Deep Learning MRI Reconstruction for Speed and Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging technologies face challenges in balancing temporal and spatial resolution, particularly in dynamic MRI applications, where high acceleration factors lead to image blurring, noise, and artifacts due to limited spatial sparsity and complex temporal behaviors.
Innovation Solution
A system and method that utilize a non-patient-specific analytical model for image reconstruction, constraining the process to adhere to a physical or physiological model while allowing deviations to accommodate underlying pathology, thereby improving image quality and reducing noise and undersampling artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parallel imaging techniques are used to accelerate MRI acquisition, then imaging speed is improved, but image quality deteriorates with blurring and artifacts
Solution Approach 1:
The patent changes the fundamental parameter of image reconstruction by transitioning from conventional parallel imaging methods to a deep learning-based approach. The neural network is trained to reconstruct high-quality images from highly undersampled k-space data, fundamentally altering how the reconstruction parameter is handled to achieve both speed and quality improvements simultaneously.
Solution Approach 2:
The patent replaces the mechanical/mathematical reconstruction algorithms (parallel imaging, compressed sensing) with an intelligent system - a deep learning neural network. This substitution allows the system to learn complex patterns and relationships in the data, enabling high-quality reconstruction from highly accelerated data without the artifacts typical of conventional methods.
2Productivity
If compressed sensing is used for high acceleration, then imaging speed is improved, but spatial resolution deteriorates due to limited spatial sparsity
Solution Approach 1:
The patent replaces the sparsity-based compressed sensing approach with a data-driven deep learning approach. Instead of relying on the mathematical assumption of spatial sparsity, the neural network learns the underlying structure and patterns directly from training data, enabling high-quality reconstruction even when spatial sparsity is limited.
Solution Approach 2:
The patent fundamentally changes the reconstruction parameter from sparsity-driven (compressed sensing) to learning-driven (deep learning). This allows the system to achieve high acceleration factors while maintaining spatial resolution by learning complex patterns that traditional sparsity methods cannot capture.
3Productivity
If temporal basis functions are learned from low-resolution images, then reconstruction speed is improved, but fidelity deteriorates due to errors in temporal behavior representation
Solution Approach 1:
The patent applies preliminary action by training the neural network offline on large datasets before actual imaging. The network learns accurate temporal patterns and relationships during the training phase, so that during actual high-speed imaging, it can directly apply this learned knowledge without needing to learn from low-resolution images at runtime, thereby maintaining both speed and fidelity.
Solution Approach 2:
The patent replaces the iterative temporal basis learning approach with a pre-trained deep learning model. Instead of learning temporal patterns on-the-fly from low-resolution images during reconstruction, the system uses a neural network that has already learned accurate temporal representations from extensive training data, eliminating the fidelity issues associated with learning from low-resolution 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
This approach enhances image sharpness, spatial resolution, and temporal resolution, providing clinically usable images with reduced noise and improved fidelity, especially in the presence of motion or complex pathologies, compared to traditional methods like parallel imaging and compressed sensing.
Implementation Method 1
magnetic resonance imaging system includes a magnet system configured to generate a polarizing magnetic field about at least a portion of a subject arranged in the MRI system and a magnetic gradient system including a plurality of magnetic gradient coils configured to apply at least one magnetic gradient field to the polarizing magnetic field
Implementation Method 2
a radio frequency (RF) system configured to apply an RF field to the subject and to receive magnetic resonance signals from the subject using a coil array
Implementation Method 3
a magnetic gradient system including a plurality of magnetic gradient coils configured to apply at least one magnetic gradient field to the polarizing magnetic field
Data Source
AI summary
A system and method for reconstructing a series of images of a subject includes acquiring medical image data from the subject with a medical imaging system and reconstructing a series of images of the subject from the acquired medical image data set. The reconstructing includes enforcing general adherence to a non-patient-specific signal model that describes a dependency of image intensity values on at least one variable that is associated with a physical or physiological property by constraining reconstruction of individual images in the series of images using the non-patient-specific model. The reconstructing also includes preserving information in the series of images that deviate from the non-patient-specific model by controlling a requirement of consistency with the non-patient-specific model.