Classification of electromagnetic signals obtained from microwave scattering over healthy and tumorous breast models

Abstract

Current breast cancer screening using mammography has a number of drawbacks. As a result, the feasibility of a portable breast microwave sensing system is being evaluated. One approach would be to automatically classify the presence of breast cancer by analysing the microwave backscatter signals from the breast. This paper shows the results of work carried out to classify 2D breast models into those containing a tumour and those without. Two different datasets were used. The first was composed of signals generated using a frequency-domain simulation of microwave scattering from breast models made up of skin, fatty tissue and tumours but containing no fibro-glandular tissue. The second dataset was based on models that included fibro-glandular patches of different sizes to produce breasts with densities that range from 0 to 25% in terms of fibro-glandular content. Support Vector Machine (SVM) and K-nearest neighbours (KNN) machine learning algorithms were used to classify the data. The SVM classification for the presence of tumours in the homogeneous dataset was correct 94% of the time, while the KNN classification was only accurate 73% of the time. For the second data set the SVM classification was correct 62% of the time, while the KNN classification showed a small improvement. Further improvements on classification features are necessary to adapt the proposed approach to clinical practice. To this extent, the use of time-domain information arising from both breasts in the classification process will be considered in the future.