Evaluation of Computed Tomography Images by Convolutional Neural Network Architectures for COVID-19 Diagnosis: A Methodological Study

Objective: For the diagnosis of coronavirus disease2019 (COVID-19) using computed tomography (CT) images, disease classification from images must be performed with high accuracy. This study aims to evaluate the classification success on CT images using fine-tuned convolutional neural network (CNN) architectures applied to the new dataset with transfer learning method, which provides accurate and fast diagnosis of COVID-19, by comparing their classification success with various performance measures. Material and Methods: The classification of the disease was performed with CNN architectures Xception, InceptionResNetV2, InceptionV3, ResNet50, VGG16, VGG19, MobileNetV2, DenseNet121, DenseNet169 and DenseNet201. The performance of the architectures was utilized with accuracy, sensitivity, specificity, precision, F1 score and Area Under the Curve (AUC). In total, the data included 7,593 CT images of 466 COVID-19 patients and 6,893 CT images of 604 nonCOVID-19 patients. Results: The architecture that had the highest performance was InceptionResNetV2, with an accuracy of 98.00%, sensitivity of 98.53%, specificity of 97.53%, precision of 97.25%, F1 score of 97.88%, and AUC of 98.03%. ResNet50 showed the lowest performance with accuracy of 97.14%, sensitivity of 98.75%, specificity of 95.70%, precision of 95.32%, F1 score of 97.01% and AUC of 97.23%. Conclusion: All architectures examined were found to work with performance measures above 95% in COVID-19 diagnostics. As a result, it was found that the preprocessing, fine-tuning, and hyperparameter optimization of the architectures we used are generalizable for COVID-19 and chest CT images, and that each of the architectures works with good performance.