Research On Key Technology For Autism Spectrum Disorder Diagnosis Based On Small-sample FMRI Datasets With Artifacts

Autism spectrum disorder(ASD)is a lifelong neurological disorder.According to the development report of China’s autism rehabilitation,there are more than 10 million ASD patients in China,including more than 3 million ASD children.Patients with ASD have deficits in social communication and interaction,with restricted repetitive patterns of behavior,interests,or activities.Relevant studies have shown that the sooner inter-vention is carried out,the more significant improvements in social skills,communi-cation skills,and behavioral patterns can be achieved in patients.Therefore,accurate diagnosis and therapy interventions are necessary for ASD patients during their child-hood,which enhance their lives quality.However,due to the lack of basis for pathophysiological testing,physicians can only make diagnostic evaluations of potential patients based on behavioral observations or interview results.This subjective assessment method,which relies on the psycho-logical experimental paradigm,leads to misdiagnosis in the clinical diagnosis of ASD.The development of neuroimaging and artificial intelligence technology in recent years has brought new opportunities for the diagnosis of autism spectrum disorder.Us-ing machine learning and deep neural network technology to analyze brain functional magnetic resonance imaging data(fMRI)has become an important tool for understand-ing brain function disorder in patients with ASD.With the support of a series of ma-chine learning and deep neural network research,the intelligent diagnosis method of ASD based on fMRI data is gradually moving towards clinical application.Advanced problems seriously affect the diagnostic accuracy of neural networks.Therefore,how to use the small samples fMRI dataset with artifacts to carry out research on the diagnosis of ASD is a key issue to realize the application of artificial intelligence technology in the field of clinical diagnosis.In this paper,the research on the diagnosis of ASD is carried out from three aspects: enlarging the number of samples,improving the quality of features,and enhancing the representation of models.The main work and innovations are as follows:(1)Aiming at the problem of small number of samples in fMRI public data set,in this paper,a data augmentation framework based on signal reconstruction is con-structed.The empirical mode decomposition(EMD)algorithm is used to decompose the real samples of fMRI,and the decomposition results of several real samples are reconstructed and assembled into artificial samples.On this basis,a Serial-EMD opti-mization algorithm for fast decomposition of multi-dimensional signals is proposed to speed up the decomposition of multi-dimensional fMRI data? a SMOTE optimization algorithm is proposed to enhance the inline characteristics of artificial samples to re-duce intra-group differences of artificial samples.The experimental results show that the data sample augmentation method based on signal reconstruction has advantages in the speed and quality of artificial sample generation.On expansion method based on signal reconstruction has advantages in the speed and quality of artificial sample gen-eration.On the fMRI sample augmentation dataset,the average diagnostic accuracy of ASD can reach 68.4%,which is 7.8% higher than the original dataset.(2)Aiming at the problem of poor feature quality of fMRI public data sets,in this paper,a feature artifact completion framework based on tensor decomposition is con-structed,which realizes artifact completion by restoring the low-rank properties of fMRI feature tensors.On this basis,a normative modelling that maintains the constraints of differences between groups is proposed to isolate atypical feature subsets? the STDC algorithm that combines Tucker decomposition and rank minimization techniques is used to perform tensor decomposition and feature completion at the same time.The experimental results show that the feature artifact completion method based on tensor decomposition has advantages in the accuracy of artifact determination and the fitting degree of feature completion.On the fMRI feature completion dataset,the average di-agnostic accuracy of ASD can achieve 75.6%,which is 15.4% higher than the original feature set.(3)Aiming at the problem of high dimension of fMRI brain function connection characteristics,in this paper,we construct a graph neural network-based framework for ASD diagnosis,which represents brain functional connections through graph represen-tation learning.On this basis,an F-score rank dimensionality reduction algorithm for high-dimensional brain functional connectivity is proposed to reduce the complexity of brain functional connectivity? a multi-modal adaptive graph convolution model is proposed to diagnose ASD in a semi-supervised learning method.Experimental results show that graph neural networks have advantages in representing high-dimensional fea-tures and accurate representation of functional connections.On the multi-modal feature dataset,the average accuracy of the ASD diagnosis method based on graph neural net-work can achieve 87.7%,which is 4.7% higher than that of traditional neural network.In conclusion,the use of neuroimaging data to execute early diagnosis of ASD is the premise for subsequent intervention and improvement of patients life quality.For the small fMRI dataset with artifacts,this paper conducts research on data enhance-ment,feature completion,graph neural network and other technologies,and has made innovative progress in the three aspects of fMRI sample quantity,functional connection feature quality,and neural network model representation.The research results are of great significance for the analysis of neuroimaging data and the diagnosis of other brain diseases.