Adversarial Network for Retrospective MRI Artifact Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning-based retrospective MRI artifact correction methods require supervised training with paired data, which is impractical due to the challenge of generating a wide range of artifacts, especially for pediatric subjects prone to motion-related artifacts.
Innovation Solution
A trained adversarial network is used for retrospective MRI artifact correction, employing unsupervised training techniques with unpaired artifact-free and artifact-containing MRI images, enabling efficient identification and removal of artifacts without the need for paired data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised training with paired data is used for deep learning-based MRI artifact correction, then the correction accuracy can be improved, but the data acquisition complexity and time increase significantly
Solution Approach 1:
Instead of requiring paired artifact-free and artifact-containing images for supervised training, the invention inverts the approach by using only artifact-containing images with motion trajectories to train the network. The system reconstructs artifact-free images by leveraging the motion information and the adversarial training framework, eliminating the need to acquire or generate paired data while maintaining correction effectiveness
2Reliability
If paired data is collected for supervised training, then the training effectiveness improves, but the complexity of generating diverse artifacts increases
Solution Approach 1:
The system uses the artifact-containing images themselves, along with their associated motion trajectories, to train the correction network. The motion information extracted from the artifact images serves as the training signal, allowing the system to self-train without requiring external paired data or complex artifact generation processes. The adversarial framework enables the network to learn from the imperfect data and reconstruct artifact-free images autonomously
3Ease of manufacture
If unpaired data is used for training, then the data collection process is simplified, but the challenge of matching corresponding features between images increases
Solution Approach 1:
The system performs preliminary extraction of motion trajectories from the artifact-containing images before the actual training process. By pre-processing the images to obtain motion information (such as displacement fields or motion vectors), the system prepares the training data in advance, making the subsequent adversarial training more effective. This preliminary action resolves the feature matching problem by providing explicit motion correspondence information rather than relying on the network to learn it from unpaired images
Data Source
AI summary
A method for performing retrospective magnetic resonance imaging (MRI) artifact correction includes receiving, as input, an MRI image having at least one artifact; using a trained adversarial network for performing retrospective artifact correction on the MRI image, wherein the trained adversarial network is trained using unpaired artifact-free MRI images and artifact-containing MRI images; and outputting, by the trained adversarial network, a derivative MRI image related to the input, wherein the at least one artifact is corrected in the derivative MRI image.