Background Transmission electron microscope (TEM) is an important method to detect virus. TEM detection often relies on manual observation by experts, and the operation steps are cumbersome. Moreover, existing machine learning methods are easily affected by background and noise, resulting in poor virus detection methods, low efficiencyand time consuming.
Objective In order to improve the efficiency of TEM virus detection, an Enhanced Graph Convolution Network (EGCN) is proposed to solve the problem of automatic identification of virus morphology in TEM images.
Methods In this model, Convolutional Neural Network (CNN) was used to extract the local feature information between pixels, and GCN was used for graph feature learning combined with the nearest neighbor relationship between sample features. In the model optimization, the group super classification loss and classification cross entropy loss were introduced to improve the feature extraction ability of the model for virus category information, and further improve the robustness of TEM virus image features compared with convolution neural network.
Results Experiments were carried out on 15 types of TEM virus image datasets through different methods, and EGCN achieved a top-1 error rate of 3.40%, a top-2 error rate of 1.88%, a precision of 96.65%, and a recall rate of 96.60%. A series of comparative experiments demonstrated that the EGCN can effectively solve the influence of background and noise in TEM virus recognition.
Conclusion By using the enhanced graph convolutional neural network, the task of virus morphology recognition can be effectively solved, providing important reference value for virus diagnosis.
