Research on the detection of COVID-19 via X-ray images based on triplet network under the condition of scarce data

With the outbreak of Omicron continuing to spread internationally, there is an urgent need to seek a prompt, effective and precise method to screen patients for infection. It was found in early studies that abnormal chest radiograph proves the vital evidence to confirm a patient infected with COVID-19 or not. Therefore, deep neural network (DNN) model trained with various and sufficient chest radiographs has become an alternative for medical screening and clinical diagnosis. The effectiveness of DNN is principally determined by the scale of chest X-ray images rather than other factors. Significantly, the more scarce training set becomes, the worse DNN will perform. In this paper, Triplet network derived from metric learning is applied to tackle the problem of scarce medical data. Using different Triplet loss functions to optimize the VGG-16 model pre-trained on ImageNet, our model enables to screen patients infected COVID-19 and distinguish normal, pneumonia and COVID-19 chest medical images rapidly, efficiently and accurately in spite of scarce data. With 91.5% accuracy in TriHard loss model and 92.4% accuracy Triplet-center loss model achieved, these all prove better result than base model.