Vision Based Screening of Children with Autism Spectrum Disorders Via Deep Learning Approach

Abstract

Researchers have discerned unique facial characteristics in children with autism, leading to the application of face recognition methods for early diagnosis. Notwithstanding the efficacy of deep learning (DL) models in image classification, issues such as overfitting and limited training datasets continue to exist. This research intends to develop a Convolutional Neural Network (CNN) Deep Learning model for the detection of faces of children with autism. The study employed the Kaggle Autistic Children Facial Dataset to create CNN models through transfer learning, utilizing MobileNet, GoogleNet, and VGG-16. Their performance was assessed using measurements of accuracy, sensitivity, and specificity. The results indicated that MobileNet attained the superior overall performance with accuracy of 0.94, sensitivity of 0.97, and specificity of 0.90, exhibiting robust generalization capabilities. Nonetheless, limits including dataset size and resource limitations were recognized, indicating that future research should prioritize the augmentation of the dataset and the utilization of cloud computing resources for enhanced model training efficiency.