Informative Feature-Guided Siamese Network for Early Diagnosis of Autism

Area of Science:

• Neuroscience
• Artificial Intelligence
• Developmental Pediatrics

Background:

• Autism spectrum disorder (ASD) is a complex developmental disability typically diagnosed around 3-4 years old.
• Early diagnosis and intervention, especially within the first two years of life, are crucial for improving ASD symptoms.
• Previous diagnostic methods primarily used intensity-based imaging features, overlooking valuable information from segmentation and parcellation maps.

Purpose of the Study:

• To develop an advanced, end-to-end AI model for early autism spectrum disorder diagnosis.
• To integrate discriminative features from segmentation and parcellation maps alongside traditional MRI data (T1w and T2w).
• To address the challenge of class imbalance in ASD datasets, common in neurodevelopmental disorder research.

Main Methods:

• Proposed an informative feature-guided Siamese network for early ASD diagnosis.
• Utilized T1w, T2w images, and features from segmentation/parcellation maps for model training.
• Incorporated a subject-specific attention module for end-to-end, automatic identification of ASD-related brain regions.
• Employed a Siamese network architecture to effectively handle class-imbalance issues by learning feature distinctions between classes.

Main Results:

• The proposed method achieved a high overall accuracy of 85.4% for early ASD diagnosis.
• Demonstrated strong performance with a sensitivity of 80.8% and specificity of 86.7%.
• Ablation studies and comparative analyses confirmed the effectiveness of the integrated approach, highlighting the value of segmentation and parcellation maps.

Conclusions:

• The developed Siamese network offers a promising, automated approach for early autism spectrum disorder diagnosis.
• Integrating multi-modal imaging features, including segmentation and parcellation maps, significantly enhances diagnostic accuracy.
• The method effectively mitigates class-imbalance issues, paving the way for more reliable ASD detection in clinical settings.