Facial Image-Based Classification of Autism Spectrum Disorder Using Pretrained CNNs and Synthetic Data

The developmental disorder known as autism spectrum disorder (ASD) is typified by notable challenges with speech, nonverbal communication, and social interaction. The symptoms will manifest between the ages of two and five years. Every person experiences the severity of the symptoms differently, and, for the present time, there is no cure. If a child is diagnosed and receives proper intervention, it can help lessen the weight of the effects on their developmental trajectory. We compared various architectures of CNNs, i.e., EfficientNetB4, VGG19, Xception, InceptionV3, DenseNet121, VGG16, and MobileNetV2, for ASD detection based on facial images. Our results indicate that VGG19 could provide a maximum accuracy of 0.84 and outperform all the models. DenseNet121 reached its peak accuracy at 0.82, while VGG16 demonstrated the highest accuracy of 0.83. EfficientNetB4 and Xception provided an accuracy of 0.80, while InceptionV3 provided the best accuracy of 0.77. MobileNetV2, although computationally lighter, had an accuracy of 0.78. The outcome is in line with existing studies to highlight the effectiveness of CNN-based approaches for classifying ASD and indicates the potential of the models in supporting early diagnosis and helping in autism learning.