A deep learning framework for classifying autism spectrum disorder from children’s facial images using a multi-scale ViT architecture and edge computing

Aim: Early screening for autism spectrum disorder (ASD) using facial images is promising but often limited by small datasets and the lack of deployable models for resource-constrained settings. To develop and evaluate a lightweight framework that combines a multi-scale vision transformer (MS-ViT) with edge optimization for ASD classification from children’s facial images. Methods: We analyzed 2,940 RGB facial images of children obtained from a publicly available Kaggle dataset. Faces were detected, aligned, and cropped (ROI extraction), then normalized; training used standard augmentations. The backbone was an MS-ViT with multi-scale feature aggregation. We performed an 80/20 stratified split (training/testing) and used five-fold cross-validation within the training set for validation (i.e., ~64% training, ~16% validation, and 20% testing per fold). Edge deployment was enabled through post-training optimization. Performance was assessed using accuracy, sensitivity, specificity, AUC-ROC, and per-image inference time. Results: The best configuration (MS-ViT + Edge + Augmented) achieved an accuracy of 96.85%, sensitivity of 96.09%, specificity of 97.92%, and AUC-ROC of 0.9874. On a Raspberry Pi-class device, the model reached ~181 milliseconds per image, supporting real-time screening. Conclusions: The proposed “MS-ViT + Edge + Augmented” framework offers near-state-of-the-art accuracy with low latency on low-power hardware, making it a practical candidate for early ASD screening in clinics and schools. Limitations include dataset size and demographic diversity; prospective clinical validation on larger, multi-site cohorts is warranted.