Detection of COVID-19 Based on Synthetic Chest X-Ray (CXR) Images Using Deep Convolutional Generative Adversarial Networks (DCGAN) and Transfer Learning

Abstract

The global COVID-19 pandemic has significantly impacted the health and lives of people worldwide, with high numbers of cases and fatalities. Rapid and accurate diagnosis is crucially important. Radiographic imaging, particularly chest radiography (CXR), has been considered for diagnosing suspected COVID-19 patients. CXR images offers quick imaging, affordability, and wide accessibility, making it pivotal for screening. However, the scarcity of CXR images remains due to the pandemic's recent emergence. To address this scarcity, this study harnesses the capabilities of Deep Convolutional Generative Adversarial Networks (DCGAN). DCGAN is a convolution-based GAN approach, has the potential to alleviate the scarcity of CXR data by generating authentic-looking synthetic images. This study combines synthetic CXR images with real CXR images to bolster model performance, resulting in an Extended Dataset. Extended Dataset comprises 7,345 images, with 34.63% being original CXR images and 65.37% being synthetic images produced by DCGAN. Expanded Dataset then utilized to train three pre-trained models: ResNet50, EfficientNetV1, and EfficientNetV2. The outcomes are remarkable, showcasing considerable enhancement in detection accuracy. Especially for the EfficientNetV1 model, it takes the lead with an impressive accuracy of 99.21% after merely ten epochs, achieved within a brief training period of 6.18 minutes. This surpasses the prior accuracy of 98.43% observed when used the Original Dataset (without synthetic CXR images). Overall, this research offers a solution to mitigate the scarcity of synthetic CXR images for COVID-19 detection. For future endeavors, refining the quality of synthetic images stands as an area for exploration, enhancing the overall efficacy of this approach.