Hierarchical MRI Reconstruction Network for Undersampling Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI reconstruction methods using deep learning networks face challenges in computational efficiency, memory requirements, and robustness to variations in MRI intensities and contrasts, leading to degraded image quality and increased complexity, especially when dealing with under-sampled multi-coil data from varying acquisition protocols and scanner models.
Innovation Solution
An iterative, hierarchal network for regularization that utilizes multiple mappings derived from reference data, such as coil sensitivity maps, to improve robustness and generalizability, reducing computational complexity and memory usage while maintaining image quality, by incorporating these mappings into the reconstruction framework.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If multiple unrolled iterations of reconstruction are performed to improve reconstruction quality, then reconstruction quality is improved, but computational time and memory requirements increase
Solution Approach 1:
The patent divides the reconstruction network into multiple hierarchical levels (coarse to fine), where each level processes features at different resolutions. This segmentation allows the network to capture both global and local features without requiring excessive iterations, thereby improving reconstruction quality while controlling computational cost.
Solution Approach 2:
The patent introduces a hierarchical dimension to the reconstruction process by processing features at multiple scales (coarse to fine resolutions). This dimensional approach enables the network to learn both low-frequency global structures and high-frequency local details in a single multi-scale pass, reducing the need for multiple sequential iterations.
2Adaptability or versatility
If U-net architectures are used to learn from heterogeneous datasets, then adaptability is improved, but computational complexity and memory requirements increase
Solution Approach 1:
The patent segments the feature processing into hierarchical levels with different resolutions. Each level processes features at its appropriate scale, allowing the network to handle heterogeneous data (different organs, protocols, resolutions) without requiring a uniformly complex architecture throughout. This reduces overall network complexity while maintaining adaptability.
Solution Approach 2:
The patent applies different processing strategies at different hierarchical levels: coarse levels capture global structures with simpler operations, while fine levels handle local details with more specialized processing. This local quality approach allows the network to be adaptive to various MRI conditions without uniformly increasing complexity across all layers.
3Reliability
If deep learning networks with enormous capacity are trained on large datasets, then robustness to MRI variability is improved, but training time and computational complexity increase
Solution Approach 1:
The hierarchical segmentation of the network allows training to proceed more efficiently by processing features at multiple resolutions simultaneously rather than requiring sequential processing. This reduces the effective training time while maintaining the network's capacity to learn from large heterogeneous datasets and achieve robustness to MRI variability.
Data Source
AI summary
For reconstruction in sampling-based imaging, such as reconstruction in MR imaging, an iterative, multiple-mapping based hierarchal machine-learned network reconstruction may produce artifact corrected images based on under-sampled scans. Two or more mappings may be used to reduce the presence of artifacts, in some cases including localized low-noise-contribution artifacts, relative to reconstructions based on fully-sampled scans.