Deep Feature Learning for Microwave Mammography with Convolutional Autoencoders

This paper introduces a novel system relying on an unsupervised approach to categorize breast impressions obtained using low-frequency electromagnetic fields (LFEMF). We designed a miniaturized metasurface as the receiver, and an electrically small antenna as the transmitter, both operating at 200 MHz, to create two-dimensional impressions, referred to as LFEMF mammograms. Leveraging the integration of our designed metasurface and applied scanning techniques, we achieved images with resolutions less than
$\lambda/150$
. Our deep learning algorithm, trained on mammogram impressions in an unsupervised manner where no human labeling is involved, clusters images to their twelve subcategories with an adjusted rand score (ARS) of 91.18%.