DeepASD: Facial Image Analysis for Autism Spectrum Diagnosis via Explainable Artificial Intelligence

Early and accurate diagnosis of Autism spectrum disorder (ASD) is crucial, but current diagnoses are subjective, time-consuming, and expensive. Recent studies used deep learning for facial images to diagnose ASD. However, the criteria are still unclear. To address these issues, we applied an explainable artificial intelligence technique to four convolutional neural networks (MobileNet, Xception, EfficientNet, and an ensemble model). We utilized gradient-weighted class activation mapping to suggest ASD diagnostic criteria based on facial morphology features. We achieved a high AUROC of 0.89 with the ensemble models. Our study provides objective and easy-to-understand diagnostic methods for early diagnosis of ASD.