Abstract:
"Early diagnosis of autism spectrum disorder (ASD) is vital for better outcomes, but current
methods rely on observation, causing delays in interventions. This emphasizes the need for more objective and reliable screening, especially for early detection of emotional challenges. Accurately recognizing emotions in individuals with ASD could lead to improved social interactions, reduced isolation, and more effective interventions, ultimately enhancing their overall well-being.
In this research, the author proposes an early detection solution for ASD and emotion detection in children using facial images and deep learning approaches, coupled with Explainable AI (XAI). The solution addresses data imbalance integrates preprocessing techniques, and leverages Keras with pre-trained models (VGG16, VGG19, ResNet50, ResNet50V2, and MobileNetV2) for feature extraction. Model accuracy is compared across these architectures to identify the most effective one. The research employs the LIME method from the XAI library to enhance interpretability, allowing a deeper understanding of the model's decision-making process. The proposed solution undergoes rigorous evaluation using various testing metrics, demonstrating high accuracy, robustness, and interpretability, highlighting its potential in effectively detecting ASD and recognizing emotions through child facial images.
The author trained a model using the Keras application for feature extraction, achieving varying accuracies with different architectures: VGG16 (93%), VGG19 (93%), ResNet50 (96%), ResNet50V2 (92%), and MobileNetV2 (94%). The highest accuracy attained is 96%. To interpret results, the author utilized the LIME XAI technique, providing explanations for the model's predictions and highlighting influential regions"
