MRI Data Processing With Phase-Normalized Complex Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In medical image processing using complex-valued neural networks, the output results are sensitive to phase modulation of the entire image, while phase information is crucial, leading to instability and reduced significance of absolute phase values.
Innovation Solution
Incorporating a division processing layer to divide input complex vector data by complex second vector data containing features, followed by a nonlinear layer and a multiplication processing layer to stabilize the learning process and improve image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If complex-valued neural network is used for medical image processing, then various applications become possible, but the output results become sensitive to phase modulation of the entire image
Solution Approach 1:
The patent introduces a phase normalization layer as an intermediary component between the input layer and subsequent processing layers. This layer receives the complex-valued input image and performs phase normalization to remove global phase modulation effects before the data is processed by convolutional layers, thereby stabilizing the output while preserving the ability to process various medical image types
Solution Approach 2:
The patent transforms the phase parameter of the complex-valued input data by applying phase normalization. Specifically, it calculates the phase of the input image and subtracts the mean phase value, thereby changing the phase parameter to be invariant under global phase modulation while maintaining the magnitude information
2Loss of information
If phase information is preserved in complex-valued neural network, then important phase information is maintained, but the absolute value of phase becomes less significant
Solution Approach 1:
The patent segments the complex-valued data into magnitude and phase components, and further processes the phase component by separating it into a normalization part (global phase) and a residual part (local phase variations). The normalization layer removes the global phase component while preserving local phase information, thereby maintaining important phase details while reducing the significance of absolute phase values
3Ease of manufacture
If standard complex-valued neural network is used, then signal processing capability is enhanced, but learning stability deteriorates
Solution Approach 1:
The patent applies phase normalization as a preliminary action before the main learning process. By pre-processing the input data to remove global phase modulation effects, it creates more stable learning conditions for subsequent convolutional layers, thereby improving learning stability while preserving the signal processing capabilities of complex-valued neural networks
Data Source
AI summary
A data processing device according to an embodiment includes a processing circuit. The processing circuit performs data processing using a learned model with a neural network including a division processing layer that divides input complex first vector data by complex second vector data containing features of the first vector data, a nonlinear layer that is disposed in the subsequent stage of the division processing layer and that performs a nonlinear operation, and a multiplication processing layer that is disposed in the subsequent stage of the nonlinear layer and that multiplies the input data by the second vector data.