Attention-Based Deep Neural Network for Screening Autism Spectrum Disorder in Children Analyzing Facial Images

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that hampers children's social, cognitive, and communication abilities. Early diagnosis is essential for improving long-term developmental outcomes. Recent advancements in facial image analysis have shown potential in supporting ASD screening, as specific facial biomarkers can help differentiate between autistic and non-autistic children. In this study, we propose a hybrid deep neural network that integrates three pre-trained convolutional neural networks Xception, InceptionV3, and EfficientNetB5 enhanced with channel and spatial attention mechanisms to improve feature extraction for ASD classification. The model was trained and evaluated on a publicly available facial image dataset, and it outperformed all other evaluated models, achieving an F1-score of 92.99%, along with precision, recall, and accuracy scores of 93.00 % each. Comparative analysis showed that MobileNetV2 is well-suited for edge deployment due to its lightweight architecture, while Xception (90.3%), ResNet50 (88.1%), and EfficientNetB5 (88.4%) also demonstrated strong performance. Additionally, experiments with different learning rates revealed that a rate of 0.0002 achieved the highest AUC of $\mathbf{0. 9 5}$, indicating the reliability and stability of the proposed framework.