Detection of COVID-19 Using Deep Convolutional Neural Network on Chest X-Ray (CXR) Images

Our study aims to investigate the best performing Convolutional Neural Networks (CNN) suitable for COVID-19 detection on Chest X-Ray (CXR) images. We applied five state-of-art CNN models in this study: DarkNet-19, ResNet-101, SqueezeNet, VGG-16, and VGG-19. These CNN models were pre-trained with natural images for classification. Therefore, we used transfer learning to modify the fully connected layer and output layer for a binary classification between COVID-19 and normal lungs. The models were trained using our combined dataset of CXR images obtained from the public domain, COVIDx, and private domain, University of Malaya (UM). The CXR images were pre-processed with reflection along the horizontal and vertical axis before being fed into the CNN models. Then another combined dataset from both COVIDx and UM was used to test the performance of the models. The numbers of correctly and wrongly predicted classes were tallied and represented with a confusion matrix. Then, the specificity, sensitivity, precision, F1-score, and accuracy were measured to evaluate the performance of each model. Our study demonstrated an accuracy above 90% for all five models. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the significant activation regions that contributed to the model's decision. We have also applied the COVID-Net-CXR-Large model to our combined dataset for testing to evaluate its performance in multiclass classification. The current CNN models require further improvement and modification before they can be applied clinically as a secondary tool for the diagnosis of COVID-19 cases.