Early Autism Detection via Multi-Stage Feature Fusion with AlexNet and Vision Transformer on Facial Images

Autism Spectrum Disorder (ASD) has seen an increasing prevalence, highlighting the importance of early detection to raise living standards and treatment outcomes. Traditional ASD diagnosis relies on observation, which can delay intervention. Machine learning (ML) and Deep Learning (DL) have shown potential in early ASD detection through the analysis of behavioural and physiological data. This study introduces a Multi-Stage Feature Fusion (MSFF) approach combining AlexNet and Vision Transformer (ViT) architectures for ASD detection using facial images. Pre-processing steps, including noise removal using a median filter, contrast enhancement using Contrast-Limited Adaptive Histogram Equalization (CLAHE), and data augmentation, are applied to facial images. The AlexNet model is employed for feature extraction, and the MSFF model is used to prevent overfitting. Furthermore, a Concentrate-and-Enhance (CE) attention model is used to decrease computational complexity and improve feature relevance. The ViT model is used for efficient autism detection by using its self-attention mechanism. The proposed approach shows promising results in identifying unique facial characteristics associated with ASD, offering a novel pathway for early diagnosis and intervention.