Abstract:
The COVID-19 pandemic has had a devastating impact on global health, economies, and societies. Early detection
of COVID-19 is crucial to prevent transmission and reduce mortality, but conventional imaging techniques such as X-rays and
computed tomography (CT) scans have limitations in accessibility, cost, and sterilization. Therefore, in this study, we explore the
use of lung ultrasound (LUS) for COVID-19 diagnosis and evaluate the performance of various deep learning (DL) models such
as AlexNet, ResNet, DenseNet, Inception, VGG, Inception-ResNet, MobileNet, and Xception with transfer learning. In the
presented study, initially, we collected 455 COVID-19 and 226 non-COVID images, including bacterial pneumonia and healthy
subjects, from a POCOVID-Net database. However, DL networks demand more data to explore and develop a significant model,
so we employ data augmentation using geometric transformations. After that, we utilize the suggested deep transfer learning
architectures for identifying COVID-19 subjects from LUS images. Finally, we estimate the performance of these models by wellreceived
metrics such as sensitivity, specificity, precision, F1-score, the area under the curve (AUC), and accuracy. From the
experimental results, we observed that DenseNet-201 achieved 100% accuracy and outperformed other models. This indicates that
deep learning with transfer learning is a promising approach for COVID-19 identification from LUS data when data is scarce.
These findings could have important implications for improving the efficiency and accessibility of COVID-19 diagnosis,
particularly in resource-limited settings.
