Proved feasibility of rapid dielectric shimming prediction in MRI using Convolutional Neural Network on enriched dataset

Motivation: Inhomogeneities of the MRI transmit field causes image shading and hinders diagnosis. Implementing dielectric shimming, high-permittivity pads are used to recover signal in low-intensity areas, but full-wave calculation of the resulting fields is slow for real-time use at the scanner and requires massive computation. Goal(s): To prove the feasibility of CNN on rapid prediction of the transmit field with dielectric pads. Approach: U-Net architecture is trained on an enriched simulated data with diversified human models. Results: We obtain a reasonably high structural similarity with a low enough mean squared error across different human models, demonstrating the robustness and potential for a real-time implementation. Impact: This study demonstrates the feasibility for AI-assisted real-time calculation of dielectric shimming effect and electromagnetic fields, which could be applied to ultra-high field strengths, where significant inhomogeneity hinders proper evaluation, providing an insightful approach to improve image shading and diagnostics.