High-efficient Bloch simulation of magnetic resonance imaging sequences based on deep learning

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence

Background:

• Bloch simulation is crucial for magnetic resonance imaging (MRI) development but computationally intensive.
• Existing graphics processing unit (GPU) acceleration methods struggle with large-scale, high-accuracy simulations.

Purpose of the Study:

• To develop a deep learning-based simulator, Simu-Net, to accelerate Bloch simulations.
• To improve the efficiency of MRI pulse sequence optimization and deep learning applications.

Main Methods:

• Developed Simu-Net, an end-to-end convolutional neural network.
• Trained Simu-Net using synthetic data from traditional Bloch simulations.
• Incorporated dynamic convolution and position encoding templates for enhanced accuracy and efficiency.

Main Results:

• Simu-Net achieved hundreds of times acceleration compared to GPU-based MRI simulation software.
• Demonstrated accuracy and robustness across traditional and advanced MRI pulse sequences.
• Successfully generated training data for deep learning-based T2 mapping, yielding comparable results to conventional methods.

Conclusions:

• Simu-Net demonstrates the potential of deep learning to approximate MRI's forward physical process.
• This approach can significantly increase the efficiency of Bloch simulations.
• Highlights future applications in optimizing MRI pulse sequences and deep learning methods.