Deep Learning-Based Feature Extraction from Children's Facial Images for Autism Spectrum Disorder Detection

Autism spectrum disorder (ASD) is a complex neurological developmental disease that appears in a variety of ways[8]. Most autistic children are not diagnosed until they are at least three years old [6] [13]. Abnormal facial traits caused by development abnormalities are strongly linked to underlying neurological issues [2]. Newborns with ASD have facial development that differs significantly from that of typically developing children [3]. Early detection is crucial to help the parents and families in denial and superstition. Differentiating facial features from normally developing children is a clear way to identify children diagnosed with ASD. In this work, we proposed a facial feature-based ASD detection method. This study uses different deep-learning models to recognize autistic children using only a facial image dataset. This work used five pre-trained CNN models as feature extractors: MobileNet, Xception, VGG16, VGG19, EfficientNet, and a binary classifier to detect autism in children effectively. To train these feature extraction models, we utilized a publicly available Kaggle dataset containing 2940 images of autistic and non-autistic children. The proposed method could detect ASD and non-ASD from the facial images of children. Using CNN models as feature extractors and Logistic regression as a binary classifier, we achieved an accuracy rate approaching 88.33% on test data.