Innovative Strategies for Early Autism Diagnosis: Active Learning and Domain Adaptation Optimization.

Area of Science:

• Computer Science
• Medical Imaging
• Developmental Psychology

Background:

• Early diagnosis of autism spectrum disorder (ASD) is crucial but hindered by domain variations in facial image datasets.
• Existing models struggle with performance degradation when applied to diverse data sources.

Purpose of the Study:

• To investigate the efficacy of active learning, specifically uncertainty-based sampling, for domain adaptation in early ASD diagnosis.
• To improve the performance of deep learning models across different facial image datasets for ASD detection.

Main Methods:

• Analysis of domain variations in Kaggle ASD and YTUIA datasets.
• Assessment of transfer learning using Xception and ResNet50V2 convolutional neural networks.
• Application of uncertainty-based active learning for domain adaptation.

Main Results:

• Pretrained models achieved high accuracy (95-96%) on individual datasets.
• Combining datasets led to a performance drop.
• Uncertainty-based active learning mitigated accuracy decline, achieving 80% (Xception) and 79% (ResNet50V2) accuracy on target datasets.

Conclusions:

• Uncertainty-based active learning is effective for domain adaptation in early ASD diagnosis.
• This approach enhances model accuracy and reduces annotation requirements.
• Findings support the development of more robust ASD detection tools for diverse datasets.