Comparison of CNN-based Approaches for Detection of COVID-19 on Chest X-ray Images

Coronavirus disease 2019 (COVID-19), a highly contagious respiratory disease, has rapidly become a global pandemic. Chest X-ray imaging could serve an important role in early diagnosis of the disease. Deep learning methods have recently shown promise in disease detection tasks. The aim of this study was to develop a deep learning-based approach for detection of COVID-19 in chest X-ray images. Data were extracted from an opensource COVID-19 database developed by Cohen JP. The data consisted of X-ray images of patients with COVID-19, with other pneumonias or with no findings. The 205 images were randomly partitioned into training, validation and test datasets containing 143, 32, and 30 images, respectively, using a 70%/15%/15% split. The performance of several deep convolutional neural network (CNN)-based architectures, including VGG16, ResNet50, DenseNet121, and InceptionV3, were evaluated on the disease detection task. These networks were first pretrained on the ImageNet dataset consisting of natural images and then further fine-tuned on the task of detecting COVID-19 in chest X-ray images. The networks were then evaluated on the test set by assessing overall accuracy, area under receiver operating characteristic curve (AUROC), sensitivity and specificity. The performance of the networks trained from scratch without pretraining on ImageNet was also compared to the performance of the networks that were first pretrained on ImageNet and then fine-tuned on the detection task. DenseNet121 had the best performance on the test set with an overall accuracy of 90.0% (95% confidence interval (CI): 78.6%, 100%), an AUROC of 0.95, a sensitivity of 91.3% and a specificity of 85.7%. The pretrained DenseNet121 also significantly outperformed the DenseNet121 trained from scratch with a 30.0% improvement in overall accuracy. The proposed deep learning-based approach showed significant promise for detection of COVID-19 in chest X-ray images.