Diagnosis of Autism Spectrum Disorder Based on Brain Network Clustering

Developments in magnetic resonance imaging (MRI) provide new non-invasive approach—functional MRI (fMRI)—to study functions of brain. With the help of fMRI, I can build functional brain networks (FBN) to model correlations of brain activities between cortical regions. Studies focused on brain diseases, including autism spectrum disorder (ASD), have been conducted based on analyzing alterations in FBNs of patients. New biomarkers are identified, and new theories and assumptions are proposed on pathology of brain diseases. Considering that traditional clinical ASD diagnosis instruments, which greatly rely on interviews and observations, can yield large variance, recent studies start to combine machine learning methods and FBN to perform auto-classification of ASD. Such studies have achieved relatively good accuracy. However, in most of these studies, features they use are extracted from the whole brain networks thus the dimension of the features can be high. High-dimensional features may yield overfitting issues and increase computational complexity. Therefore, I need a feature selection strategy that effectively reduces feature dimensions while keeping a good classification performance. In this study, I present a new feature selection strategy that extracting features from functional modules but not the whole brain networks. I will show that my strategy not only reduces feature dimensions, but also improve performances of auto-classifications of ASD. The whole study can be separated into 4 stages: building FBNs, identification of functional modules, statistical analysis of modular alterations and, finally, training classifiers with modular features for auto-classification of ASD. I firstly demonstrate the whole procedure to build FBNs from fMRI images. To identify functional module, I propose a new network clustering algorithm based on joint non-negative matrix factorization. Different from traditional brain network clustering algorithms that mostly perform on an average network of all subjects, I design my algorithm to factorize multiple brain networks simultaneously because the clustering results should be valid not only on the average network but also on each individual network. I show the modules I find are more valid in both views. Then I statistically analyze the alterations in functional modules between ASD and typically developed (TD) group to determine from which modules I extract features from. Several indices based on graph theory are calculated to measure modular properties. I find significant alterations in two modules. With features from these two modules, I train several widely-used classifiers and validate the classifiers on a real-world dataset. The performances of classifiers trained by modular features are better than those with whole-brain features, which demonstrates the effectiveness of my feature selection strategy