Detection of Coronavirus in Electron Microscope Imagery Using Convolutional Neural Networks

Abstract

Object recognition has always been one of the most crucial areas of research in modern times when it comes to industries like healthcare. Not only it is useful to automate certain kinds of jobs to improve efficiency, but also it also takes away the scope of human errors that are ever-present within such fields. A multitude of factors like differences in skills, experience, and more critically, basic human errors contributes to inadequacy and thereby inefficiency. These are not merely hypotheticals, but legitimate concerns raised in research papers and by professionals within their respective sectors as laid out in (Hopfer et al. in Histopathology 78:358–370, 2021 [1]). Yet, no significant work has been done so far in combining object detection with the identification of the coronavirus. This is due to some significant obstacles. Recognizing these viruses is a tedious task because substantial datasets of electron microscope images are not available. In this paper, we propose a method of detecting the virus through electron microscope imagery using the You Only Look Once (YOLO) convolutional neural network (CNN) model. We trained our model on an amalgamation of two datasets. The first is an open-access dataset named ‘electron microscopy of SARS-CoV particles Dataset’ by Laue, Michael; Kauter, Anne; Hoffmann, Tobias; Michel, Janine; Nitsche, Andreas. The second is ‘The SARS-CoV-2 Dataset’ by Northeastern University, made available to us on request by Prof. Jiawei Zhang. Overall, we achieved a mean average precision of 86.5%. Such a system is beneficial as it assists in a large amount of research work, both in terms of increasing efficiency and minimizing errors. However, we do believe there exists a lot of scope to build upon the work with a larger dataset in order to achieve higher accuracy.