Diagnosis of COVID-19 based on chest X-ray images using pre-trained deep convolutional neural networks

The novel coronavirus (COVID-19) that emerged and transmitted from China (Wuhan City) had a staggering effect on public health and the world economy. The early diagnosis of COVID-19 has become more important for its treatment and for controlling its spread due to its highly transmissible nature. In addition, the restricted supply of test kits calls for an alternative system for diagnosis. Since radiological images of chest of patients with COVID-19 show abnormalities, it is possible to diagnose COVID-19 utilizing chest X-ray images. Therefore, by applying deep convolution neural network (CNN), we have presented a diagnosis of COVID-19 based on chest X-ray images in this paper. For the diagnosis of COVID-19, an exhaustive comparative performance analysis of 16 state-of-the-art models is presented. Moreover, each model is trained with three approaches: transfer learning, fine tuning and scratch learning. The experiments were conducted on the dataset that comprises of 127 images of COVID-19, 500 images of Pneumonia and 500 images of normal cases. We have performed the experiments in two scenarios: binary classification (COVID-19 vs. Normal) and multiclass classification (COVID-19 vs. Pneumonia vs. Normal). Further, we have applied cost-sensitive learning technique to handle the class imbalance issue. In this study, InceptionResNetV2 model with fine-tuning approach achieved highest classification accuracy of 99.20% in binary classification and Xception model achieved classification accuracy of 89.33% in multiclass classification among all considered models. To validate our approach, we have presented the performance of our model on three other datasets and achieved adequate classification accuracy. Hence, the promising results demonstrate that the fine-tuning of deep CNN models is an effective way for diagnosis of COVID-19 and therefore, it can be deployed in diagnostic centers to assist radiologist after its validation with more prominent datasets.