The dramatically fast evolution of COVID-19 has triggered scientists around theworld to conceive fast effective responses in favor of a better understanding of thisnovel disease. Radiography examination is an available, accessible, and portable alternative screening method that enables rapid triaging. In this sense, deep neuralnets can provide recognition and classification tasks to aid radiologist experts and improve Chest X-ray (CXR) images analyses for faster medical treatment. Wang’s(2020) work deploys a customized network, a.k.a. COVID-Net, conceived from an automatic network search (AutoML) consisting of a deep-CNN composed of several PEPX modules. This work aims to reproduce Wang’s (2020) network in PyTorch without accounting for pre-training on the ImageNet dataset. Furthermore, the challenge of dealing with small data and the duty to provide reasonable accuracy for medical purposes drives the investigation of deep neural methods to go on various strands. We provide a performance comparison with ResNet50 as we account for possible extensions and improvements through generative models. Further additions, such as calibration, GradCAM algorithm, and an interface forsimplified usability were added on the top of results analysis.
Keywords:COVID-19, COVID-Net, ResNet, CNNs, GradCAM, calibration.