COVID-19 Image Classification Method Based on Model Fusion

Abstract

In 2020, the world witnessed a new and severe global health crisis: the outbreak of Covid-19 and the number of positive cases and deaths around the world rose at a frightening rate throughout 2021. Given its highly contagious, convenient and efficient detection means are significant. At present, RT-PCR testing is the common diagnostic method for COVID-19 cases, but the process is time-consuming and inefficient. The recent COVID-19 radiology literature has focused on CT imaging because of its higher sensitivity, but it leads to high costs compared to X-ray imaging. Nowadays, many AI applications are focused on quantification and identification of infections to fully automate diagnoses to assist medical experts. Therefore, we compared seven classic network models including ResNet50, VGG16, VGG19, InceptionV3, InceptionResNetV2, MobileNetV1, DenseNet169 by the diversity measure. DenseNet169 performed pretty well with an accuracy of 97.5% on the training set and 96.58% on the test set. After comparing the results of different model fusion methods, stacking these models by four folds and selecting the tree classifier as second layer models outweighed other methods which reach 100% on the test set, which is helpful in the diagnosis of COVID19.