Deep Learning Approach for Screening Autism Spectrum Disorder in Children with Facial Images and Analysis of...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Developmental Pediatrics

Background:

• Autism spectrum disorder (ASD) presents significant social, communication, and behavioral challenges.
• Early intervention in ASD can enhance intellectual abilities and reduce symptoms.
• Clinical studies indicate distinct facial phenotypic differences between children with ASD and typically developing (TD) children.

Purpose of the Study:

• To propose a practical screening solution for ASD using facial images.
• To leverage deep learning, specifically VGG16 transfer learning, for ASD detection.
• To validate the model on a unique dataset of clinically diagnosed children.

Main Methods:

• Collected a unique dataset of facial images from clinically diagnosed children with ASD.
• Applied VGG16 transfer learning, a deep learning technique, to analyze facial images.
• Developed and evaluated a classification model for ASD screening.

Main Results:

• The deep learning model achieved 95% classification accuracy.
• The model obtained an F1-score of 0.95, indicating high performance.
• Results align with clinical observations of facial differences in ASD.

Conclusions:

• Deep learning models are viable for screening ASD using facial images.
• Facial phenotypic differences between ASD and TD children are supported by the study.
• Racial and ethnic factors are critical for the accuracy and viability of AI-based ASD screening.