Extract image markers of autism using hierarchical feature selection technique

Abstract

Autism spectrum disorder (ASD) can be described as a cognitive and behavioral impairment associated with a group of polygenetic brain disorders. ASD affects 1.4% of the children in the population. Structural magnetic resonance imaging (sMRI) has provided researchers with several means to examine structural changes in the brains of individuals with ASD. Using sMRI, we can extract morphological features such as surface area, cortical thickness, cortical curvature, folding index, and volume to describe the structural changes in autism. The biggest hindrance to applying machine learning in autism, or neuroimaging in general, is the curse of dimensionality, or sometimes being called small-n-large-p problem. A novel framework is designed and implemented to perform feature selection recursively and use the resulted features to train different machine learning models. The aim of this framework is to highlight the feasibility of having a set of imaging markers for autism from the morphological features of the brain. Classification accuracy of up to 82% using neural network and up to 72% using support vector classifier have been achieved on ABIDE I preprocessed dataset and FreeSurfer.