A deep learning model for the estimation of RF field trained from an analytical solution

In the article we describe a neural network able to estimate the electric field from SNR and B1+ maps. The network was trained using the images obtained from a 2D analytical solution of an infinitely long circular cylinder excited by a conductive wire. The predicted images are in good agreement with the targets with a maximum mean square error of 0.28%. The performance of the network was also tested on a human body model with an error of 13.96%. These preliminary results are very promising and open the possibility to use the network on numerically simulated and experimentally acquired images.