Small-Sized Neural Network for Detecting COVID-19 from Chest X-rays

ABSTRACT

COVID-19 is a highly contagious infection that has now reached almost all countries in the world infecting over 33M and killing 1M people as of the time of writing. Therefore, it is essential to diagnose it early so that health care professionals can prevent the chance of a person spreading the virus. Because the disease often presents with respiratory symptoms, one method for detecting it is by radiology examination using chest radiography. Healthcare professionals examine the chest X-ray for abnormalities that are characteristics of those infected with COVID-19, which must be distinguished from other conditions with similar presentation such as pneumonia. This requires significant expertise, which may not be available in all parts of the world, so computer assisted diagnosis would be highly beneficial. We propose a deep neural network for extracting those abnormalities as features and classifying the infection. In this study, we examine the efficiency of small-sized deep neural network tailored for the detection of COVID-19 infection from chest X-ray (CXR) images. We designed a modified version of SqueezeNet and Capsule Network and show that even with a relatively small number of free parameters, it can achieve a competitive result while having modest hardware requirements. We use a modified version of fire modules to ensure better convergence. For our Capsule network, we used fire modules as two of its upper layers. To our knowledge, this is the first time that a fire module has been used in conjunction with capsules. Without any pretraining or transfer learning, our SqueezeNet was able to achieve an accuracy of 94.8 %, sensitivity of 88.0 %, and specificity of 98.4%. Additionally, our CapsNet achieved an accuracy of 93.8 %, sensitivity of 88.0 %, and specificity of 96.9 %. © 2020 IEEE.