Deep-Learning Optimized Reconfigurable Metasurface for Magnetic Resonance Imaging

Reconfigurable metamaterials (MTMs) allow for manifold applications in Magnetic Resonance Imaging (MRI). Most importantly, the full control in the spatio-temporal domain enables the realization of arbitrary sensitivity patterns during the transmit (Tx) as well as the receive (Rx) phase of an MRI scan via (re-)shaping the magnetic field distribution. The optimization of linking the reconfigurable N degrees of freedom to a desired field profile, however, is a non-trivial problem. In this work, a one-dimensional prototype (N = 14) is presented that is digitally controlled in combination with deep learning (DL)-driven optimization. Both, the forward and inverse problem are addressed. The forward problem, involving the mapping of capacitance values to a desired magnetic field distribution, is well-posed and can be effectively solved through a variety of methods. The inverse problem, i.e., deriving capacitance values from observed magnetic fields, is tackled by deep learning, specifically through a Multi-Layer Perceptrons (MLPs) approach. The feasibility of training a neural network with simulation data, achieving sufficient agreement between numerical and experimental results is demonstrated. Deep learning-driven optimization of MTMs in MRI applications holds a huge potential for many applications in the future and it represents a significant step towards bridging the gap between simulated and experimental results.