Purpose
The aim of this study was to establish and evaluate a fully automatic deep learning system for the diagnosis of COVID-19 using thoracic computed tomography (CT).
Materials and methods
In this retrospective study, a novel hybrid model (MTU-COVNet) was developed to extract visual features from volumetric thoracic CT scans for the detection of COVID-19. The collected dataset consisted of 3210 CT scans from 953 patients. Of the total 3210 scans in the final dataset, 1327 (41%) were obtained from the COVID-19 group, 929 (29%) from the CAP group, and 954 (30%) from the Normal CT group. Diagnostic performance was assessed with the area under the receiver operating characteristic (ROC) curve, sensitivity, and specificity.
Results
The proposed approach with the optimized features from concatenated layers reached an overall accuracy of 97.7% for the CT-MTU dataset. The rest of the total performance metrics, such as; specificity, sensitivity, precision, F1 score, and Matthew Correlation Coefficient were 98.8%, 97.6%, 97.8%, 97.7%, and 96.5%, respectively. This model showed high diagnostic performance in detecting COVID-19 pneumonia (specificity: 98.0% and sensitivity: 98.2%) and CAP (specificity: 99.1% and sensitivity: 97.1%). The areas under the ROC curves for COVID-19 and CAP were 0.997 and 0.996, respectively.
Conclusion
A deep learning–based AI system built on the CT imaging can detect COVID-19 pneumonia with high diagnostic efficiency and distinguish it from CAP and normal CT. AI applications can have beneficial effects in the fight against COVID-19.