Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by various challenges in social interaction, communication, and behavior. Recent advancements in deep learning and facial feature extraction have significantly enhanced the diagnostic process for ASD. This research seeks to combine geometric feature extraction methods with deep learning models like ResNet50v2 to improve the accuracy of identifying facial traits associated with ASD. Geometric features help pinpoint specific facial landmarks, such as the distances between key points like the eyes, nose, and mouth, which may reveal subtle morphological differences in individuals with ASD. At the same time, ResNet50v2, a deep residual network, supports effective feature learning and classification. By merging these approaches, the model captures both manually crafted geometric features and those obtained through DL, leading to a more comprehensive representation of facial characteristics. The model is trained on datasets that include images of children with and without ASD, resulting in improved accuracy compared to traditional methods. This combined approach highlights the potential of machine learning in aiding the early diagnosis of ASD, providing a non-invasive and scalable option for healthcare providers. The results suggest that integrating geometric feature analysis with ResNet50v2 can significantly enhance the identification of facial traits linked to ASD, thereby promoting more effective and accurate screening practices.