Performance Evaluation of Deep Learning in Detection COVID-19 Based on Different Types of Datasets

Abstract

Background:

In this paper, we conducted a comparison of a number of deep learning networks in detecting Covid-19 disease based on X-ray and CT-scan images.  

Materials and Methods:

 An eight different deep learning network types (ResNet50, Mobilenetv2, Densnet201, SqueezeNet, Efficientb0, Googlenet, VGG19, and Alexnet25) are investigated to detect the COVI9-19 based on the X-ray  and CT-scan  images. A transfer-learning technique is adopted in the training phase for all networks with different datasets. The networks are trained and tested on 70% and 30% of the dataset, respectively.

Results:

The confusion matrix is calculated in the testing phase and the evaluation metrics including F1 score, accuracy, precision, specificity and sensitivity are calculated from the confusion matrix. The comparison results demonstrate that the classification accuracy when using X-ray dataset is better than that of using CT-scan datasets. Moreover, the Mobilenetv2 delivered the greatest results for different datasets in which the accuracy is greater than 99% for X-ray images and the accuracy is less than 80% for CT-scan images .

Conclusion:

Our conclusion is that the using of Mobilenetv2 network with the X-ray images is more suitable than the others.