Development of an artificial intelligence method to detect COVID-19 pneumonia in computed tomography images

INTRODUCTION: This study was aimed to construct an artificial intelligence system to detect COVID-19 pneu-monia on computed tomography (CT) images and to test its diagnostic performance.
METHODS: Data was acquired between March 18-April 17, 2020. CT data of 269 RT-PCR-proven patients were extracted and 173 studies (122 for training, 51 testing) were finally used. Most typical lesions of COVID-19 pneumonia were labeled by two radiologists using a custom tool to gen-erate multiplanar ground-truth masks. Using a patch size of 128x128 pixels, 18.255 axial, 71.458 coronal, and 72.721 sagittal patches were generated to train the datasets with the U-Net network. Lesions were extracted in orthogonal planes and filtered by lung segmentation. Sagittal and coronal predicted masks were reconverted to the axial plane and were merged into the inter-sect-ed axial mask using a voting scheme.
RESULTS: Based on axial predicted masks, the sensitivity and specificity of the model were found as 91.4% and 99.9%, respectively. The total number of positive predictions has increased by 3.9% by the use of intersected predicted masks whereas the total number of negative predictions has only slightly decreased by 0.01%. These changes have resulted in 91.5% sensitivity, 99.9% specificity, and 99.9% accuracy.
DISCUSSION AND CONCLUSION: This study has shown the reliability of the U-Net architecture in diagnosing typical pulmonary lesions of COVID-19 in CT images. It also showed the slightly favorable effect of the intersec-tion method to increase the model’s performance. Based on the performance level presented, the model may be used in the rapid and accurate detection and characterization of the typical COVID-19 pneumonia to assist radiologists.