European Journal of Molecular & Clinical Medicine

Background:Reading radiographic images for Covid-19 identification by an expert radiologist requires significant time, therefore the development of an automated analysis system to assisting and saving time in diagnosing Covid-19 is important.
Objective: The purpose of this study is to implement the GoogleNet architecture with various epochs in hope achieving higher level of accuracy in Covid-19 detection.
Methods: We retrospectively used 813 images, i.e. 409 images indicating Covid-19 and 404 normal images. The deep TL model with GoogleNet architecture was implemented.The training was carried out several times to get the best acquisition value with a learning rate of 0.0001 for all levels. The network training was carried out with different epochs, i.e. 12, 18, and 24 epochs, and each epoch with 65 iterations.
Results: It was found that accuracy was determined by changes in the number of epochs. The classification accuracy was 96.9% in epoch 12, 98.2% in epoch 18, and 99.4% in epoch 24.
Conclusion: An increase in the number epochs increases the accuracy in the detection of Covid-19. In this study, the accuracy of the method reached 99.4%. These results are promising for the automation of Covid-19 detection from radiographic images.

The 2019 novel coronavirus (COVID-19), which originated from China, has been declared a pandemic by the World Health Organization (WHO) as it has surpassed over eighty three million cases worldwide, with nearly two million deaths. The unexpected exponential increase in positive cases and the limited number of ventilators, personal safety equipment and COVID-19 test kits, especially in Low to Middle Income Countries (LMIC), had put undue pressure on medical staff, first responders as well as the overall health care systems. The Real-Time Reverse Transcriptase-Polymerase Chain Reaction (RT-PCR) test is the decisive test for the diagnosis of COVID-19, but a significant percentage of positive tests return a false negative result. For patients in LMICs, the availability and affordability of routine Computerized Tomography (CT) scanning and chest X-rays is better compared to an RT-PCR test, especially in rural areas. Chest X-rays and CT scans can aid in the prognosis and management of COVID-19 positive patients, but are not recommended for diagnostic purposes. Using Deep Convolutional Neural Networks (CNN), three network based pre-trained models (AlexNet, GoogleNet and Resnet50) were used for the automatic classification of positive COVID-19 chest X-Rays and CT scans based on their severity into three classes- normal, mild/moderate, severe. This classification can aid health care workers in performing expeditious analysis of large numbers of thoracic CT scans and chest X-rays of COVID-19 positive patients, and aid in their prognosis and management. The images were obtained from public repositories, and were verified and classified by trained and highly experienced radiologist from Agha Khan University Hospital prior to simulations. The images were augmented and trained, and ResNet50 was concluded to achieve the highest accuracy. This research can be used for other lung infections, and can aid the authorities in the preparations of future pandemics.