A Bayesian nonparametric approach to tumor detection using UWB imaging

Abstract

We develop a tumor detection and discrimination algorithm for Ultra-Wideband (UWB) microwave imaging of breast cancer based on a Bayesian nonparametric approach. We model the UWB backscattered signal as a mixture of distinct scatterer contributions, and use a Dirichlet Process mixture model (DPMM) to describe the amplitudes and delays of the backscattered returns. Because of the unbounded complexity afforded by the DPMM, model under-fitting is avoided and parameters like the clutter covariance matrix in other commonly used approaches, need not be estimated. The DPMM allows us to perform discrimination when there are multiple tumor and clutter sources that present as extended radar targets. After performing discrimination, we distinguish the tumor sources from other clutter sources using a generalized likelihood ratio test (GLRT). We perform experiments on a breast phantom with realistic dielectric contrast ratios, and compare the performance of our algorithm with a direct GLRT approach. Our numerical results show performance improvement in terms of tumor detection probability and Signal to Interference and Noise Ratio (SINR) gain of approximately 2.2 dB at a probability of detection of 0.9 over the GLRT method.