| With the increasing concern of human health and the continuous improvement of deep learning method,the use of deep learning models for diagnostic analysis of med-ical images is also in full swing.The graph neural networks for their powerful feature extraction ability are widely used in the diagnostic research of medical images.The anal-ysis of medical images needs not only to give prediction results,but more importantly,the model should be interpretable.Furthermore,it is able to use multiple medical obser-vations for comprehensive analysis to further improve the diagnostic performance.This thesis proposes two models for diagnostic analysis of medical images based on graph neu-ral networks as follows:（1）Interpretable Dual-Graph Convolutional NetworkThis thesis proposes an interpretable dual-graph convolutional neural network,namely FSNet,to perform features and samples interpretation for the diagnose of medical images.Specifically,FSNet is a network with learning and explaining together.Firstly,the initial features of the nodes are aggregated by two sub-networks based on graph convolution.Then,FSNet use L2,1-norm and sparse mechanism to calculat and filter the weights of features and samples to obtain interpretable results.Finally,the feature weights and sam-ple weights are assigned to each node,and then feed the weighted nodes into the graph neural network for diagnostic classification.The experimental results on four binary clas-sification Alzheimer’s datasets show that,the proposed model,FSNet,outperforms other comparison methods in terms of both classification and interpretability because it weights the features and samples meanwhile.（2）Interpretable Multi-task Graph Convolutional NetworkThis thesis proposes an interpretable multi-task graph convolutional network,namely IMTGCN,to provide the classification of medical images and regression prediction of the clinical observables.IMTGCN also provides the interpretability required for medi-cal diagnosis.Specifically,IMTGCN begins by utilizing FSNet’s interpretable module to calculate the features and samples weight of each task.Subsequently,the attention mech-anism is employed to complete the sharing of feature information between tasks.Finally,IMTGCN designs the specific network of each task to perform the corresponding task.The experimental results show that on four datasets of Alzheimer’s disease,the proposed model,IMTGCN,takes advantage of the correlation information between multiple tasks to achieves better classification,regression of ADAS-Cog ans MMSE,and interpretability performance than analyzing each task individually.In summary,this thesis presents the diagnostic study of medical images using graph neural networks.This thesis proposed an interpretable dual-graph convolutional network and an interpretable multi-task graph convolutional Network to solve both classification and regression tasks in medical images analysis and provide the interpretability necessary for medical diagnosis.Experimental results on the Alzheimer’s disease dataset verified the validity of the two proposed models. |
