AI For Early Detection of ASD: A VGG16-Based Child Facial Image Classification

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition whose
global prevalence continues to rise, yet conventional diagnostic procedures
remain costly, time-consuming, and inaccessible in many regions. This study
develops an ASD classification system based on children’s facial images
using a Convolutional Neural Network (CNN) with VGG16 transfer
learning, motivated by clinical evidence of phenotypic facial differences
between children with ASD and typically developing children. The model
was trained using the Piosenka Autism Image Dataset, comprising 2,936
facial images divided into training (2,536 images), validation (100 images),
and testing (300 images) sets with a balanced class distribution. The images
underwent preprocessing steps including resizing, pixel normalization, and
data augmentation to improve generalization. Experimental results show that
the model achieved an overall accuracy of 84%, with a precision of 0.80 and
recall of 0.90 for the ASD class, indicating relatively balanced performance
across both classes. Evaluation using a confusion matrix and ROC–AUC
curve further indicates good discriminative capability. The trained model was
deployed as an interactive, Streamlit-based web application that allows users
to upload facial images and receive real-time predictions. These findings
suggest that the proposed VGG16-based approach has potential as a
supportive, non-invasive tool for early ASD screening.