Enhancing Autism Spectrum Disorder Screening through Learning Geometric Features of Retinal Blood Vessels Using Graph Representation

Abstract

Autism Spectrum Disorder (ASD) is a neurological and developmental disorder that has effects on how people behave and interact. It is a lifelong condition, and those born with it show symptoms typically during one’s early childhood. Over the past two decades, the cases and prevalence of ASD have been on a steady increase, with those born in the United States in 2012 being around 4 times more likely to have ASD compared to those born 20 years before. The treatment of ASD is shown to be more effective at earlier ages, which is why being able to perform an early diagnosis is crucial for those born with the condition. However, most of the current tests for child ASD diagnosis rely on the analysis of behavioral patterns by other people. This becomes more of an issue the younger the child gets, because at the youngest ages, there are not many clear actions or traits that a baby can physically show, making early subjective judgment vague and unreliable. To address this problem, I proposed an ASD screening framework that utilizes geometric patterns from retinal blood vessel scans. The framework contains two modules, the vessel segmentation and the ASD prediction network. The segmentation network processes the retinal images and generates a vessel segmentation map, which is then used for the ASD prediction network for a prediction of the severity. Through experimental results, it is demonstrated that the proposed approach is feasible by achieving an accuracy of 94.2% on a public dataset.