Autism Spectrum Disorder Detection in Children Via Deep Learning Models Based on Facial Images

Abstract

Autism spectrum disorder (ASD) is a complicated psychiatric disease that causes difficulty in communicating with others, and restricted behavior, speech, as well as nonverbal interaction. Children with autism have unique facial characteristics that distinguish them from ordinarily developing children. Therefore, there is a requirement for a precise and automated system capable of early detection of autism in children, yielding accurate results. The objective of this research is to assist both families and psychiatrists in diagnosing autism through a straightforward approach. Specifically, the study employs a deep learning method that utilizes experimentally validated facial features. The technique involves a convolutional neural network along with transfer learning for the detection of autism. MobileNetv2, Xception, ResNet-50, VGG16 and DenseNet-121 were the pretrained models used for detection of autism.  The evaluation of these models utilized a dataset sourced from Kaggle, comprising 2,940 facial images. We evaluated the five deep learning models using standard measures like recall, precision, accuracy, F1 score, and ROC curve. The proposed DenseNet-121 model outperformed existing transfer learning models, with 96% accuracy rate. With respect to performance evaluation, the proposed method exhibited superiority over the most recent models. Our model possesses the capability to support healthcare professionals in validating the precision of their initial screening for Autism Spectrum Disorders (ASDs) in pediatric patients.