Multi Label Classification And Unknown Disease Detection Of Medical Image Based On Knowledge Graph

With the excellent results of deep learning methods in data analysis,the interdisciplinary field of medical image processing has attracted a lot of attention from researchers.At present,most image disease classification is still based on supervised learning,so it will face problems such as excessive dependence on a large number of labeled data and the recognition ability of the model will be limited when a new disease appears.At the same time,for the traditional image processing model,it mainly focuses on the fine-grained classification of imaging features,and lacks the internal relationship analysis of the entity objects in the image.With the continuous promotion of knowledge graph,it describes the facts of the objective world with its powerful representational ability,linking many potential knowledge to the actual task requirements.Therefore,this thesis constructs a medical image knowledge graph based on knowledge graph theory and technology.The lesion entity association is used as the base knowledge to simulate the global perception of image information by human brain.The machine is equipped with the ability of reasoning between multiple associated information as well as the cognition and prediction of the unknown domain to realize the model's deep cognition of image content.Based on the above problems,this thesis explores a method similar to the process and application of knowledge construction by human brain based on the knowledge graph theory and thoughts.And this knowledge architecture is applied to the analysis of multi-label discrimination and unknown domain.The work of this thesis is as follows.1?In this thesis,we use the intrinsic connection between entity objects in medical images to construct medical image knowledge graphs,and give the representation of entities in the knowledge graphs and the calculation of node features.The image features are weighted and averaged with the node features,and the corresponding weights among the nodes are dynamically adjusted.The correlation problem between multiple labels is effectively solved.2? In the unknown disease detection task,this thesis uses the structural characteristics of the knowledge graph and the knowledge features to participate in the inference calculation,which motivates the machine to simulate the human imagination of the unknown,so as to effectively solve the knowledge migration problem between known and unknown diseases.In the thesis,the knowledge graph is constructed by extracting the semantic association information between images and text among lesion entities,and assigning weights to the graph edges by calculating the association values.The combination of image features and map node information is used to complete the inference calculation of the map;multi-label classification of images and detection of unknown diseases are achieved.The experimental results show that the proposed model in this thesis can obtain higher accuracy,and both are improved compared with the traditional model.