Research On Identification Of Multi-site Autism Spectrum Disorder Based On Domain Adaptation

Autism spectrum disorder(ASD)is a neurodevelopmental disorder characterized by social deficits and repetitive behaviors.Traditional diagnostic tools lack objectivity.In order to promote objective ASD-assisted diagnosis,the Autism brain imaging data exchange(ABIDE)database proposes to conduct ASD prediction research based on MRI data from multiple imaging sites.However,due to the inconsistent collection standards of multiple imaging sites,heterogeneity among the multi-site data has been introduced,and the research on ASD identification based on the multi-site data has not achieved good performance.The domain adaptation method can reduce the data offset between domains by aligning the feature distribution of data from different sources,which helps to improve the performance of multi-site ASD prediction.In order to reduce the heterogeneity of multi-site data,this paper takes one of the centers as the target domain and the other centers as the source domain,and proposes three multi-site ASD recognition methods based on domain adaptation.The specific work is as follows:(1)The traditional multi-site ASD prediction methods usually use the common subspace projection method to align the data distribution of multiple sites,and do not consider the existence of the common subspace of multiple centers.To this end,this thesis proposes an ASD identification method based on a two-stage multi-site domain adaptation.First,the manifold feature transform is used to align the data distribution of each pair of source and target domains,and then multiple source domain-specific classifiers are learned,and the results of multiple classifiers are combined using decision integration strategies to obtain the prediction of the target sample.The average classification accuracy achieved by this method is 69.93%,which is an increase of about 5% compared with the results of the traditional multi-site ASD identification methods.(2)In order to solve the problem that it is difficult to find the shared subspace between the source and target domains,this thesis proposes an ASD identification method based on the landmark-based low-rank basis transfer method.The proposed method does not involve shared subspace,and directly uses the basis vector of the target domain data to represent the source domain data,so that the data distribution of the two domains becomes similar.To eliminate the influence of abnormal points in the target domain,a clustering algorithm is used to select the landmark points of the target domain to obtain more representative low-rank basis vectors of the target domain data.Aiming at the problem of imbalance in the sample size of the source and target domains,the source domain data is sampled by category.This method has achieved a diagnostic accuracy of 82.01%,and the ASD identification performance has been significantly improved.(3)Based on the adversarial network and classifier ensemble learning method,this thesis proposes a multi-source deep adversarial domain adaptation method.Two different convolutional neural networks for each pair of source and target domains are designed to extract the shared feature between the sites and the private feature of the site.Then through the gradient reversal layer to balance the domain discriminator loss and the classifier loss,,domain invariant and discriminative shared feature representation is obtained.Combining the class label and domain label information,a new loss function is designed,which introduces the classifier difference loss,so that the prediction of the target domain data by multiple source domain specific classifiers is as consistent as possible.In order to improve the ASD identification performance,the maximum mean discrepancy between the source domain and target domain data is also introduced as the classifier weight,and the results of multiple source domain-specific classifiers are integrated to predict the label of the target domain data.The proposed method achieves an identification accuracy of 65.93%,which is better than other traditional domain adaptation methods.