A deep learning‐based approach for diagnosing COVID‐19 on chest x‐ray images, and a test study with clinical experts. (17th May 2022)

• Record Type: Journal Article
• Authors: Sevli, Onur
• Is Part Of: Computational intelligence. Volume 38:Number 5(2022)
• Page Start: 1659
• Page End: 1683
• Publication Date: 2022-05-17
• Subjects: chest x‐ray analysis -- convolutional neural network -- COVID‐19 diagnosis -- pneumonia detection; Artificial intelligence -- Periodicals; Computational linguistics -- Periodicals; 006.3
• DOI: 10.1111/coin.12526
• Languages: English
• ISSNs: 0824-7935

Abstract

Abstract: Pneumonia is among the common symptoms of the virus that causes COVID‐19, which has turned into a worldwide pandemic. It is possible to diagnose pneumonia by examining chest radiographs. Chest x‐ray (CXR) is a fast, low‐cost, and practical method widely used in this field. The fact that different pathogens other than COVID‐19 also cause pneumonia and the radiographic images of all are similar make it difficult to detect the source of the disease. In this study, automatic detection of COVID‐19 cases over CXR images was tried to be performed using convolutional neural network (CNN), a deep learning technique. Classifications were carried out using six different architectures on the dataset consisting of 15, 153 images of three different types: healthy, COVID‐19, and other viral‐induced pneumonia. In the classifications performed with five different state‐of‐art models, ResNet18, GoogLeNet, AlexNet, VGG16, and DenseNet161, and a minimal CNN architecture specific to this study, the most successful result was obtained with the ResNet18 architecture as 99.25% accuracy. Although the minimal CNN model developed for this study has a simpler structure, it was observed that it has a success to compete with more complex models. The performances of the models used in this study were compared with similar studies in the literature and it was revealed that they generally achieved higher success. The model with the highest success was transformed into a test application, tested by 10 volunteer clinicians, and it was concluded that it provides 99.06% accuracy in practical use. This result reveals that the conducted study can play the role of a successful decision support system for experts.