Classification of COVID-19 in CT Scans Using Image Smoothing and Improved Deep Residual Network

ABSTRACT

The global spread of coronavirus disease has become a major threat to global public health. There are more than 137 million confirmed cases worldwide at the time of writing. The spread of COVID-19 has resulted in a huge medical load due to the numerous suspected examinations and community screening. Deep learning methods to automatically classify COVID-19 have become an effective assistive technology. However, the current researches on data quality and the use of CT data to diagnose COVID-19 with convolutional neural networks are poor. This study is based on CT scan data of COVID-19 patients, patients with other lung diseases, and healthy people. In this work, we find that data smoothing can improve the quality of CT images of COVID-19 and improve the accuracy of the model. Specifically, an interpolation smoothing method is proposed using the bilinear interpolation algorithm. Besides, we propose an improved ResNet structure to improve the model feature extraction and fusion by optimizing the structure of the input stem and downsampling parts. Compared with the baseline ResNet, the model improves the accuracy of the three-class classification by 3.8% to 93.83%. Our research has particular significance for research on the automatic diagnosis of COVID-19 infectious diseases. © 2021, Springer Nature Switzerland AG.