Using Convolutional Neural Networks for Autism Spectrum Disorder Detection

Abstract

Research in this area focuses on convolutional neural networks (CNNs) and their potential utility in
the diagnosis of autism spectrum disorder. To begin, the provided image is enhanced by pre-processing. In
doing so, we eliminate background noise and boost contrast. The data augmentation technique is relocating a
pre-processed image to increase the dataset's diversity and resilience. After data augmentation, feature
extraction is performed using a scale-invariant feature transform (SIFT). SIFT maintains unique image points
regardless of changes in scale, rotation, or illumination. By using these features throughout training, the CNN
model becomes better at spotting ASD patterns. The objective of building and training the CNN model is to
accurately classify input images as either having ASD or not. Through a series of iterative model improvements
and validations, the CNN achieves impressive performance in identifying ASD-related traits in facial images.
Thanks to the trained CNN model's seamless integration, users may see the outcome inside the app interface of
the Streamlit online application.