Detection of autistic individuals using facial images and deep learning

Abstract

Autism spectrum disorder(ASD) is a medical condition that causes major impairments to the neurology of the autistic individual. An autistic child has difficulty responding to their name, avoids maintaining eye contact, and lacks the ability to show emotions. Humans are social animals and the limitations brought about by ASD mars an individual’s overall development. ASD is normally diagnosed using brain images in childhood. However, this proves to be very expensive and takes a large amount of time. Recent studies have shown that ASD can be detected by making use of facial images. In this paper, deep learning models are pre-trained to classify facial images of children as either healthy or potentially autistic. Features such as eyes, nose, and lip distance in a child’s image and its arrangement can be an indicator of autism. Unlike the previous methods used to detect autism, the proposed method performs extensive pre-processing by removing the duplicate images, thereby, making it suitable for real-world applications. On training, the MobileNet model on facial images gave a maximum of 87% testing accuracy.