Convolutional Residual Network for MRI Artifact Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI techniques face challenges in reducing scan time due to off-resonance artifacts, which cause image blurring and signal dropout, especially with non-stationary artifacts from long readouts, and existing correction methods are either inefficient or computationally complex.
Innovation Solution
A convolutional residual network is used to process 3D MRI images, correcting non-stationary off-resonance artifacts in real-time, allowing for faster scan trajectories without compromising image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If non-Cartesian trajectories with long readouts are used to reduce scan time, then scan time is reduced, but off-resonance artifacts increase causing image blurring and signal dropout
Solution Approach 1:
The patent applies preliminary action by correcting off-resonance artifacts through a convolutional neural network after image reconstruction but before final image display. The network is trained in advance on pairs of accelerated and reference images to learn the mapping from artifact-containing images to corrected images, enabling rapid correction without requiring preliminary adjustment of scan parameters
Solution Approach 2:
The patent introduces an intermediary solution by using a convolutional neural network as a mediator between the accelerated k-space data and the final corrected image. The network acts as an intermediate processing step that transforms the artifact-prone reconstructed image into a corrected image, separating the acceleration process from the final image quality requirement
2Manufacturing precision
If traditional autofocus correction methods are used to correct off-resonance artifacts, then image quality is improved, but computational complexity increases and processing time extends
Solution Approach 1:
The patent replaces the traditional mechanical/mathematical autofocus correction system with a data-driven neural network system. Instead of using iterative phase correction algorithms that require complex calculations of off-resonance frequency maps and phase unwrapping, the system substitutes these with a pre-trained convolutional neural network that directly maps artifact-containing images to corrected images through learned features
Solution Approach 2:
The patent uses copying by training the neural network on pairs of accelerated and reference images, where the reference images serve as copies of the ground truth. The network learns to replicate the quality of reference images from accelerated input images through supervised learning, copying the correction pattern from training data to test data
3Productivity
If more samples are collected per readout to reduce number of readouts, then scan time is reduced, but non-stationary off-resonance artifacts are introduced
Solution Approach 1:
The patent converts the harmful off-resonance artifacts into a beneficial training signal for the neural network. The artifacts present in accelerated images are not merely corrected but used as the basis for teaching the network what corrections are needed. The harmful artifacts become the very feature that enables the network to learn and apply appropriate corrections in new images
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The method reduces scan time by a factor of 2.46 while maintaining image quality, outperforming traditional autofocus correction in efficiency and accuracy, and can process images in under a minute, making it suitable for clinical use.
Implementation Method 1
the body is placed in a strong, spatially homogeneous, and time-invariant magnetic field B0 created by a polarizing magnet. This magnetic field is briefly oscillated using RF transmit coils to excite the hydrogen nuclei, causing them to precess at the Larmor frequency ω=−γB0
Implementation Method 2
Magnetic resonance imaging (MRI) is an important medical imaging modality for imaging soft tissue in the body
Implementation Method 3
to provide spatial encoding of the signal, a smaller magnitude linearly varying magnetic field, referred to as a gradient field G(t), is superimposed on the primary field B0, resulting in an applied field Br(t)=B0+r·G(t)
Implementation Method 4
A convolutional residual network is used to process 3D MRI images, correcting non-stationary off-resonance artifacts in real-time
Data Source
AI summary
A method for magnetic resonance imaging corrects non-stationary off-resonance image artifacts. A magnetic resonance imaging (MRI) apparatus performs an imaging acquisition using non-Cartesian trajectories and processes the imaging acquisitions to produce a final image. The processing includes reconstructing a complex-valued image and using a convolutional neural network (CNN) to correct for non-stationary off-resonance artifacts in the image. The CNN is preferably a residual network with multiple residual layers.