Research On Multi-task Model For Pulmonary Nodule Classification Based On GCN/HGNN And 3D U-Net

Lung cancer has the highest incidence and mortality rates among all cancers,and a considerable part of lung cancer is manifested as pulmonary nodules in the early stage.Pulmonary nodules are a common and difficult to diagnose disease in chest medicine.Its diagnosis and treatment have always been a clinical challenge.Due to the shortage of medical resources in our country,and manual diagnosis is highly dependent on the work experience of doctors,the use of artificial intelligence technology for the diagnosis of pulmonary nodules has attracted the attention of many researchers.In tasks such as classification,detection and segmentation of pulmonary nodules,artificial intelligence technology has made substantial progress.However,most research work has not fully utilized the correlation between pulmonary nodule characteristics,resulting in poor interpretability of diagnostic results and limiting their practical application prospects.Firstly,how to effectively obtain the correlation between pulmonary nodule characteristics,and then use the correlation to establish a graph adjacency matrix and a hypergraph incidence matrix,and then construct a graph convolutional neural network and a hypergraph neural network,and finally construct two multi-task classification models.The main work results are as follows.(1)Carrying out how to effectively obtain the correlation between the characteristics of pulmonary nodulesAiming at the problem of how to effectively obtain the correlation between pulmonary nodule characteristics,in this paper,the LIDC-IDRI dataset is first analyzed,and nine pulmonary nodule characteristics marked by experts in the dataset are extracted.By analyzing the level labeling of each characteristic,it is concluded that the internal structural are not significant among different pulmonary nodules and need to be excluded.Finally,eight pulmonary nodule characteristics selected:subtlety,sphericity,margin,lobulation,spiculation,texure,malignancy and calcification for correlation study.The main work of correlation research is as follows:by calculating the correlation coefficient between pulmonary nodule characteristics to determine the degree of correlation between each pair of pulmonary nodule characteristics;by constructing multi-order transfer learning models,the optimal source task set for each pulmonary nodule characteristic scoring task is found,and finally the correlations between individual characteristics and characteristic combinations is analyzed.(2)Building a multi-task model for pulmonary nodule classification using the correlation between pulmonary nodule characteristicsAiming at the problem of how to apply the correlation between pulmonary nodule characteristics,in this paper,two multi-task models are proposed,namely multi-task model for pulmonary nodule classification based on graph convolutional neural network and 3D U-Net,and multi-task model for pulmonary nodule classification based on hypergraph neural network and 3D U-Net.i)Constructed multi-task model for pulmonary nodule classification based on graph convolutional neural network and 3D U-NetFirstly,a fully-connected neural network is used to obtain the initial label embeddings containing the correlation between pulmonary nodule characteristics.Afterwards,the graph adjacency matrix that can reflect the correlation of pulmonary nodule characteristics is constructed by three methods:data-driven,multi-order transfer learning and adaptive.A feature fusion module that fuses label embeddings and image features,and a cross-channel attention module that replaces U-Net skip connections are designed respectively.Finally,a multi-task model for pulmonary nodule classification based on graph convolutional neural network and 3D U-Net is proposed.The experimental results show that,compared with other existing models,the proposed multi-task model for pulmonary nodule classification based on graph convolutional neural network and 3D U-Net can make full use of the correlation between pulmonary nodule characteristics and the similarities between different tasks for pulmonary nodules.It can better complete the benign and malignant classification,characteristic scoring and three-dimensional segmentation tasks of pulmonary nodules,and the model outputs the diagnostic results of these tasks at the same time,providing comprehensive and credible diagnosis basis for doctors.ii)Constructed multi-task model for pulmonary nodule classification based on hypergraph neural network and 3D U-NetDue to some deficiencies in multi-task model for pulmonary nodule classification based on graph convolutional neural network and 3D U-Net,in clinical diagnosis,doctors often distinguish between benign and malignant pulmonary nodules based on the multiple characteristics of pulmonary nodules instead of a single feature.However,graph convolutional neural networks cannot model the complex relationship between pulmonary nodule characteristics.To this end,a hypergraph neural network is further proposed to extract the high-order correlation between pulmonary nodule characteristics,and three methods of Euclidean distance,ridge regression,and multi-order transfer learning are respectively used to construct a hypergraph incidence matrix that can reflect higher-order dependencies between pulmonary nodule characteristics.Finally,a multi-task model for pulmonary nodule classification based on hypergraph neural network and 3D U-Net was designed.Experimental results show that multi-task model of pulmonary nodule classification based on hypergraph neural network and 3D U-Net further improves the performance of multi-task model on the tasks of benign and malignant pulmonary nodules classification,characteristic scoring and three-dimensional segmentation.The effectiveness of the method of using highorder correlations between pulmonary nodule characteristics to improve the performance of classification models is verified.