Manuscript received February 1, 2026; revised February 24, 2026; accepted March 19, 2026; published June 12, 2026.

Abstract—Autism Spectrum Disorder (ASD) is a neurodevelopmental condition affecting social interaction, communication, and behavior. Early detection is critical, yet conventional diagnostic methods are often time-consuming and resource-intensive. This study proposes a deep learning framework based on decision fusion of multiple Convolutional Neural Network(CNN)architectures to classify facial images of children with and without ASD. Six architectures (ResNet, DenseNet, Xception, ShuffleNet, MobileNetV2, and EfficientNet) were fine-tuned through transfer learning, and their predictions were integrated via decision fusion to improve generalization and classification accuracy. Experiments were conducted on a primary dataset of 1,050 facial images collected from 70 children (35 ASD, 35 typically developing) aged 5 to 15 years, recruited from special and regular elementary schools in Banda Aceh, Indonesia. Data augmentation was applied to address the limited sample size. Model performance was evaluated using accuracy, sensitivity, specificity, precision, and F1-Score. Among the combinations tested, ResNet and EfficientNet together yielded the best results, achieving 90% accuracy and a 91% F1-Score. These findings suggest that decision fusion is a practical and scalable approach for early ASD screening in resource-limited settings.