Classification of Chest CT Images for COVID-19 Detection Using an Ensemble of Twin CNNs with Decision Fusion

COVID-19 pandemic has revealed a challenging area of research regarding artificial intelligence for medical applications. This paper presents an approach of COVID-19 detection for chest Computed Tomography (CT) scan images using an ensemble of two twin Convolutional Neural Networks (CNNs) with the same architecture and training technique. For our investigation model, both CNN modules are intentionally trained with asymmetric volumes of training images belonging to the two considered categories (COVID-19 and NORMAL). Each of the CNN modules has been trained using N chest CT scan COVID-19 labeled images and (N/ß) NORMAL labeled images, where ß is the class asymmetry parameter defined as ratio of training volumes per class. The classification model uses a simple decision fusion technique, based on maximization of the NET functions in the output layer of each of the two twin CNNs. We have evaluated the influence of the asymmetry parameterß on decision performances. By selecting a small training image set, one obtains best COVID detection performances, corresponding to indices OA of 99.67%, MAR of 0.65% and FAR of 0%.