Research On Brain Functional Connections Classification Based On Deep Learning

Brain functional connection classification is a technique to determine whether a subject suffers from a mental disorder through feature extraction and analysis of human brain functional connection data.It provides a new idea for finding the cause of mental disorders,so has important application value and practical significance.However,the high dimensionality and the small sample size of brain functional connection data are great challenges to this study.In recent years,the classification method of brain functional connections based on deep learning has become a research hotspot.Stacked autoencoders(SAE)has a simple structure and can extract deep features from the data layer by layer.It is one of the most commonly used deep learning methods in brain function connections classification tasks.However,at present,the classification accuracy of such methods needs to be further improved,and the multiview information provided by different types of brain region templates is not fully utilized,so there is still much room for improvement in classification performance.Therefore,this thesis focuses on the following two research work:(1)In order to improve the ability of the deep model to extract discriminative features,and to comprehensively use the different levels of features extracted by the deep model during the classification stage,we propose a brain functional connections classification method based on prototype learning and deep feature fusion.Firstly,we use stacked autoencoders to extract lower-to-higher deep features from brain functional connections.Then the prototype learning is used to extract the distance feature of the sample category from each hidden layer of the stacked autoencoders.Finally,the deep feature fusion strategy is adopted to fuse these distance features and the fused feature is applied for the brain functional connections classification.The experimental results on the ABIDE dataset show that compared with other methods,the proposed method not only has a higher classification accuracy,but also can locate the brain areas related to diseases more precisely.(2)In order to further improve the classification performance of brain functional connections,we propose a brain functional connections classification method based on multiview learning.Firstly,we use different types of pre-defined templates to identify regions of interest to build brain functional connections in different views.Secondly,the proposed multiview feature selection method is used to select the most discriminative features from each view with the assistance of other views.Thirdly,using SAEs to extract deep features of brain functional connections from various views.Finally,the proposed multiview fusion layer is used to fuse the deep features from multiple views and make full use complementary information from each view.Experimental results on ABIDE I data set show that the brain functional connections classification method based on multiview learning can further improve the classification performance of brain function connections.The above researches in this thesis not only improve the classification performance of brain functional connections and promote the research and development of deep learning in brain functional connections classification,but also play a positive role in the accurate diagnosis and the discovery of biomarkers for brain diseases.