Diagnosis Of Autism Spectrum Disorder Via Multi-template Multi-center Learning

Autism spectrum disorder(ASD)is a group of widespread developmental disorders without effective diagnosis and treatment method.Modeling research on ASD helps doctors understand the pathological mechanism of the disease and make correct decisions,so as to effectively intervene in the disease and improve the living standards of patients.Resting state functional magnetic resonance imaging(rs-f MRI)is currently widely used for the research of the pathogenesis of ASD due to its non-invasive and relatively high spatial and temporal resolution.Neuroimaging-based machine learning methods provide an effective way for the study of ASD,but there are many limitations in the existing researches,such as the model does not consider the prior knowledge of the brain network,the complementarity between multi-template feature representations,and the heterogeneity issues between multi-center data and the relationship between non-imaging information and disease.To handle these problems,this paper proposes a multi-template multi-center learning method for ASD diagnosis,and verifies the effectiveness of the proposed method on the ABIDE dataset.The method can be divided into two research points:(1)In order to model the brain network reasonably and correctly express the relationship between brain regions,this study compared different data fitting methods and incorporated the prior knowledge of the brain while constructing functional connectivity network.In order to improve the diagnosis results of ASD,this research proposes a new multi-template multi-center ensemble classification framework based on network construction,which unifies sparse feature learning,manifold learning and classification in the same framework.Specifically,we use multi-task sparse feature learning framework to learn disease-related discriminant features from multiple perspectives,where feature learning in each template space is a task.In addition,to learn the local manifold structure of each template space,we designed a unique manifold regularization term.Here,our similarity matrix is obtained by adaptive process,which can reduce noise and learn manifold structure better.Meanwhile,to balance the contribution of each task,the weight corresponding to each template space is automatically assigned without additional parameters.Finally,to effectively alleviate the problem of data heterogeneity,we use multi-template multi-center ensemble strategy to obtain the final diagnosis result.This method has been experimentally verified for different imaging centers,and has obtained average accuracy of 77.63%,82.73%,78.11% and 89.13%.And a large number of experimental results show that our algorithm is quite attractive and advantageous.(2)In order to study the complementarity of multi-template features and combine population phenotypic information,we constructed a multi-task graph convolutional network model.First,to obtain effective and complementary features,we construct multiple functional connectivity networks for each subject.Second,to explore the relationship between non-imaging information and disease diagnosis,we use network features combining collection device information and population phenotypic information to construct a population phenotypic feature map.Finally,we use multi-task sparse feature learning framework to learn disease-related discriminant features from multiple perspectives,and regard feature learning in each template space as a task.Multiple graph convolutional network models are trained to learn feature graphs to obtain diagnostic results.The method achieved an average classification accuracy of 67.91% on a dataset consisting of four imaging centers,which indicates that the proposed method can be effectively applied to the ASD diagnosis and analysis with multi-template multi-center data.