QSM Reconstruction Using Unsupervised CycleGANs for Artifact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing QSM reconstruction methods face challenges such as high computational complexity, streaking artifacts, difficulty in hyperparameter tuning, and underestimation of susceptibility values due to the requirement of matched pairs in supervised learning, especially when training data lacks variability and structure.
Innovation Solution
A quantitative susceptibility mapping image processing method using an unsupervised learning-based neural network with a cycle-consistency generative adversarial network (cycleGAN) structure, trained on non-matching data, employing optimal transport theory, cycle-consistency loss, gradient difference loss, total variation loss, and adversarial loss to reconstruct QSM images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning methods are used for QSM reconstruction, then training accuracy can be improved with matched pairs, but susceptibility values are underestimated and data variability is reduced
Solution Approach 1:
The patent inverts the traditional supervised learning approach by using unsupervised learning with cycle-consistency loss. Instead of requiring ground truth QSM labels for training, the system trains the generator to reconstruct the input phase image through a cycle process, eliminating the need for matched pairs and preventing underestimation of susceptibility values.
Solution Approach 2:
The system uses self-service by enabling the neural network to train on its own data without external ground truth labels. The cycle-consistency mechanism allows the network to evaluate its own reconstruction quality by comparing the original input with the reconstructed output after passing through the generator and discriminator.
2Measurement precision
If multiple head orientation sampling (COSMOS) is used for dipole inversion, then reconstruction accuracy is improved, but acquisition time and subject burden increase significantly
Solution Approach 1:
The patent extracts the essential information needed for accurate QSM reconstruction from a single head orientation, eliminating the need for multiple orientations. The unsupervised learning framework with cycle-consistency loss extracts sufficient constraints from the magnitude image and phase image relationship to achieve accurate reconstruction without the time-consuming multi-orientation acquisition.
Solution Approach 2:
The system changes the fundamental parameter from multiple head orientations to a single orientation with enhanced processing. By using the cycle-consistency generative adversarial network, the method transforms the reconstruction problem to work effectively with reduced acquisition parameters while maintaining or improving accuracy.
3Productivity
If classical dipole inversion algorithms are used, then computational speed is improved, but streaking artifacts and hyperparameter tuning difficulties increase
Solution Approach 1:
The patent replaces classical iterative dipole inversion algorithms with a deep learning-based generative adversarial network. This substitution eliminates the need for manual hyperparameter tuning and streaking artifact suppression techniques, as the neural network learns the inversion process directly from data, achieving both speed and artifact-free reconstruction.
Solution Approach 2:
The system uses a composite approach combining generator and discriminator networks in a cycle-consistency framework. This composite structure integrates multiple functions (reconstruction, validation, and artifact suppression) into a unified system that outperforms individual classical algorithms.
4Device complexity
If supervised learning with matched pairs is used, then training data structure is simplified, but adaptability to varying data distributions is reduced
Solution Approach 1:
The unsupervised learning framework provides universality by enabling the model to adapt to various data distributions without requiring matched pairs. The cycle-consistency mechanism and adversarial training make the system versatile across different imaging conditions, anatomies, and scanner parameters, eliminating the need for dataset-specific training.
Data Source
AI summary
Disclosed is a quantitative susceptibility mapping image processing method using an unsupervised learning-based neural network and an apparatus therefor. The quantitative susceptibility mapping image processing method includes receiving a phase image and a magnitude image for reconstructing the quantitative susceptibility mapping image, and reconstructing the quantitative susceptibility mapping image corresponding to the received phase image and the received magnitude image using an unsupervised learning-based neural network, and the neural network may be generated based on an optimal transport theory.