Research On Brain Networks Classification Based On Convolutional Neural Networks

Brain networks classification is a technology that can automatically determine whether a subject has neuropsychiatric diseases by mining and analyzing the features of brain network data,which provides an effective tool for understanding the pathogenesis and early diagnosis of brain diseases.Therefore,brain networks classification has a very important theoretical significance and application value.However,the high dimensionality and the small sample size of brain network data are great challenges to this study.In recent years,brain networks classification based on deep learning has attracted extensive attention.In particular,convolutional neural networks have gradually become a hot topic in brain networks classification due to their excellent feature learning ability for high-dimensional data.However,this kind of methods do not fully consider the topological characteristics of brain network data,nor do they make use of other existing data resources,such as clinical phenotype data,to improve the classification performance.To overcome the above shortcomings,this thesis focuses on the following two research works:(1)Considering the topological characteristics of brain networks,we present a brain networks classification method based on Convolutional Neural Network with Element-wise filters(CNN-EW).CNN-EW mainly includes two kinds of specially designed layers for brain network to extract topological characteristics of the brain network level by level,i.e.,the edge-to-edge layer with element-wise filters,the edgeto-node layer with element-wise filters.More specifically,the element-wise filter gives a unique weight to each edge of brain network,starting from the independence of edges in the brain network,which may reflect the topological structure information more realistically.Moreover,we combine two feature analysis principles for deep neural networks to further identify biomarkers associated with brain diseases.The experimental results on simulated datasets and the real dataset of ABIDE I show that CNN-EW models can not only improve the classification performance significantly,but also more precisely identify the abnormal brain regions compared to some state-ofthe-art methods.(2)To further improve the performance of the brain networks classification based on CNN-EW,we propose single-task learning and multi-task learning CNN-EW brain networks classification methods combining clinical phenotype features.Firstly,phenotype features that may be related to diseases are selected from dozens of clinical phenotype features(e.g.,gender,age and IQ score)to provide auxiliary information for brain networks classification.Then,the brain networks classification methods combining clinical phenotype features of single-task learning CNN-EW,multi-task learning CNN-EW and adaptive multi-task learning CNN-EW are studied and explored from shallow to deep.Among them,the proposed adaptive multi-task learning CNNEW method can automatically adjust the weight of each clinical phenotype auxiliary task to reduce the experimental cost and classification errors caused by the human operation.The experimental results on the real ABIDE I dataset show that the classification performance can be effectively improved by the single-task learning CNN-EW,multi-task learning CNN-EW and adaptive multi-task learning CNN-EW brain networks classification methods combining clinical phenotype features,which compared to CNN-EW without clinical phenotype features.In particular,the adaptive multi-task learning CNN-EW methods can accurately assign the weight for each clinical phenotype auxiliary task,thereby obtaining higher performance than the other two methods.The above researches in this thesis not only improve the performance of brain networks classification based on the convolutional neural network and promote the research and development of deep learning in brain networks classification,but also play a positive role in the accurate diagnosis and the discovery of biomarkers for brain diseases.