Neural Network Model for Breast Tissue Thickness Estimation

This paper explores the use of optimization methods to estimate thicknesses of subcutaneous tissues using a non-invasive near-field antenna with microwave low-power reflectometry from 2-13.5 GHz. An application is in breast cancer detection at microwave frequencies. The initial model assumes a planar tissue stack and is implemented with a Long Short-Term Memory (LSTM) neural network trained on full-wave electromagnetic simulation data. Tissue phantoms for skin, fat and muscle with 2 and 3 layers are used for experimental validation. The model developed for this paper predicts thicknesses of a stackup of Playdough, Rogers Duroid 6010 and a commercial skin phantom within 7%, 41%, and 24% respectively, with a neural network loss value of 0.29. This preliminary model is a proof of concept for a higher fidelity model that can additionally predict dispersive permittivities and conductivities of layered tissues.