Resources

The 2019 coronavirus disease (COVID-19) is responsible for millions of casualties. While the outbreak impacts not only individuals but also our nation's healthcare and economic systems, efforts to combat it are hampered by testing that is insufficient and expensive. Chest X-rays are common radiographic imaging tests used to identify respiratory diseases including COVID-19 and pneumonia. However, getting the strong performance required in the majority of contemporary medical applications often necessitates a huge number of samples. Convolutional neural networks (CNN) have shown the potential to be good at classifying X-rays for helping ailment diagnosis. In this paper, the researchers demonstrate the effectiveness of the Cycle-Generative Adversarial Network (Cycle - GAN), commonly used for neural style transfer, to generate a set of samples from X-ray images of healthy individuals (widely available) so that they appear to belong to an X-ray image set of patients with diabetes. Neural style transfer refers to techniques able to adopt key characteristics from one set of samples to another to perform image transformation. The researchers used transfer learning methods to remodel some of the most effective CNNs in order to assess this strategy. Using six different alternatives in which Covid-synthetic images are gradually added to the training collection of images, the macro f1-score was evaluated for each choice. The paper concludes that this method is effective at enhancing the performance of COVID-19 classifiers for a number of popular convolutional neural networks when combined with conventional transfer learning approaches. This work made a contribution by demonstrating the effectiveness of neural style transfer as a method for addressing the lack of labeled examples in imbalanced datasets when modeling picture classification issues.