Unsupervised Deep Learning MRI Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI reconstruction techniques, such as parallel imaging and compressed sensing, are computationally expensive and require ground-truth images for training, making them inefficient for applications where fully-sampled data is difficult or impossible to acquire, especially in dynamic contrast-enhanced MR scans.
Innovation Solution
An unsupervised deep-learning method trains a deep neural network using under-sampled data without ground truth images, employing a loss function that enforces data consistency and additional image features like sharpness and spatial-temporal low rank, allowing for real-time model updates and flexible loss function design.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If parallel imaging and compressed sensing are used for MRI reconstruction, then reconstruction speed is improved, but computational cost increases significantly
Solution Approach 1:
The patent replaces traditional iterative optimization algorithms (mechanical/computational processes) with a trained neural network model that performs reconstruction through direct forward propagation. The neural network learns the mapping from undersampled k-space to image domain during training, enabling fast reconstruction without expensive iterative optimization during inference, thus reducing computational cost while maintaining high speed.
2Speed
If supervised deep learning approaches are used for MRI reconstruction, then reconstruction speed is improved, but ground-truth images are required which are difficult or impossible to acquire in many applications
Solution Approach 1:
The patent inverts the traditional supervised learning approach by using unsupervised learning. Instead of requiring ground-truth images to train the network, the network is trained to reconstruct images from undersampled data while enforcing data consistency through a specially designed loss function. This inversion removes the dependency on ground-truth images, enabling application to dynamic contrast-enhanced and other scans where fully-sampled reference images are unavailable.
Solution Approach 2:
The neural network performs self-supervised learning by using its own outputs to compute the loss function. The network reconstructs images from undersampled k-space, then the loss is computed by comparing the reconstructed images against the actual undersampled measurements through data consistency constraints. This self-service mechanism eliminates the need for external ground-truth labels while still enabling effective training.
3Measurement precision
If fully-sampled scans are collected for training, then training data quality is improved, but acquisition time increases significantly making it clinically impractical
Solution Approach 1:
The patent applies partial action by training the neural network on undersampled data rather than requiring fully-sampled data. The loss function is designed to enforce data consistency with the available undersampled measurements, allowing the network to learn effective reconstruction patterns from partial data. This approach achieves sufficient training quality without the excessive time cost of acquiring fully-sampled training scans.
Data Source
AI summary
A method for magnetic resonance imaging performs unsupervised training of a deep neural network of an MRI apparatus using a training set of under-sampled MRI scans, where each scan comprises slices of under-sampled, unclassified k-space MRI measurements. The MRI apparatus performs an under-sampled scan to produce under-sampled k-space data, updates the deep neural network with the under-sampled scan, and processes the under-sampled k-space data by the updated deep neural network of the MRI apparatus to reconstruct a final MRI image.