REFENet: A Novel Hybrid Deep Learning Framework for Accurate and Efficient Autism Spectrum Disorder Detection

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by impairments in communication, social interaction, and behavioral flexibility. Early diagnosis is essential to ensure timely intervention. The problem is traditional diagnostic practices are often subjective, time-consuming, and limited in scalability. This research proposes a robust and efficient hybrid deep learning model designed for early ASD detection from facial images, integrates two state of the art architectures modified ResNet152 and modified EfficientNetB7 through a trainable fusion neural network is REFENet. The ensemble leverages the complementary feature extraction strengths of both models to improve accuracy and generalization with strong balance between accuracy and computational intensity without the need for high-end hardware or resource-intensive configurations. Experiments conducted on a publicly available ASD facial dataset show that REFENet outperforms standalone models, achieving 92% accuracy, 95% sensitivity, and an F1-score of 92%. These results highlight REFENet's potential as a reliable tool for real-world clinical and educational screening applications.