Image-Based Prediction of Respiratory Diseases Including COVID-19 Using Convolutional Neural Networks

Abstract

Respiratory diseases such as COVID-19, Pneumonia, SARS, and Streptococcus have caused severe worldwide public health concerns. Specifically, COVID-19, as an emerging worldwide pandemic, imposed the most critical challenge to all scientists and researchers for prognosis, diagnosis, and treatment of COVID-19 infection. This study aims to predict the aforementioned 4 respiratory diseases and normal people with chest X-ray and CT scan images using convolutional neural networks. A total of 1,156 images has been collected from 3 published databases. The combined dataset was enriched by empowering augmentation techniques and visual filters such as rotation and lung segmentation. The noises for augmentation include Gaussian and Speckle noises with zero mean and variance of 0.05, 0.10, and 0.20, and Salt and Pepper noise with 50% and 75% ratio. The customized convolutional neural network reached a prediction accuracy of 94% in classifying the test images into the normal and 4 disease categories, and 92%, 93%, and 92% as average precision, recall, and F1-score over all categories, respectively.