Investigation of Classifiers for Tumor Detection With an Experimental Time-Domain Breast Screening System

Abstract

Abstract-In this work we examine, for the first time, the use of classification algorithms for earlystage tumor detection with an experimental time-domain microwave breast screening system. The experimental system contains a 16-element antenna array, and testing is done on breast phantoms that mimic breast tissue dielectric properties. We obtain experimental data from multiple breast phantoms with two possible tumor locations. In this work, we investigate a method for detecting the tumors within the breast but without the usual complexity inherent to image-generation methods, and confirm its feasibility on experimental data. The proposed method uses machine learning techniques, namely Support Vector Machines (SVM) and Linear Discriminant Analysis (LDA), to determine whether the current breast being scanned is tumor-free. Our results show that both SVM and LDA methods have promise as algorithms supporting early breast cancer microwave screening.

Investigation of classification algorithms for a prototype microwave breast cancer monitor

“…As was shown in [5] and [8] the dielectric properties of the breast phantom vary over time. In [5], measurements over a period of 7 days demonstrated that the relative permittivity, ε r , at 3 GHz of the healthy breast phantom varied between 5.5 and 19. For "healthy" breast phantoms, the radome is completely filled with the adipose tissue.…”

“…Similar to [4], [5], we apply both LDA and SVM classifiers, individually, on selected data features that are extracted using Principle Component Analysis (PCA).…”

“…However, these papers have only focused on numerically obtained data. Most recently, we have demonstrated that these classification algorithms, both SVM and linear discriminant analysis (LDA), can be applied to data obtained from realistic breast phantoms using an experimental time-domain system [5].…”

“…We use a 16-element antenna array to perform measurements in the time-domain on dielectricallyrealistic breast phantoms. Unlike in [5], we now investigate the use of differential signals from successive scans of the breast phantoms over time. In an experimental system, variation of both the direct pulse signal (signals travelling directly between antennas) and the background environment signal degrades the classifier performance as these variations are not associated with changes within the breast structure.…”

Initial study for the investigation of breast tumour response with classification algorithms using a microwave radar prototype

“…The main author has successfully used classification algorithms, such as Linear Discriminant Analysis, Quadratic Discriminant Analysis, Support Vector Machines [1], Spiking Neural Networks [2] and Self-Organising Maps [3], to determine the size and shape of benign and malignant tumour models inside breast models. Other authors have also previously investigated the use of classifiers to aid in determining if a breast model represents a healthy breast or a breast with tumours, in simulation [4] and experimental scenarios [5].…”