LOUPE Neural Network Optimizes MRI Under-sampling Pattern
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Compressive sensing techniques, such as those used in Magnetic Resonance Imaging (MRI), face challenges in accurately producing high-quality images from under-sampled data due to aliasing artifacts and the computational demands of iterative reconstruction methods, which are not optimized in a data-driven fashion.
Innovation Solution
A machine learning-based approach that collectively optimizes the under-sampling pattern and reconstruction model using a neural network, enabling efficient and fast image reconstruction by identifying the optimal sampling pattern and corresponding reconstruction model for MRI images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative numerical optimization methods are used for reconstruction, then image quality can be improved, but computational complexity and time consumption increase significantly
Solution Approach 1:
The patent replaces iterative numerical optimization methods with a trained neural network model for image reconstruction. The neural network learns the mapping from under-sampled k-space data to high-quality images during training, and during inference, it directly produces reconstructions without requiring iterative optimization. This substitution of mechanical/mathematical iteration with a learned model reduces computational complexity while maintaining image quality.
Solution Approach 2:
The patent performs preliminary training of the neural network model using synthetic data generated from fully-sampled images. During this offline training phase, the network learns optimal reconstruction strategies. When actual under-sampled data arrives, the pre-trained model can immediately perform reconstruction without requiring iterative optimization, thus solving the computational complexity issue while maintaining high image quality.
2Measurement precision
If regularized regression with non-convex objective functions is used, then reconstruction accuracy improves, but computational demands increase
Solution Approach 1:
The patent substitutes the computationally intensive regularized regression process with a neural network that has been trained to perform the same reconstruction task. The neural network learns the complex optimization landscape during training and can perform reconstructions efficiently during inference, avoiding the need to solve non-convex optimization problems in real-time while maintaining high reconstruction accuracy.
3Productivity
If standard under-sampling patterns are used, then data collection efficiency is improved, but reconstruction quality deteriorates due to aliasing artifacts
Solution Approach 1:
The patent changes the approach to under-sampling by training the neural network to learn optimal sampling patterns rather than using fixed standard patterns. The network can adaptively determine which k-space locations to sample based on the specific imaging scenario, and learns corresponding reconstruction strategies. This flexibility allows maintaining high data collection efficiency while avoiding the aliasing artifacts that plague standard fixed patterns.
Solution Approach 2:
The patent enables different parts of k-space to be sampled with different densities based on their importance. The neural network learns to identify which regions require denser sampling and which can be sparsely sampled, creating a non-uniform sampling pattern optimized for the specific imaging task. This local optimization improves reconstruction quality while maintaining overall data collection efficiency.
4Loss of time
If the number of samples is reduced below Shannon-Nyquist rate, then scanning time is reduced, but image quality deteriorates due to insufficient data
Solution Approach 1:
The patent performs preliminary training of the neural network using synthetic data generated from fully-sampled images. This pre-training enables the network to learn how to reconstruct high-quality images from under-sampled data. During actual scanning, even though fewer samples are collected (reducing scanning time), the pre-trained network can compensate for the missing data and produce high-quality reconstructions.
Solution Approach 2:
The patent uses synthetic training data that copies the statistical properties and structures of real fully-sampled images. By training on these synthetic copies, the neural network learns to recognize and reconstruct the essential features of real images even when actual data is under-sampled, thus maintaining image quality while reducing scanning time.
Data Source
AI summary
Systems and methods are disclosed for optimizing a sub-sampling pattern for efficient capture of a sub-sampled image to be reconstructed to form a high-resolution image, in a data-driven fashion. For example, Magnetic Resonance Imaging (MRI) scans can be accelerated by under-sampling in k-space (i.e., the Fourier domain). Since the reconstruction model's success depends on the sub-sampling pattern, optimization of the sub-sampling pattern can be combined with optimization of the model, for a given sparsity constraint, using an end-to-end learning operation. A machine-learning model may be trained using full-resolution training data that are under-sampled retrospectively, yielding a sub-sampling pattern and reconstruction model that are customized to the type of images represented in the training data. The disclosed Learning-based Optimization of the Under-sampling PattErn (LOUPE) operations may implement a convolutional neural network architecture, appended with a forward model that encodes the under-sampling process.