Real-Complex Neural Networks for MRI Phase Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical imaging technologies face challenges in balancing image quality with reduced acquisition time and radiation dosages, particularly in magnetic resonance imaging (MRI) systems, where super-resolution methods for phase information are ill-defined and complex-valued datasets are not effectively processed with standard real-valued convolutional kernels.
Innovation Solution
A method combining neural networks to process both real-number-based and complex-number-based images, utilizing subnetworks to enhance image resolution and reduce artifacts, involving a combination of real and complex deep neural networks to generate higher resolution images from lower resolution inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If super-resolution methods are applied to phase information, then image resolution is improved, but the accuracy of phase information is compromised because standard real-valued convolutional kernels cannot effectively process complex-valued datasets
Solution Approach 1:
The patent changes the fundamental parameter of the neural network from real-valued to complex-valued processing. By using complex convolutional kernels that can process complex-valued inputs, the system maintains phase information accuracy while achieving super-resolution enhancement of phase images.
Solution Approach 2:
The patent employs a hybrid neural network architecture that combines real-valued and complex-valued processing components. The network uses complex convolutional layers for phase information processing while maintaining real-valued pathways for magnitude information, creating a composite system that handles both aspects optimally.
2Measurement precision
If higher resolution imaging is used to improve shim behavior accuracy, then measurement precision is improved, but acquisition time increases
Solution Approach 1:
The patent applies super-resolution neural networks to process low-resolution phase images acquired during rapid shim scans, transforming them into high-resolution images before final analysis. This preliminary enhancement allows accurate shim behavior assessment without requiring time-consuming high-resolution acquisitions.
Solution Approach 2:
The system creates high-resolution copies of low-resolution phase images through neural network-based super-resolution. Instead of acquiring actual high-resolution data that would take longer, the network generates synthetic high-resolution versions that maintain the necessary measurement precision.
3Productivity
If low-resolution imaging with large voxels is used to reduce scan time, then productivity is improved, but measurement precision deteriorates due to partial volume error
Solution Approach 1:
The patent performs super-resolution enhancement as a preliminary processing step on the low-resolution phase images acquired during fast scans. By enhancing the resolution before analyzing shim behavior, the system recovers detail that would otherwise be lost to partial volume effects, maintaining measurement precision despite rapid acquisition.
Data Source
AI summary
The present disclosure relates to a real-number-based neural network operating in combination with a complex-number-based neural network to perform image processing (e.g., using phase-based medical images). In one embodiment, a method includes, but is not limited to, applying, to inputs of a first trained neural network trained to process real-number-based images, first image data generated from real-number-based measurements obtained by imaging a subject; applying, to inputs of a second trained neural network trained to process complex-number-based images, second image data generated from complex-number-based measurements obtained by imaging the subject; and combining a first output of the first trained neural network and a second output of the second trained neural network to produce a combined image, based on the first image data and the second image data.