Application And Research Of Graph Neural Network In Multi-label Chest Radiography

The efficient development of medicine is a particularly important link to ensure the life safety of citizens.Chest X-ray(chest film)diagnosis,as the most common imaging examination method,has the disadvantages of low efficiency and easy misdiagnosis,and deep learning technology can effectively complete the chest film detection task.For multi-label chest film classification,the main work of this paper includes:preprocessing CheXpert data set;according to the characteristics of CheXpert data set and the correlation between labels,MIA-GRGN for node classification is designed and DROP-DIR-RES network applied to multi-label chest film classification;The two networks are tested under the benchmark data set to verify the effect of the designed network.From the research status of chest disease detection and graph neural network at home and abroad,this paper analyzes the advantages and disadvantages of different methods,and summarizes how to better use GNN in multi-label chest film classification.Secondly,it expounds the pathological characteristics and label correlation of common disease types in multi-label chest film chexpert data set.In order to make chexpert data set suitable for multi-label network,through U-Zeros,U-Ones and U-Prob uncertainty label table The processing method is shown to process the training set;The word vector embedding set of disease category is obtained from glove corpus;The tag co-occurrence matrix is constructed according to the number of tag pairs.In order to more effectively capture the label dependency in multi-label images,Graph convolution network is improved,and a graph residual generation network MIA-GRGN based on multi information aggregation is proposed.By studying the defects of GCN,DIRGCN is designed to extract graph structure features,which retains part of graph structure information for each layer of features through initial residual connection,which effectively alleviates the over smoothing problem.The random graph generation method RGGM is designed to optimize the loss function of dirgcn in the form of generation framework.The optimized loss function promotes model learning and forms MIA-GRGN combined with graph structure,node characteristics and label joint distribution.The experimental results show that the classification accuracy of Citeseer dataset is improved by 1.3%in the semi-supervised node classification task of MIA-GRGN;In the fully-supervised node classification task,the classification accuracy of Cornell dataset is improved by 5.77%.In order to efficiently classify multi-label chest radiographs,a multi-label classification network DROP-DIR-RES is designed.Combining DIRGCN with residual network The combination of ResNet enables the interdependent label relationship to be better mapped to the network classifier;secondly,the DROP method is designed to refine the threshold division and flexibly generate the label incidence matrix through random loss,which improves the robustness of the network.The experimental results show that the map is increased to 93.8%on the benchmark dataset Pascal VOC 2007.The DROP-DIR-RES network is applied to the preprocessed network The AUC index reached 88.3%on the CheXpert data set.Compared with the existing models,the two networks proposed in this paper have certain advantages,and provide a research basis for the classification of multi-label chest radiographs.Figure[21]table[9]reference[81]...