Evaluating Deep Learning Models for Autism Detection in Children Using Facial Images

This study develops and evaluates a comprehensive deep-learning framework for early detection of Autism Spectrum Disorder (ASD) through facial image analysis. Five state-of-the-art convolutional neural network (CNN) architectures, VGG16, VGG19, ResNet50, InceptionV3, and MobileNet, were systematically assessed using a balanced dataset of 5,000 images (2,500 ASD, 2,500 non-ASD). Transfer learning and data augmentation enhanced model generalization. VGG19 achieved the highest overall accuracy (77.89%) and F1-score (0.7962), ResNet50 attained the best precision (82.53%), and InceptionV3 produced the highest recall (99.67%), indicating strong screening potential. The findings confirm that deep CNNs can capture subtle facial morphological cues linked to ASD, supporting their feasibility as non-invasive diagnostic tools. This work provides a benchmark for future multimodal, explainable, and clinically validated AI systems for autism detection.