ASDFace: Face-based Autism Diagnosis via Heterogeneous Domain Adaptation

Abstract

While the prevalence of children with autism spectrum disorder (ASD) has emerged as a major public health concern, approximately 25% of children with ASD are not being diagnosed. The standard instruments to diagnose ASD are time-consuming and labor expensive, resulting in long wait times for a diagnosis. There is strong evidence that facial morphology is associated with autism phenotype expression. We hypothesize that the use of deep learning on facial images can speed the diagnosis without compromising accuracy. However, collecting and labeling large-scale facial images of autistic is a complicated and expensive process, which makes it inapplicable to train accurate deep learning-based diagnostic tools. To address this problem, we present a heterogeneous domain adaptation model that adopts sufficient individuals’ labeled characteristic and behavioural data as source domain to execute the facial classification task in the target domain. We also deploy this model to a web-based platform named ASDFace. ASDFace aims to provide a free preliminary ASD screening tool that can aid diagnosis and help parents decide whether they should take their children to an ASD specialist for further consultation.