Integrating Facial Image Data for Autism Identification: A Comparative Evaluation of Deep Learning Classifiers

Autism spectrum disorder (ASD) is characterized by difficulties in social relations, communication, and behavior, and as was shown by the results of several studies, deep learning models may be used for automated diagnosis of subjects with this disorder. This study aims to explore how deep learning models can be used to classify images of autistic individuals' facial features. For this purpose's objectives, a comparative analysis through several deep learning classifiers namely VGG16, VGG19, VGG Faces, and Resnet50 was performed to assess their ability to classify images of faces of children. The dataset includes 2926 images and the balance between autistic and non-autistic people was maintained. To improve model performance and to counter overfitting, we performed some preprocessing operations such as image normalization and resizing and used some data augmentation techniques such as rotation, shearing, and flipping. The results showed that the ResNet50 model accounted for all other models' limitations and achieved the highest score of 97.14%. The results of the study are linked to the recruitment of deep learning tools for the detection of ASD providing a great hope for improved tools for earlier interventions.