Estimation of Two Dimensional Electric Field Distribution Through Deep Learning: Preliminary Study

This study presents a deep learning approach capable of directly generating the distribution of a two-dimensional (2D) electric field (EF) distribution using electrical permittivity and conductivity maps. This approach aims to save time by bypassing the time-consuming traditional numerical calculation stage, particularly in studies involving a static system where computation difficulty is prevalent. Highlighting the critical importance of accurately determining EF distribution for various medical applications, our study presents preliminary results of deep learning models employing UNet and ResNet architectures. In the study, deep learning models process 2D cross-sections of breast models to directly predict EF distribution with high accuracy using electrical conductivity (σ) and permittivity (ϵ) values of numerical breast models. The proposed method also incorporates a masking-based loss function focusing the model’s learning efforts on significant regions where the desired electric fields exist. By applying this technique, the aim is to significantly reduce the time required to produce EF distributions from hours to seconds without compromising the accuracy of EF distribution predictions. Preliminary results indicate that the signal-to-noise ratio obtained with UNet is up to 3.45 dB higher compared to ResNet. The progress made in this study could pave the way for faster diagnosis and treatment planning without the need to wait for prolonged computation results.