MAMVCL: Multi-Atlas Guided Multi-View Contrast Learning for Autism Spectrum Disorder Classification

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Autism Spectrum Disorder (ASD) is a neurodevelopmental condition with significant early childhood plasticity.
• Early interventions including behavioral therapy, language, and social skills training can mitigate ASD symptoms.
• Accurate ASD classification is crucial for timely and effective intervention.

Purpose of the Study:

• To introduce a novel Multi-Atlas Guided Multi-View Contrast Learning (MAMVCL) framework for ASD classification.
• To leverage functional connectivity (FC) matrices from multiple brain atlases to improve diagnostic accuracy.
• To integrate imaging and phenotypic data for a comprehensive ASD diagnostic approach.

Main Methods:

• The MAMVCL framework integrates imaging and phenotypic data using a population graph structure.
• Graph convolution extracts global features, while a Target-aware attention aggregator captures local brain region dependencies.
• A graph contrastive learning strategy aligns global and local feature representations for consistency.

Main Results:

• The MAMVCL model achieved an accuracy of 85.71% on the ABIDE-I dataset for ASD classification.
• The proposed framework demonstrated superior performance compared to existing methods.
• Experimental results confirm the effectiveness of the multi-atlas and multi-view learning approach.

Conclusions:

• The MAMVCL model shows superior performance in ASD classification.
• Multi-atlas and multi-view learning approaches hold significant potential for enhancing diagnostic precision in ASD.
• The findings support the development of improved early intervention strategies for individuals with ASD.