Deep Learning Models for Classification and Localization of COVID-19 Abnormalities on Chest Radiographs

COVID-19 pandemic is five times more deadly than flu and other disease. It causes serious morbidity and mortality across the world. Like other pneumonias, pulmonary infection with COVID-19 results in fluids in the lungs and inflammation. Equally, the disease looks very similar to other bacterial and viral pneumonias on chest radiographs;as such it is very difficult to be diagnosed. In this work, Convolutional Neural Network (CNN), Faster Region Based Convolutional Neural Network (Faster R-CNN) and Chest X-ray Network (CheXNet) deep learning algorithms were used to develop models for classification and localization of COVID-19 abnormalities on chest radiographs models for normal and opacity (typical, atypical, indeterminate) cases in order to help medical doctors, radiologists and other health workers to provide fast and confident diagnosis of the COVID-19. Hence, CheXNet based model has comparatively outperformed other models for being able to classify chest radiographs as negative for pneumonia or typical, indeterminate and atypical for COVID-19 pandemic with 97% accuracy and more so for its ability to correctly classify chest radiographs for typical, indeterminate and atypical COVID-19 pandemic cases the model has comparatively outperformed other models with 93% precision. However, for the ability to correctly classify the chest radiographs as negative for pneumonia, Faster R-CNN based model outperformed other models with 94% recall. © 2021 IEEE.

Deep Learning Models for Predicting COVID-19 Using Chest X-Ray Images

The outbreak of COVID-19 has cost the world a lot of lives and causes the shutdown of businesses which get most of the countries gone into economic recession. Despite the fact that some of the vaccines of the pandemic are now available, immediately after the first wave of the COVID-19 pandemic, the second wave of the pandemic has now started and causes a lot of lives and grounds a lot of businesses that have resumed. Therefore, in order to contain its further spread among humans, testing and screening of a large number of suspected COVID-19 cases for appropriate quarantine and treatment measures are of high priority to all governments around the world. However, most of the countries are facing inadequate and standard laboratories for testing a large number of suspected COVID-19 cases in their countries despite the fact that the virus is now endemic like other communicable diseases. Therefore, alternatives in non-medical diagnosis of COVID-19 techniques using artificial intelligence which include deep learning, data mining, machine learning, expert system, software agent, and other techniques are urgently needed in the cause of the diagnosis, containing and combatting the further spread of the pandemic. In this study, deep learning algorithms were used to develop models for predicting COVID-19 using chest x-ray images, and models were able to extract COVID-19 imagery features and provide clinical diagnosis ahead of the pathogenic test with a view to saving time, thereby complementing COVID-19 testing laboratories. ResNet50-based model was found to have the highest accuracy, sensitivity, and AUC score of 99%, 89%, and 96%, respectively. In contrast, EfficientNet B4-based model was found to have the highest specificity of 89%. Therefore, ResNet50-based model which has the highest sensitivity of 89% can be used for diagnosis of COVID-19 infection as well as an adjuvant tool in radiology department in hospitals. © 2022, Springer Nature Switzerland AG.