Image-to-Image Neural Network for MR Artifact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current magnetic resonance (MR) imaging techniques face challenges in effectively reducing motion artifacts, off-resonance effects, and trajectory infidelity, which lead to image blurring, ghosting, and distortion, especially in high-field scanners and fast imaging, and existing post-processing methods are limited by computational power and model inaccuracies.
Innovation Solution
A deep learning-based image-to-image neural network is trained to generate artifact-reduced MR images by learning from pairs of artifact-free and artifact-contaminated data, using auxiliary maps, sequence metadata, and discriminators to produce contrast-invariant features, allowing for real-time processing with lower computational burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex post-processing methods are used to reduce artifacts, then artifact reduction performance is improved, but computational power requirements increase
Solution Approach 1:
The patent transforms the complex post-processing approach into a deep learning-based image-to-image network that learns artifact patterns during training and automatically removes them during inference. This parameter change from model-based processing to data-driven processing reduces computational burden while maintaining or improving artifact reduction performance
Solution Approach 2:
The patent replaces traditional mechanical signal processing methods with a neural network-based system. The image-to-image network substitutes complex iterative algorithms with a trained model that performs artifact reduction in a single pass, significantly reducing computational power requirements
2Power
If simple post-processing methods are used, then computational power requirements are reduced, but artifact reduction performance is limited
Solution Approach 1:
The patent performs preliminary action by training the image-to-image network offline using paired artifact-free and artifacted images. During this training phase, the network learns complex artifact patterns and removal strategies. When deployed, the pre-trained network efficiently removes artifacts without requiring complex real-time computation, thus achieving both high performance and low computational power requirements
3Productivity
If model-based post-processing is used, then processing speed is improved, but model accuracy limitations reduce artifact reduction effectiveness
Solution Approach 1:
The patent substitutes model-based processing with a data-driven deep learning approach. The image-to-image network is trained on diverse artifact examples and learns to handle various artifact types and severities, overcoming the limitations of simplified models while maintaining fast processing speeds through efficient network architecture
Data Source
AI summary
For artifact reduction in a magnetic resonance imaging system, deep learning trains an image-to-image neural network to generate an image with reduced artifact from input, artifacted MR data. For application, the image-to-image network may be applied in real time with a lower computational burden than typical post-processing methods. To handle a range of different imaging situations, the image-to-image network may (a) use an auxiliary map as an input with the MR data from the patient, (b) use sequence metadata as a controller of the encoder of the image-to-image network, and/or (c) be trained to generate contrast invariant features in the encoder using a discriminator that receives encoder features.