EXPLAINABLE MULTI-TASK VISION TRANSFORMER FOR AUTOMATED AUTISM SPECTRUM DISORDER DETECTION FROM FACIAL IMAGES
Keywords:
Autism Spectrum Disorder (ASD); Early Diagnosis; Autism Screening; Multi-Task Learning; Vision Transformer (ViT); DeiT; Facial Image Analysis; Emotion Recognition; Explainable Artificial Intelligence (XAI); Attention-Based Explainability; Deep Learning; Medical Artificial Intelligence.Abstract
Timely interventions and cognitive, social, and emotional growth among children with Autism Spectrum Disorder (ASD) can only be achieved by early and accurate diagnoses of the condition. Traditional methods of diagnostic like behavioral testing and clinical observations are usually subjective, lengthy and require specialists. This paper suggests a Multi-Task Vision Transformer (ViT) model of automated ASD recognition in face images. The model classifies and recognizes auxiliary emotions together with ASD through a common DeiT based transformer backbone, which improves feature representations and generalization. Attention based explainability identifies the regions of a face that make the predictions and enhance transparency and clinical understandability. The model is compared to the conventional convolutional neural networks (VGG16, ResNet18) and single-task ViTs, which prove worse performance and stability. The proposed approach should be offered as a scalable, non-invasive, and interpretable solution to screening ASD, integrating multi-task learning and transformer-based feature extraction with explainable AI, which can be used both in the academic and medical contexts.
