Deep Learning Transfer with AlexNet for chest X-ray COVID-19 recognition

Authors

Keywords:

alexnet, x-ray chest, convol, COVID-19, Deep Learning, recognition

Abstract

Deep Learning Transfer is an efficient method of recognition problems with a large amount of data. In response to the COVID-19 pandemic, we have applied learning transfer to a convolutional neural network known as AlexNet (AlexNet CNN) for chest X-ray recognition. We have fine-tunned AlexNet CNN for our specific problem. The first layer, which works with RGB images, is replaced for images in a single intensity (grayscale). The last layer is replaced with another fully connected layer to recognize only two classes (normal, pneumonia), instead of recognizing 1000 classes as the original AlexNet does. To train the network, a database of 5,216 chest X-ray images was used, among them are samples of healthy people and samples that present pneumonia caused by bacteria (streptococcus), acute respiratory system conditions (ARDS), severe acute respiratory syndrome (SARS) and viruses (COVID-19). For validation, we have used another set of chest X-ray sample data with the same characteristics. The results prove that Deep Learning with chest X-ray images can extract significant biomarkers related to COVID-19, since the obtained accuracy, sensitivity and specificity were 96.4%, 98.0%, and 91.7%, respectively. ROC analysis and confusion matrices are used to validate the results of the fine-tunned AlexNet network.

Published

2020-11-21

Issue

Section

Special Issue on Fighting against COVID-19
Bookmark and Share