Autism Detection for Toddlers using Facial Features with Deep Learning

Autism Spectrum Disorder is a neuro-developmental disorder linked to challenges in social interaction, communication, and behavior. Children with autism often struggle with speech, exhibit challenges in maintaining eye contact, and face difficulties expressing emotions. Early detection and intervention in ASD are necessary to reduce the adverse effects of this disorder. Traditionally, ASD diagnosis involves costly and time-consuming brain imaging procedures. But nowadays, deep learning algorithms have advanced, providing a new way to detect ASD early by analyzing facial images. Unique facial features are found in Autistic toddlers that could be utilized with deep learning, but there has been limited attention given to this area. This research aims to improve the effectiveness of deep learning hybrid models, such as Vision Transformer (ViT) combined with Principal Component Analysis (PCA) and Support Vector Machine (SVM), ViT with CatBoost and SHAP, ViT with XGBoost and SHAP, and VGG16 with XGBoost. These modifications are specifically designed for image recognition tasks, such as detecting ASD in children’s facial images. The proposed hybrid models bring about significant improvements over traditional architectures, decreasing the risk of overfitting and improving the models’ capacity to identify subtle facial features linked to ASD. The proposed approach involves important preprocessing steps, which makes it appropriate for real-life scenarios. The accuracy of the hybrid models’ performance is measured, and the findings show that the suggested hybrid models perform better than other conventional models at identifying ASD from images of children’s faces.