Lightweight Deep Learning Model for Autism Spectrum Disorder Detection and Expression Recognition in Children Using Facial Images

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental illness that affects social skills, pronunciation, and communication abilities. Early diagnosis of ASD relies on detecting brain function abnormalities, which may be modest or absent at the beginning of the condition. Because children with ASD frequently exhibit distinct patterns that set them apart from normally developing children, facial expression analysis has emerged as an alternative and successful tool for the early detection of ASD. Identifying autism using facial expressions is difficult for both parents and physicians. However, Deep Learning (DL) can help address this issue. In this study, we developed a DL system to detect children's conditions, distinguishing between normalcy and autism. Additionally, the system identifies expressions such as happiness, sadness, and anger from facial images of children. For ASD detection, we employed MobileViT, while for expression recognition, we utilized VGG-16. The dataset used for training and testing comprised over 600 facial images sourced from the internet. The DL models' performance was assessed using standard evaluation metrics like accuracy and loss. During training and validation, the MobileViT model achieved peak accuracies of 99% and 98%, respectively, while the VGG-16 model attained peak accuracies of 96% and 99%, respectively. With these promising results, we proceeded to deploy the models on a website. The website interface allows users to simply upload a photo of a child, whereupon the models, MobileViT and VGG-16, discern the child's condition and expression, facilitating easy assessment.