Research On Prediction Of Autism Spectrum Disorder Based On Structural Magnetic Resonance Imaging

Autism spectrum disorder(ASD)is a highly heterogeneous neurodevelopmental disorder that severely affects the normal life of patients.Presently the clinical diagnosis of ASD is mainly based on the behavior scale,which has a certain degree of subjectivity.Structural magnetic resonance imaging(MRI)data can provide an objective basis for the diagnosis of ASD.Finding the brain structural biomarkers that can identify ASD based on structural MRI is helpful to improve the objective auxiliary diagnosis of ASD and to understand its neurobiological mechanisms.In this thesis,we use the structural MRI data from the autism brain imaging data exchange(ABIDE)dataset to carry out ASD qualitative prediction from two aspects of morphological features and gray matter texture features.Additionally,we conducted quantitative prediction of ASD severity based on morphological features.The main work is as follows:(1)The volume and cortical thickness features were extracted to perform ASD qualitative prediction based on single-center and multi-center datasets,respectively.Firstly,the t-test was used to select the features that are significantly related to ASD.Then,we used support vector machine(SVM)to perform ASD prediction on multiple single-center datasets,and the accuracy of most centers was over 70%.In order to further improve the prediction performance,a new method was proposed to use multi-kernel SVM to fuse the two types of morphological features:1)cortical curvature,and 2)volume and cortical thickness.Due to using the complementary information,the prediction accuracy of ASD for the NYU center achieved 79.46%,4.46%higher than the previous method.Due to the high heterogeneity of multi-center data,the prediction performance is not high.To reduce the heterogeneity of the data,the samples were divided into 9 groups according to age and verbal intelligence quotient.As a result,the prediction accuracy of most groups was significantly improved compared to when the samples were not grouped.(2)Most of the methods adopt morphological features to predict ASD,and rarely utilize texture features of gray matter images.In this thesis,independent component analysis(ICA)was used to extract fine texture features for ASD qualitative prediction.Firstly,the t-test was implemented to select the local discriminative sub-blocks.Then the exture features of these sub-blocks were extracted by ICA,and further refined with the t-test for prediction.The prediction accuracy of ASD for the NYU center was 68.75%.The use of texture features did not achieve better performance than morphological features.The preprocessing methods require to be improved to retain more texture information in the image,and more subtle information was extracted to boost prediction performance.(3)The current ASD studies mainly focus on qualitative prediction,but rarely predict disease severity of ASD.This thesis conducted quantitative prediction of ASD based on regional and interregional morphological characteristics.The prediction scores of regional and interregional morphological features were merged at the decision-making level.Then support vector regression was used to predict the severity scores of 71 ASD patients from the NYU center.The correlation between the predicted score and the true score was 0.5712.The results reached the advanced level in the current studies,which helps to predict ASD severity clinically.