Simulation of spin dephasing in arbitrary susceptibility fields using physics-informed neural networks.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Computational Physics
• Artificial Intelligence

Background:

• MRI signal dynamics are complex, influenced by dephasing, diffusion, and susceptibility effects.
• Analytical solutions for MRI signal dynamics are limited to simple models, necessitating advanced simulation techniques.
• Accurate simulation of MRI signal dynamics is crucial for improving image acquisition and interpretation.

Purpose of the Study:

• To develop and validate a physics-informed neural network (PINN) framework for predicting MR signal dynamics.
• To enable simulations for arbitrary local Larmor field distributions.
• To provide a more flexible and accurate alternative to conventional MRI simulation methods.

Main Methods:

• Implemented a PINN framework to solve the Bloch-Torrey equation, incorporating diffusion and susceptibility terms.
• Represented spatiotemporal magnetization as a neural network output, using spatial coordinates and time as inputs.
• Utilized reflecting boundary conditions and initially isotropic magnetization within the complex Bloch-Torrey operator.

Main Results:

• Demonstrated the feasibility of predicting local and total magnetization using the PINN framework for intermediate oscillation frequencies.
• Identified the need for adapted neural networks (e.g., dual encoders, Fourier features) to learn higher-frequency solutions, as conventional MLPs failed.
• Observed spectral bias, partially mitigated by Fourier features, and noted the failure of gradient-enhanced PINNs to converge.

Conclusions:

• Simulating MRI signal dynamics is achievable with a PINN framework, offering high accuracy and enhanced flexibility.
• PINNs represent a promising advancement for complex MRI simulations where analytical solutions are intractable.
• Further research into network architectures and training strategies is warranted to fully address challenges like spectral bias.