Research On Prediction And Analysis Of Autism Spectrum Disorder Based On Resting-state Functional Magnetic Resonance Imaging

Autism Spectrum Disorder(ASD)is a complex and highly heterogeneous neurodevelopmental disorder.Its etiology and neurobiological mechanism are still unsolved problems in world medicine.At present,ASD mainly relies on semi-structured assessment scales for clinical diagnosis,which is subjective and lacks objective biomarkers.Resting-State Functional Magnetic Resonance Imaging(rs-f MRI)technology has become a powerful tool for clinical auxiliary diagnosis and biomarker search in ASD due to its non-invasiveness,high temporal and spatial resolution,and convenient acquisition.In recent years,machine learning technology has greatly promoted the development of ASD research,and promising results have been achieved in the qualitative diagnosis and quantitative assessment of ASD.However,due to the complexity and high heterogeneity of ASD,there are still many problems and challenges in the field of ASD research.First,most of qualitative diagnostic studies only use features extracted from a single network to identify ASD and cannot comprehensively characterize the subtle perturbation of the brain functional network caused by neuropsychiatric disorders.Secondly,the multi-center integrated ASD data have low identification performance due to the problem of inter-center heterogeneity.Although some studies have used domain adaptation methods to reduce the data distribution differences between multiple centers to a certain extent,they ignore the class discrimination of multi-center data.Finally,the vast majority of ASD studies are concerned about qualitative diagnosis,but after subjects are diagnosed with ASD,it is important to focus on how severe their conditions are.Quantitative assessment of ASD severity can make a more detailed diagnosis for patients,but currently there are fewer studies on quantitative assessment of ASD severity.To solve the above problems,this dissertation proposes new methods in both aspects of feature extraction of brain functional network and multi-center qualitative diagnosis,and explores quantitative assessment of ASD severity.The main research contents of this dissertation are as follows.(1)The features extracted from a single functional network cannot comprehensively represent the subtle disturbance of the brain functional network caused by neuropsychiatric diseases.To solve this problem,inspired by the idea of multi-view learning,this dissertation proposes an identification framework of fusing features from multiple group-sparse functional networks,which effectively fuses the consistent and complementary information of multiple group-sparse functional networks and enhances the feature representation of brain functional network.First,multiple brain function networks with different sparsity levels are constructed for each sample by changing the value of regularization parameter of the group-sparse regression model,and the topology of the brain function network at each sparsity level is the same among all subjects,which facilitates inter-subject comparisons.Then,multiple groups-sparse brain functional networks with adjacent values of regularization parameter are fused to enhance the common intrinsic topology and limit the error rate caused by different networks.Meanwhile,these group-sparse networks with different sparsity levels provide complementary information different from each other in a granular fashion to capture subtle disease-associated alterations in functional connectivity.To obtain a set of more meaningful and discriminative features,this dissertation proposes improved local weighted clustering coefficients that can quantify the subtle differences of each groupsparse functional network at local level.Experimental results show that the proposed multiple group-sparse functional network fusion method can effectively improve the qualitative diagnosis performance of ASD compared with the single network methods and other multiple network methods.Compared with the conventional local weighted clustering coefficients,the improved local weighted clustering coefficients can extract more discriminative between-group difference information and improve the performance.In addition,the brain region biomarkers of ASD are located and analyzed to provide reference for the clinical diagnosis and neurobiological mechanism of ASD.(2)The multi-center ASD identification methods based on domain adaptation do not consider the class discrimination information of multi-center data to more fully reduce the marginal distribution differences and conditional distribution differences between data from multiple centers.To solve this problem,this dissertation proposes an ASD identification method based on unsupervised multi-source domain adaptation via low rank and class discriminative representation(LRCDR).LRCDR first utilizes the low rank representation to project the source domain data from multiple centers and target domain data from a single center into a potential common subspace to reduce the marginal distribution differences between domains.Then,to reduce the conditional distribution differences of data from all centers to effectively improve the identification performance of multi-center ASD data,LRCDR further learns the class discriminative representation of data from multiple source domains and target domain,so that the data from all centers can enhance intra-class compactness and inter-class separability.For multi-center source domain data with known labels,minimizing the intra-class distance and maximizing the inter-class distance makes the multi-center source domain data have good discrimination.For single center target domain data with unknown labels,the intra-class distance is minimized by clustering and the pseudo-class centroids of the target domain are obtained.To obtain good inter-class separability for the target domain data,the closest source domain class centroid is searched for each pseudo class centroid of the target domain by centroid matching,and the distance between them is minimized,which can make the target domain data enhance inter-class separability via source domain data whose interclass distance has been maximized.The experimental results show that LRCDR achieves superior identification performance compared with the state-of-the-art domain adaptation methods and multi-center ASD identification methods.Ablation experiments also demonstrate that the class discriminative representation plays a crucial role in improving the identification performance of multi-center ASD data.In addition,based on large sample data,the abnormal resting-state functional connectivity biomarkers in ASD are located and analyzed.(3)There are fewer studies on the quantitative assessment of ASD severity,and the existing quantitative assessment studies based on rs-f MRI data have used limited sample data,unstandardized severity labels,and regional brain connectivity features for severity assessment,and so on.To address these problems,this dissertation uses whole-brain resting-state functional connectivity(RSFC)features and multivariate pattern analysis method to quantitatively assess ASD severity based on multi-center integrated ASD data and standardized surrogate-calibrated severity scores,and investigate a subset of connectivities among whole-brain RSFCs that are more contributive to ASD severity estimation.The experimental results show that ASD brain functional connections undergo notable alterations with the severity of ASD.There are obvious abnormalities in the connection modes of inter-network and intra-network connections,and the abnormal functional connections are mainly related to sensorimotor network,default mode network,and inter-hemisphere connectivities,while exhibiting significant left hemisphere lateralization,providing additional evidence of large-scale functional network alterations and neurobiological mechanisms in ASD.