Frequency Sinogram Based Cancer Recognition via Convolutional Neural Network for Microwave Breast Cancer Diagnosis

Abstract

In this paper, we present a frequency sinogram data-based cancer recognition model employing a convolutional neural network (CNN) scheme for microwave-based breast cancer screening. As most cancer recognition schemes are based on radar image exploitation, these approaches have difficulty in discriminating a fibroglandular tissue from a malignant tumor, particularly in dense breasts, due to low contrast between tumor and fibro-glandular, in terms of dielectric properties. Thus, we introduce a straightforward recognition scheme involving the exploitation of a potential characteristic of backscattered frequency sinogram data. Furthermore, data augmentation schemes along an array rotation angle are introduced with a Fourier-based upsampling scheme to ensure high-accuracy recognition. The two-dimensional finite-difference time domain method using a realistic numerical phantom validates the effectiveness of our proposed approach.