Physics-constrained deep-learning framework for MRI metasurfaces.

Area of Science:

• Metasurface design
• Artificial intelligence in scientific research
• Medical imaging hardware

Background:

• Metasurfaces offer solutions for signal-to-noise ratio (SNR) and radio frequency (RF) inhomogeneity in magnetic resonance imaging (MRI).
• Conventional metasurface design is hindered by ill-posed problems due to strong near-field coupling, leading to high computation costs and non-unique solutions.

Purpose of the Study:

• To develop a novel physics-constrained bidirectional deep learning framework for the inverse design of MRI metasurfaces.
• To address the challenges of computation cost and non-unique mappings in metasurface design.
• To enable automated and patient-adaptive MRI hardware design.

Main Methods:

• Implementation of a Spectral-Feature U-Net surrogate model to replace time-consuming physics simulations.
• Utilization of a tandem training strategy with physics-consistency constraints for a residual multi-layer perceptron (Res-MLP).
• Guiding the Res-MLP towards physically equivalent solutions to manage multi-valued mappings in inverse design.

Main Results:

• Simulations demonstrated an AI-designed metasurface achieving precise 63.8 MHz resonance and excellent field homogeneity.
• The designed metasurface maintained specific absorption rate (SAR) within established safety limits.
• The framework successfully addressed the ill-posed nature of metasurface design problems.

Conclusions:

• The proposed physics-constrained deep learning framework offers a robust solution for MRI metasurface design.
• This AI-driven approach facilitates automated and patient-adaptive MRI hardware development.
• Establishes a powerful 'AI for Science' toolchain for advanced medical imaging applications.