Covidx: Covid-19 Detection From Chest X-Rays Using Transfer Learning

  • Revanth Krishna, Joel James, Balaji Ravindaran, Abheendra S, Sourav Sudheer


Since, its outbreak in November 2019, the COVID-19 pandemic has affected almost all the countries in the world, and have had a deleterious impact on health and lives of people across the globe. A major reason for the widespread of COVID-19, is the slow pace of testing or lack of ability to detect the disease at the earliest. Therefore, a faster way to diagnose the patients, would be to detect the disease from X-ray images or Radiography. Several earlier research works have reflected certain anomalies in Chest X-rays of patients infected with COVID-19. Motivated by these research works, we study the possibility of detecting COVID-19 from Chest X-rays by the application pre-trained deep learning models. A customized dataset has been created for this study consisting of 6669 Chest X-rays adapted from datasets available for public or general use. The study aims to train fourteen significant convolutional neural network architectures, such as DenseNet-121, DenseNet-161, Googlenet, Inception-v3, MnasNet, MobileNet, ResNet18, ResNet50, ResNet101, ResNet152, SqueezeNet, ShuffleNet, VGG16, and VGG19on a training set consisting of 3469 radiographs, in order to detect COVID-19 from chest X-ray radiograms. These trained models are then validated on the test set consisting of the remaining 3200 images and in most of ourmodels, weobserved a sensitivity rate of over 90%, and obtaininga specificity rate over 94%. Apart from metric parameters like sensitivity and specificity, we observed the Receiver Operating Characteristic (ROC) curve along with AUC score, accuracy and loss for every epoch, and confusion matrix for each model. We also analysed the probability distribution for each prediction. Even though the observed performance of each model is satisfactory, a wider or broader research is necessary on a dataset with more number of COVID-19 images, for implementation of the model for real-time predictions. The dataset, all models along with its saved weights, and observations, are all added under the below mentioned Github repository for the benefit of other peers involved in research.