Deep Dense Model for Classification of Covid-19 in X-ray Images

Novel Coronavirus, SARS-CoV-2, can be fatal for humans and animals. The ease of its propagation, with its extraordinary ability to cause disease and even death in humans, makes it a hazard to humanity. Chest X-ray is the most popular but difficult to apprehend radiographic analysis for immediate diagnosis of COVID-19. It yields significant anatomical and physiological information. However, extracting the appropriate information from it is seldom difficult, even for radiologists. Deep CNN architectures can assist in reliable, swift, and accurate results. We propose a deep dense model fine-tuned from scratch and statistically analyzed its results using paired two-sided t-test with state-of-the-art deep learning models, namely, SqueezeNet, AlexNet, DenseNet201, and MobileNetV2. Current datasets are limited and generally unbalanced. However, we devised a larger and well-balanced dataset for training the model. Moreover, as the dataset is still not significant, thus data augmentation and finetuning approaches are employed to evade overfitting and generate a better-generalized model. Our deep dense model produces better performance from analyzed deep learning models to generate Specificity, Recall, FScore, and Accuracy of 97.33%, 92.01%, 92.00%, and 96.01%, when trained on a significantly larger and balanced dataset, while employing 5- Folds cross-validation. The statistical analysis also shows that our model is better than its competing methods. Our deep model can help radiologists in the correct identification of COVID-19 in X-rays. That can contribute toward speedy and reliable diagnosis, thereby saving precious lives and minimizing the socio-economic burden on society.