconstruction; second, a dedicated refinement network processes the preliminary result to recover high-frequency details. This segmentation allows each stage to optimize for its specific function without compromise.
Solution Approach 2:
The refinement network acts as an intermediary between the deep learning reconstruction and the final image output. It takes the preliminary reconstruction as input and transforms it by adding back high-frequency details that were lost during the initial reconstruction process.
2Manufacturing precision
If generative adversarial networks are used to enhance high-frequency details, then image realism is improved, but training difficulty increases and artificial structures may be hallucinated
Solution Approach 1:
The refinement network uses a simple autoregressive architecture with basic convolutional layers and pixel-shuffle operations, avoiding the complex adversarial training mechanisms of GANs. This simpler approach is easier to train and less prone to generating artificial structures.
Solution Approach 2:
The refinement network is trained in a self-supervised manner using the preliminary reconstruction and the corresponding ground truth image. The training process automatically learns to enhance high-frequency details without requiring manual intervention or complex hyperparameter tuning.
3Manufacturing precision
If generative adversarial networks are used for detail enhancement, then high-frequency visual details are improved, but reliability decreases due to hallucination of artificial structures
Solution Approach 1:
The refinement network uses a pixel-shuffle operation that provides feedback between different spatial resolutions, allowing the network to maintain consistency with the input structure while enhancing details. The loss function directly compares the refined output with the ground truth, ensuring accurate structure recovery.
Solution Approach 2:
The network employs pixel-shuffle operations that change the spatial arrangement of pixels to upscale the image while preserving structural information. This parameter transformation approach maintains reliability by ensuring that enhanced details are consistent with the underlying anatomy.
Data Source
AI summary
A method for magnetic resonance imaging (MRI) includes acquiring under-sampled k-space measurements from an MRI apparatus using multiple receiver coils; reconstructing an MRI image from the under-sampled k-space measurements and coil sensitivity maps using an unrolled neural network; generating reconstructed multi-coil k-space data from the MRI image by multiplying the MRI image by the coil sensitivity maps followed by performing a Fourier transform; estimating a k-space null-space convolutional kernel from fully-sampled k-space measurements in autocalibration signal lines of the under-sampled k-space measurements; solving a convex optimization problem to produce refined k-space data from the k-space null-space kernel, the under-sampled k-space measurements, and the reconstructed multi-coil k-space data; and producing a refined MRI image from the refined k-space data by performing an inverse Fourier transform followed by a coil combination using the coil sensitivity maps.