Deep Learning-Based CT Image Classification Model with Application to COVID-19 Diagnosis

In this paper, aiming at the computed tomography (CT) images, a deep learning (DL)-based classification model is proposed, which is applied to realize the accurate diagnosis of the novel coronavirus disease COVID-19. In particular, the ResNeXt is adopted as the backbone, where the grouped convolution facilitates the extraction of discriminative features, and the residual structure can effectively address the gradient vanishing problem. Moreover, by improving and introducing the residual block in ResNet-D, the network can flexibly pay attention to features with different importance. Named as ResNeXt-vd, the proposed model is proven competitive in COVID-19 diagnosis task, which outperforms some other classic DL models by yielding the highest accuracy of 0.975, thereby showing the potential application values in providing reliable diagnostic references for the clinicians.