Microwave Breast Imaging Based on Deep Learning

In order to address the problems of high computing cost and inability to real-time imaging in the traditional iterative methods of microwave breast imaging, a composite autoencoder network is proposed in this paper. It reconstructs the images from the scattered field arrays obtained by illuminating the breast dielectric constant images with antennas. The composite autoencoder network consists of two networks, the first being an autoencoder that mainly compresses high-resolution breast permittivity images into 256×3 vectors. The second neural network maps the scattered field arrays to compressed features 256×3, which are upsampled to high- resolution images. In this paper, a number of realistic breast phantoms are used to obtain a two-dimensional breast permittivity image dataset by slicing 3-D phantoms with a thickness of 2 mm. The proposed network can achieve real-time imaging compared to the improved traditional iterative method.