End-to-End Residual Stage Network for COVID-19 Detection

Currently, the COVID-19 virus is in a low-level epidemic state but still spreading in society. To timely cut off the transmission chain, early screening for COVID-19 is crucial. However, adding extra network modules during training may lead to module stacking errors and increased complexity. In this paper, we propose an optimization method based on residual stage networks (ResStage) and polarized self-attention mechanism to improve the detection ability of the model. This method provides better paths for information propagation across different layers of the network and extraction more refined and high-quality features. Results show that the fusion model based on ResStage and polarized self-attention can improve the detection accuracy of COVID-19 without increasing excessive parameters, achieving an AUC score of 86.53%. Moreover, this paper also subdivides the training tasks, including COVID-19 positive patients, COVID-19 positive patients with coughing, healthy individuals with coughing, asthma patients with coughing, and healthy asymptomatic individuals. These different tasks help to evaluate the performance of model and provide more references for practical applications.