Breast Cancer Detection using Thermal Infrared Image Analysis based on Dempster-Shafer Decision Fusion of CNN Classifiers

Abstract

Thermography is a promising technology for breast cancer detection. We propose a new model to detect breast cancer based on thermography using an ensemble composed by two Convolutional Neural Networks (CNNs). The considered classifier applies Dempster-Shafer decision fusion. The two CNN modules have an identical architecture, but they use an asymmetric training procedure. The ratio between the number of cancer training thermograms and the normal training thermograms corresponding to first CNN module is denoted by β. The corresponding ratio for the second CNN module is chosen to be (1/β). The influence of the asymmetry training parameter β over the decision fusion classifier performances is evaluated. We have obtained the best result concerning overall accuracy (OA) of 98.02%, by choosing the parameter β of 1.2.