Repository logo
 

DIAGNOSIS OF ASD FROM RS-FMRI IMAGES BASED ON BRAIN DYNAMIC NETWORKS

Date

2020-09-28

Journal Title

Journal ISSN

Volume Title

Publisher

ORCID

Type

Thesis

Degree Level

Masters

Abstract

Autism spectrum disorder (ASD) is a prevalent and heterogeneous childhood neuro-developmental disease with an estimated prevalence of 1% of the global population and 1 in 54 children aged 8 years in the United States. The resting-state functional magnetic resonance imaging (rs-fMRI) is a non-invasive technique showing fluctuations of functional activities of a whole brain through measuring blood oxygen level-dependent (BOLD) signals. Rs-fMRI does not require the active participation of the subject, hence, it is suitable to investigate aberrant neurobiological function in ASD. The application of rs-fMRI images in the ASD research has been growing for the past two decades. Machine learning is initially designed and used to analyze medical datasets. In the past two decades, lots of studies have been done in combining machine learning technology with medical diagnosis in specialized diagnostic problems. Through the application of machine learning classifiers for analyzing fMRI data, some exciting new information can be extracted from neuroimaging data. Furthermore, the autism brain imaging data exchange I (ABIDE I) has pooled neuroimaging and phenotypic data from 1112 subjects across 17 sites and shared it for the ASD research community. This has facilitated the development of machine learning models towards the automated diagnosis of ASD. In this thesis, a method for diagnosing ASD based on brain dynamic network (BDN) was proposed. The BDNs are constructed with time series rs-fMRI brain images. BDN can model directed influences among multiple regions of interest (ROIs) across the whole brain based on the assumption that the current state of specific ROI depends on the linear combination of the previous states of multiple ROIs. The least squares method with the forward model selection method was used to establish BDNs, and the Bayesian information criterion (BIC) was adopted as the model selection criteria to avoid overfitting. Since, BDN is constructed with the temporal dependency of rs-fMRI brain image data, it can capture more complex interactions cross multiple brain ROIs than the traditional functional connectivity networks. Subsequently, graph theory and complex network analysis were applied on DBNs, which are the weighted and directed dynamic networks, to extract more representative and discriminated features. Besides, a representative partition of the whole brain was discovered from BDNs. Salient ROIs and a representative module of the whole brain were proposed, which can help to unveil the latent biomarkers for ASD. Lastly, machine learning classifiers were trained with whole ABIDE I cohort to identify ASD. Especially, the accuracy of 88.1% was achieved, which is higher than any previously reported methods.

Description

Keywords

Autism Spectrum Disorder, Resting-state fMRI, Brain Dynamic Network, Least Square Regression, Complex Networks Analysis, Machine Learning, Support Vector Machine, Logistic Regression.

Citation

Degree

Master of Science (M.Sc.)

Department

Biomedical Engineering

Program

Biomedical Engineering

Part Of

item.page.relation.ispartofseries

DOI

item.page.identifier.pmid

item.page.identifier.pmcid