In Procedia computer science
With the COVID-19 pandemic sweeping the globe, an increasing number of people are working on pandemic research, but there is less effort on predicting its severity. Diagnostic chest imaging is thought to be a quick and reliable way to identify the severity of COVID-19. We describe a deep learning method to automatically predict the severity score of patients by analyzing chest X-rays, with the goal of collaborating with doctors to create corresponding treatment measures for patients and can also be used to track disease change. Our model consists of a feature extraction phase and an outcome prediction phase. The feature extraction phase uses a DenseNet backbone network to extract 18 features related to lung diseases from CXRs; the outcome prediction phase, which employs the MLP regression model, selects several important features for prediction from the features extracted in the previous phase and demonstrates the effectiveness of our model by comparing it with several commonly used regression models. On a dataset of 2373 CXRs, our model predicts the geographic extent score with 1.02 MAE and the lung opacity score with 0.85 MAE.
Zheng Yongchang, Dong Hongwei
COVID-19, CXRs, deep learning method, estimate severity, feature extraction