Neural Network Dimensionality Reduction for MR Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for processing magnetic resonance (MR) images face challenges due to the high computational and memory demands of training neural networks with 3D or higher-dimensional data, limited availability of high-quality 3D training data, and trade-offs between temporal and spatial resolutions in dynamic MR imaging, which hinder the generation of high-quality processed images.
Innovation Solution
A neural network model is trained with lower-dimensional data and applied to process MR images of higher dimensions, allowing for input and output in the higher dimension without requiring higher-dimensional training data, and utilizing multiple combinations of dimensions to increase inference confidence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If neural networks are trained with high-dimensional MR data to improve image processing quality, then manufacturing precision is improved, but use of energy and computational resources increases
Solution Approach 1:
The patent segments the high-dimensional MR data processing task by training separate neural networks on different dimensionality-reduced representations of the data. Instead of training one large network on full 3D or higher-dimensional data, the system divides the processing into multiple smaller networks that each handle specific aspects of the data in reduced dimensions, thereby reducing overall computational resource requirements while maintaining image processing quality.
Solution Approach 2:
The patent applies dimensionality reduction techniques to transform high-dimensional MR data into lower-dimensional representations for training neural networks. By changing the dimensional representation of the input data (e.g., projecting 3D data onto 2D planes or using other dimensionality reduction methods), the system reduces computational strain during training while preserving essential features needed for high-quality image processing.
2Manufacturing precision
If neural networks are trained with high-dimensional MR data to improve image processing quality, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
The system segments the complex high-dimensional processing task into multiple simpler neural networks operating in reduced dimensions. This segmentation reduces the complexity of individual network components and their training requirements, making the overall system more manageable despite achieving high processing quality through the coordinated operation of multiple networks.
Solution Approach 2:
By transforming the problem from high-dimensional to lower-dimensional spaces for training purposes, the patent reduces the complexity of the training process and the computational infrastructure required. The dimensionality change allows standard neural network architectures to be used effectively without requiring complex custom-designed systems.
3Reliability
If more training data is used to improve neural network performance, then reliability is improved, but loss of time increases
Solution Approach 1:
The patent reduces training time by transforming the training data into lower-dimensional representations. This dimensionality reduction allows the neural networks to be trained faster while still maintaining reliable performance on the original high-dimensional MR images. The essential features and patterns are preserved in the reduced dimensional space, enabling efficient training without sacrificing model reliability.
Solution Approach 2:
By dividing the training task into multiple smaller networks trained on dimensionality-reduced data, the system reduces the overall training time required to achieve reliable performance. Each smaller network trains faster than a single large network would, and their combined results provide reliable image processing performance.
Data Source
AI summary
A magnetic resonance (MR) image processing system is provided. The system includes an MR image processing computing device that includes at least one processor. The processor is programmed to execute a neural network model configured to receive crude MR data as an input and output processed MR images associated with the crude MR data, the crude MR data and the processed MR images having the first number of dimensions. The processor is also programmed to receive a pair of pristine data and corrupted data both having a second number of dimensions lower than the first number of dimensions. The corrupted data are the pristine data added with primitive features. The processor is further programmed to train the neural network model using the pair of the pristine data and the corrupted data. The trained neural network model is configured to change primitive features associated with the crude MR data.