Research On Multi-label Classification Based On Graph Convolution

Medical images have very complex visual features.In the traditional medical industry,human doctors classify pathological types in images based on personal experience.There is a lack of uniform standards and the classification accuracy cannot be guaranteed.Neural networks can surpass doctors in medical image classification tasks.The development of computer-aided diagnosis systems through deep learning algorithms is an important technological development direction.Multi-label and uncertainty label are two natural attributes of medical images.Research on how to conduct data mining on multi-label datasets with uncertain labels is an indispensable content in the extensive and in-depth application of neural networks in medical scenarios.This paper conducts multi-label classification research on chest radiology image data sets containing uncertain labels,and proposes a deep convolutional neural network-graph convolutional network structure model that is suitable for multi-label classification,and designs a new model for uncertain-label image convolutional information dissemination and aggregation method.The proposed model regards uncertain labels as samples that are difficult to judge from the image features,predicts its groundtruth category based on the label dependency,and establishes a classification that maintains a dynamic balance between image feature-driven prediction and label-dependent prediction system.The model achieves an average AUC of 0.899 on the Che Xpert dataset,which is 3% higher than the best of the baseline.The proposed model holds a more conservative attitude when predicting samples with similar characteristics to the uncertainty labels in the dataset,reflecting the confidence that it is more neutral compared with predicting other samples,and there is no need to evaluate the uncertainty of the model.The degree of certainty of the classification result is visually observed by the prediction confidence degree of the model.