Comparative study for automated coronavirus detection in CT images with transfer learning

Abstract

Purpose: To design a computer-aided diagnosis system with transfer learning methods to serve as decision support system for automated coronavirus detection in CT images.
Methods: Four pre-trained deep convolutional neural networks (ResNet-18, SqueezeNet, ShuffleNet, MobileNet-v2) have been investigated to diagnose coronavirus with CT scans. To evaluate the pre-trained deep convolutional neural network, we used the COVID-CT dataset, which contains 349 CT images of COVID-19 from 216 patients, and 397 CT images obtained from non-COVID-19 subjects.
Results: Considering binary classification performance results, it has been seen that the pre-trained ResNet-18 model provides the highest classification performance (97.0470 ± 5.5466 accuracy, 98.7342 ± 2.1925 sensitivity, 95.1429 ± 9.3460 specificity, and 0.9737± 0.0489 F1-score) among other three used models.
Conclusion: ResNet-18 model can be employed as a supportive decision-making system to assist radiologists at clinics and hospitals to screen patients swiftly.