Research On Identification Methods Of Autism Spectrum Disorder Based On Multi-modality Magnetic Resonance Imaging

Autism spectrum disorder(ASD)is a complex neurodevelopmental disorder and the clinical diagnosis is subjective.Structural magnetic resonance imaging and functional magnetic resonance imaging can provide information about brain structure and function from two different modalities,which can provide objective biomarkers for the identification of ASD.Compared with the identification method based on single-modality image data,the accuracy of ASD objective auxiliary diagnosis can be improved by multi-modality datasets fusion.In order to utilize the shared and modality-specific features of the two modalities,a multi-modality fusion method based on feature representation learning is proposed for ASD identification in this article.Afterwards,for the characteristics of small samples of ASD multi-modality datasets,a method of multi-scale generative adversarial network is proposed for ASD multi-modality fusion.Finally,in order to reduce the computational complexity while learning the fusion information of multi-modality datasets,an ASD multi-modality fusion method based on low-rank tensor learning is proposed in this article.The specific work of this article is as follows:(1)A multi-modality fusion method based on coupling feature representation and latent feature representation is proposed for ASD identification in this article.In order to alleviate the problem of the large difference in feature dimension between modalities and to mine more discriminative features of structural data,the coupling feature representation is introduced to learn the coupling relationship of structural data features.The existing latent feature representation learning methods only consider the shared features or specific features of multi-modality datasets.This article introduces a set of latent spaces to jointly extract the shared features of multi-modality and the specific features of each modality.The proposed method achieves 70.03% accuracy of ASD identification,which is superior to the existing ASD multi-modality fusion methods.(2)Aiming at the characteristics of small samples of ASD multi-modality datasets,and considering from the perspective of obtaining data feature distribution information,a multi-modality fusion method based on multi-scale generative adversarial network is proposed in this article to identification ASD more accurately.In order to make full use of the data samples of each modality,this method introduces the concept of multi-scale in the traditional generative adversarial network.By dividing the samples from multiple scales,the generative adversarial network is used to capture the features distribution information of multi-modality datasets at different scales.The experiments results show that the proposed method achieves better performance than the existing multi-modality fusion methods,and the identification accuracy of ASD is 74.36%.(3)The method based on tensor fusion can explicitly capture the fusion information of multi-modality datasets by calculating the Cartesian product of multi-modality tensors and the method is stable,but the computational complexity is high.A low complexity multi-modality fusion method based on low-rank tensor learning is proposed for ASD identification in this article.The feature expansion and weight distribution in the coupled feature representation of structural data are improved to linear calculations;In the process of low-rank tensor multi-modality fusion,the multi-modality tensor representation are combined by direct summation.The identification accuracy of the proposed method is 76.06%,which is superior to the existing multi-modality fusion methods.The ASD multi-modality fusion methods proposed in this article can improve the identification accuracy of ASD,find the objective biomarkers,help to improve the objective auxiliary diagnosis of ASD,and deepen the understanding of the pathological mechanism of the disease.