Appearance-Based Gaze Estimation for ASD Diagnosis.

Area of Science:

• Computer Vision
• Neuroscience
• Developmental Psychology

Background:

• Current autism spectrum disorder (ASD) diagnosis relies on specialized equipment like MRI and EEG, requiring hospital visits.
• There is a need for more accessible and convenient diagnostic tools for ASD.

Purpose of the Study:

• To introduce AttentionGazeNet, an appearance-based gaze estimation algorithm for accurate 3-D gaze tracking from raw video.
• To evaluate the algorithm's performance on benchmark datasets and its effectiveness in classifying ASD in children.

Main Methods:

• Developed AttentionGazeNet for appearance-based 3-D gaze estimation from video.
• Validated performance on MPIIGaze and EYEDIAP datasets, comparing against state-of-the-art methods.
• Applied accumulated histogram for spatial-temporal analysis of gaze and head pose data.
• Classified ASD using a self-collected dataset (ACVD) of children's videos.

Main Results:

• AttentionGazeNet demonstrated competitive performance on the MPIIGaze dataset.
• Achieved significant improvements on the EYEDIAP dataset: 14.7% for static head pose and 46.7% for moving head pose.
• On the ACVD dataset, the method achieved 94.8% accuracy, 91.1% sensitivity, and 96.7% specificity for ASD classification.

Conclusions:

• AttentionGazeNet provides an effective and efficient method for diagnosing ASD.
• The algorithm offers a convenient, appearance-based approach, reducing reliance on specialized medical equipment.
• This technology has the potential to improve early detection and intervention for ASD.