Lightweight deep learning framework for autism spectrum disorder detection from children's facial images

Abstract

Autism spectrum disorder (ASD) refers to a neurodevelopmental disorder that affects children’s social interaction, responsiveness, and behavioral patterns. Early identification of ASD is necessary for prompt treatment of children. The existing diagnosis procedures are subjective, take a lot of time, and require interpretation by an expert. In this study, an efficient network baseline 0 (EfficientNetB0)-based deep learning (DL) approach is developed for automatic ASD detection from children's facial images. The model is trained on enhanced training data by using a preprocessing-aware data augmentation strategy. The proposed augmentation method includes image processing operations such as face alignment, Gaussian smoothing, hue, saturation, value (HSV) normalization, contrast-limited adaptive histogram equalization (CLAHE), and horizontal flipping. The experimental analysis demonstrated the effectiveness of the proposed predictive model against existing methods on the same dataset, where the model achieved 93% test accuracy and an area under the curve (AUC) score of 0.97. Additionally, gradient-weighted class activation mapping (Grad-CAM)-based interpretation demonstrated that the model focused on semantically relevant facial regions, which offers visual support and transparency for the classification. The proposed predictive system is lightweight and provides a non-invasive and scalable solution to support ASD screening using facial images.