Reinforcement Learning for MRI Trajectory Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI techniques face inefficiencies in data acquisition due to slow scan times and lack of optimal undersampling patterns, which are often predetermined and do not effectively utilize real-time data for improved image quality.
Innovation Solution
The implementation of online deep reinforcement learning to dynamically update the undersampling pattern during an MRI scan, using a deep convolutional neural network and fully connected layers to optimize data collection based on real-time image quality metrics, allowing for arbitrary sampling trajectories and minimizing L2 error.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If predetermined undersampling patterns are used, then device complexity is reduced, but image quality and scanning efficiency deteriorate because real-time data is not utilized
Solution Approach 1:
The patent implements feedback by using reinforcement learning where the agent continuously observes image quality metrics from reconstructed images and adjusts the undersampling pattern accordingly. The reward function provides feedback signal that guides the agent to optimize sampling decisions based on real-time scanning data, resolving the contradiction between simple predetermined patterns and efficient adaptive sampling.
Solution Approach 2:
The system performs self-service by autonomously determining optimal sampling patterns without external intervention. The reinforcement learning agent independently evaluates the current scanning state, assesses image quality metrics, and selects subsequent sampling trajectories autonomously, eliminating the need for complex external optimization systems while improving scanning efficiency.
2Manufacturing precision
If more data is acquired to improve image quality, then manufacturing precision improves, but loss of time increases due to longer scan durations
Solution Approach 1:
The patent applies partial action by acquiring only the necessary amount of data needed to achieve acceptable image quality rather than complete sampling. The reinforcement learning agent determines when sufficient data has been collected by evaluating image quality metrics, allowing the scan to terminate early when quality thresholds are met, thus reducing scan time while maintaining adequate image quality.
Solution Approach 2:
The system dynamically changes sampling parameters based on real-time feedback. The agent adjusts the undersampling ratio, sampling trajectories, and acquisition priorities according to the evolving image quality metrics, enabling optimal balance between data acquisition volume and scan time reduction throughout the scanning process.
3Productivity
If online reinforcement learning is implemented to optimize sampling patterns, then productivity improves, but device complexity increases due to additional computational requirements
Solution Approach 1:
The patent applies preliminary action by pre-training the reinforcement learning agent offline using simulated scanning data and synthetic images. This pre-training phase prepares the agent with prior knowledge of optimal sampling strategies, reducing the computational burden during actual online scanning. The agent only requires lightweight inference and adaptation during real scans, balancing online optimization capability with acceptable computational complexity.
Data Source
AI summary
A magnetic resonance imaging scan performs an MRI acquisition using an undersampling pattern to produce undersampled k-space data; adds the undersampled k-space data to aggregate undersampled k-space data for the scan; reconstructs an image from the aggregate undersampled k-space data; updates the undersampling pattern from the reconstructed image and aggregate undersampled k-space data using a deep reinforcement learning technique defined by an environment, reward, and agent, where the environment comprises an MRI reconstruction technique, where the reward comprises an image quality metric, and where the agent comprises a deep convolutional neural network and fully connected layers; and repeats these steps to produce a final reconstructed MRI image for the scan.