A Novel Study for Automatic Two-Class and Three-Class COVID-19 Severity Classification of CT Images using Eight Different CNNs and Pipeline Algorithm

Abstract

SARS-CoV-2 has caused a severe pandemic worldwide. This virus appeared at the end of 2019. This virus causes respiratory distress syndrome. Computed tomography (CT) imaging provides important radiological information in the diagnosis and clinical evaluation of pneumonia caused by bacteria or a virus. CT imaging is widely utilized in the identification and evaluation of COVID-19. It is an important requirement to establish diagnostic support systems using artificial intelligence methods to alleviate the workload of healthcare systems and radiologists due to the disease. In this context, an important study goal is to determine the clinical severity of the pneumonia caused by the disease. This is important for determining treatment procedures and the follow-up of a patient’s condition. In the study, automatic COVID-19 severity classification was performed using three-class (mild, moderate, and severe) and two-class (non-severe and severe). In the study, deep learning models were used for classification. Also, CT images were utilized as radiological images. A total of 483 COVID-19 CT-image slices, 267 mild, 156 moderate, and 60 severe, were used. These images and labels were used directly for the three classifications. In the two-class classification, the mild and moderate images were accepted as non-severe. A total of eight classifications were made with convolutional neural network (CNN) architectures. These architectures are MobileNetv2, ResNet101, Xception, Inceptionv3, GoogleNet, EfficientNetb0, DenseNet201, and DarkNet53. In the study, the results of the top four CNN architectures with the best performance were combined using a pipeline algorithm. In this way, it is seen that significant improvements have been achieved in the results of the study. Before using the pipeline algorithm for the three-class classification, the results of weighted recall-sensitivity (SNST), specificity (SPCF), accuracy (ACCR), F-1 score (F-1), area under the receiver operating characteristic curve (AUC), and overall ACCR were obtained: 0.7785, 0.8351, 0.8299, 0.7758, 0.9112 and 0.7785, respectively. After using the pipeline algorithm for the three-class classification, the results of these parameters were obtained: 0.8095, 0.8555, 0.8563, 0.8076, 0.9089, and 0.8095, respectively. Before using the pipeline algorithm for the two-class classification, the results of SNST, SPCF, ACCR, F-1, and AUC were obtained: 0.9740, 0.8500, 0.9482, 0.9703, and 0.9788, respectively. After using the pipeline algorithm for the two-class classification, the results of these parameters were obtained: 0.9811, 0.8333, 0.9627, 0.9788, and 0.9851, respectively.