Iterative Hierarchical Medical Image Reconstruction with Auxiliary Information
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning-based medical image reconstruction methods, particularly in MRI, face challenges with high computational complexity and reduced robustness due to the need for large datasets and complex architectures like U-nets, which are inefficient and require extensive training time.
Innovation Solution
An iterative, hierarchal network for regularization that incorporates auxiliary information, such as noise maps or gfactor maps, within the reconstruction process to improve robustness and reduce computational complexity by using a machine-learned model with a hierarchical design.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If U-net architectures are used for image regularization, then learning from heterogeneous datasets is effective, but the overall size of feature maps increases, resulting in decreased receptive field and increased computational complexity
Solution Approach 1:
The network is divided into multiple processing stages with separate feature extraction and regularization components. Instead of using a single large U-net architecture, the method segments the processing into distinct modules that handle different aspects of image regularization independently, reducing overall computational complexity while maintaining adaptability
Solution Approach 2:
The method extracts and utilizes auxiliary information (such as intermediate feature maps or specific image properties) separately from the main reconstruction process. This extraction allows the network to leverage additional data without increasing the overall network size or computational burden proportionally
2Reliability
If deep learning networks with enormous capacity are used to cover MRI variability, then robustness is improved, but training time and computational complexity increase
Solution Approach 1:
The network is pre-trained or configured with prior knowledge about MRI variability from training data. This preliminary action allows the network to handle diverse MRI protocols and scanner models without requiring enormous capacity or extensive retraining for each new dataset, thus reducing both model size and training time
Solution Approach 2:
The regularized network is designed to perform multiple functions simultaneously - handling different MRI protocols, scanner models, and image types within a single unified architecture. This multi-functionality eliminates the need for separate specialized networks for each MRI variability scenario, reducing overall computational complexity while maintaining robustness
3Manufacturing precision
If multiple unrolled iterations of reconstruction are performed, then reconstruction quality is improved, but computational time and memory requirements increase proportionally
Solution Approach 1:
Instead of performing multiple full unrolled iterations, the method applies a single iteration or partial iteration with carefully selected regularization parameters that achieve sufficient reconstruction quality. This partial action maintains acceptable reconstruction quality while significantly reducing computational time and memory requirements
Solution Approach 2:
The method optimizes regularization parameters and network configuration to achieve the best possible reconstruction quality with minimal iterations. By carefully tuning parameters such as regularization strength, network depth, and learning rates, the system achieves high-quality reconstruction with fewer computational steps, improving processing speed without sacrificing quality
Data Source
AI summary
For reconstruction in medical imaging, such as reconstruction in MR imaging, an iterative, hierarchal network for regularization may decrease computational complexity. To further maintain computational complexity while improving robustness, auxiliary information is used in the regularization and corresponding reconstruction. The auxiliary information is in put to the machine-learned network.