Graph Convolutional Neural Network Based ADHD Classification

Attention Deficit and Hyperactivity Disorder(ADHD)is the disease with the symptoms of inattention,hyperactivity,or impulsivity.It is one of the most common neurodevelopmental disorders and mental disorders in children and adolescents.ADHD tends to make negative influence on their study and life.With the development of medical imaging technology,some imaging approaches are available to the researches and diagnosis for brain diseases.Because of high spatial and temporal resolution,Resting State Functional Magnetic Resonance Imaging(rs-f MRI)technology has been widely used in the research of ADHD.Among these work,it is a popular method to modeling the brain network with graph theory and be recognized by machine learning algorithms in ADHD classification research.However,the classification accuracy of ADHD still needs to be improved.Single measurement was used for functional connection in most work,it would cause the false connection and the loss of information.In order to solve the problems in the existing algorithms,we divide our work in two parts:(1)We propose A Cross-layer node feature extraction based Multi-scale Graph Convolution Neural Network(CM?GCN)which is a deep learning algorithm,to classify ADHD and normal individuals.The multi-scale graph convolution algorithm enriches the the feature information of the connection edge and the cross-layer node feature extraction algorithm combines the information of the low layer and the high layer to obtain a more accurate connection relationship between the nodes,so that the feature of nodes can be extracted accurately.We perform comparative experiment in the ADHD-200 dataset and the above two algorithms and the CM?GCN model were proved effective.(2)To solve the limitation of the single functional connection measurement on the model classification performance in the existing method,We propose a multi-branch graph convolution neural network model.First,construct five types of functional connection matrices(FCM)based on the subject's rs-f MRI time series samples,which would solve redundant information problem.Finally the classification result is obtained in the output.We test the model in the NYU and PEK databases from the ADHD-200 database and it has obtained 91% and 89% accuracy rates,92% and 86% F1 scores respectively,which is superior to existing related work and proves the effectiveness of the algorithm.In summary,our study optimizes the graph convolutional neural network model,and has a better solution to the instability of the single modeling method of the brain network and the redundancy problem of multi-feature information.