Classification of COVID-19 Infection in Posteroanterior Chest X-rays With Common Deep Learning Architectures
Classification of COVID-19 Infection in Posteroanterior Chest X-rays With Common Deep Learning Architectures
The objective of this study is to assess three configurations of two convolutional deep neural network architectures for the classification of COVID-19 PCX images.
The December 2019 outbreak of COVID-19 has now evolved into a public health emergency of global concern. Given the rapid spread of infection, the rapid depletion of hospital resources due to high influxes of patients, and the current absence of specific therapeutic drugs and vaccines for treatment of COVID-19 infection, it is essential to detect onset of the disease at its early stages. Radiological examinations, the most common of which are posteroanterior chest X-ray (PCX) images, play an important role in the diagnosis of COVID-19. The objective of this study is to assess three configurations of two convolutional deep neural network architectures for the classification of COVID-19 PCX images. The primary experimental dataset consisted of 115 COVID-19 positive and 115 COVID-19 negative PCX images, the latter comprising roughly equally many pneumonia, emphysema, fibrosis, and healthy images (230 total images). Two common convolutional neural network architectures were used, VGG16 and DenseNet121, the former initially configured with off-the-shelf (OTS) parameters and the latter with either OTS or exclusively X-ray trained (XRT) parameters. The OTS parameters were derived from training on the ImageNet dataset, while the XRT parameters were obtained from training on the NIH chest X-ray dataset, ChestX-ray14. A final, densely connected layer was added to each model, the parameters of which were trained and validated on 87% of images from the experimental dataset, for the task of binary classification of images as COVID-19 positive or COVID-19 negative. Each model was tested on a hold-out set consisting of the other 13% of images. Performance metrics were calculated as the average over five random 80%-20% splits of the images into training and validation sets, respectively.
Inclusion Criteria:
Exclusion Criteria: