Automated Diagnosis of Autism Spectrum Disorder: A Deep Learning Approach

Abstract: The Autism Spectrum Disorder (ASD) is an intellectual and developmental disorder that has a significant impact on behavior, communication, and social interactions. Conventional diagnosis is based on subjective assessment, which may result in late diagnosis. Due to improvements in Artificial Intelligence, automatic detection of ASD has gained significant interest in recent years. This study proposes a web-based system that uses deep learning Convolutional Neural Networks (CNNs) to detect whether children 12 years old and younger have features of autism spectrum disorders. This system utilizes pre-trained deep models such as VGG19, EfficientNet, and Inception to examine facial expressions and behavioral patterns and also process static images and live video streams with high accuracy in screening. The training set consists of approximately 6,000 face images included equal numbers of ASD and non-ASD children. The images are obtained from publicly accessible research and medical databases and hence bring diversity in terms of varying age groups and ethnic groups. To ensure maximum performance of the system, the set is split into 80% training and 20% testing with an equal proportion of ASD and non-ASD samples. This system also includes real-time analysis through an integrated live-streaming module, enhancing the convenience of parents and clinicians. According to the experimental findings, EfficientNet performed better than the other deep learning models in terms of classification accuracy. Adding real-time video analysis further enhances our system's reliability. For increased accessibility through web-based platforms, this research will create a mobile application for ASD detection on smart phones. This study demonstrates the potential of AI-powered tools to enable early ASD diagnosis and reduce the application of traditional screening tools. The future work will cover the model's generalizability through the use of a more diverse and large set of data, multi modal input such as speech and eye-tracking, and optimizing the mobile application for real-world deployment. Through the creation of AI-powered autism screening, this study contributes to the cause of early intervention and improved access to healthcare.