ASDVIT: Advancing Autism Spectrum Disorder Classification Through Robust Vision Transformers

Abstract

This work presents a comprehensive evaluation of four state-of-the-art Vision Transformer (ViT) architectures, DaViT, MaxViT, Swin, and MViTv2, and one CNN-based architecture, Residual Dense Networks (RDNet), for automated Autism Spectrum Disorder (ASD) classification using facial images from the Autistic Children Facial Dataset. Among the tested models, MaxViT achieved the highest accuracy of 91.2%, outperforming traditional CNN-based methods and establishing a new benchmark for AI-assisted ASD diagnosis. The study employs a robust two-stage training strategy, incorporating rigorous data preprocessing, extensive augmentation techniques, and strict evaluation protocols, including accuracy, F1-score, confusion matrices, and ROC curve analysis. The unified pipeline leverages the advanced attention mechanisms and hierarchical feature extraction capabilities of transformer models to capture both local facial features and global contextual information. Comparative analysis demonstrates that transformer-based models consistently outperform conventional approaches, with top-performing architectures achieving 89.1-91.2% accuracy. These results demonstrate the potential of Vision Transformers to advance medical AI diagnostics by delivering objective, scalable, and clinically viable tools for early ASD detection. The findings underscore a paradigm shift from subjective behavioral assessments to AI-driven diagnostic support, offering healthcare professionals improved precision and accessibility in neurodevelopmental disorder identification while establishing a foundation for future research and real-world applications.