COVID-19 X-ray Image Diagnostic with Deep Neural Networks

Abstract

The COVID-19 pandemic impacted all spheres of our society. The outbreak increased the pressure on public health systems, urging the scientific community to develop and evaluate methods to reliably diagnose patients. Driven by their effectiveness in medical imaging analysis, deep neural networks have been seen as a possible alternative to automatically diagnose COVID-19 patients from chest X-rays. Despite promising initial results, most analyses so far have been performed in small and under-represented datasets. Considering this, in this work, we evaluate state-of-the-art convolutional neural network architectures proposed in recent years by the deep learning field on images from COVIDx, a dataset consisting of 13,975 chest X-ray from COVID-19, pneumonia, and healthy patients. In our experiments, we investigate the effect of data pre-processing steps and class unbalancing for this task. Our best model, an ensemble of several networks, achieved an accuracy above 93% in the testing set, showing promising results in a challenging dataset.

@inproceedings{oliveira2020bsb,
    author       = "Gabriel Oliveira and Rafael Padilha and Andr{é} Dorte and Luis Cereda and Luiz Miyazaki and Maur{í}cio Lopes and Zanoni Dias",
    title        = "COVID-19 X-ray Image Diagnostic with Deep Neural Networks",
    booktitle    = "Brazilian Symposium on Bioinformatics (BSB)",
    year         = 2020
}