An Approach of Dempster-Shafer Decision Fusion to Diagnose COVID-19 in Chest X-ray Imagery by Using Controlled Asymmetric Training of the two CNNs Ensemble

Abstract

This paper proposes a model of Dempster-Shafer decision fusion based on controlled training of the ensemble of two Convolutional Neural Networks (CNNs) by the asymmetry parameter k, defined as the ratio of the numbers of training data per class assigned to each CNN module. The proposed model is dedicated to COVID-19 diagnose in chest X-ray imagery. We have considered two CNN modules with identical architectures. First CNN module has been trained with 2837 COVID-19 labeled images and (2837/k) NON-COVID images. Second CNN module has been trained with (2837/k) COVID-19 labeled images and 2837 NON-COVID images. We have evaluated the influence of control parameter k on the diagnosis performances. As a result of Dempster-Shafer fusion, for k=2.1, one obtains a maximum Overall Accuracy (OA) of 95.18% The above performance is clearly better than the corresponding OA obtained by a single CNN (92.26 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> ) for the same k, and at the same time it is better than OA obtained by any single CNN module for any considered k. Moreover, one can remark, that by controlled training, for k=20, a CNN module can lead to an incredible low Missing Alarm Rate (MAR) of only 0.63%!