M³ASD: Integrating Multi-Atlas and Multi-Center Data via Multi-View Low-Rank Graph Structure Learning for Autism...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Engineering

Background:

• Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition requiring accurate diagnosis for timely intervention.
• Resting-state functional magnetic resonance imaging (rs-fMRI) is vital for ASD diagnosis but faces challenges with data heterogeneity across multiple centers and reliance on single brain atlases.
• Existing methods often overlook variations in imaging devices, acquisition parameters, and processing pipelines, limiting diagnostic generalizability.

Purpose of the Study:

• To develop an automated diagnostic method for ASD that integrates multi-atlas and multi-center rs-fMRI data.
• To address data heterogeneity issues inherent in multi-center neuroimaging studies.
• To improve the accuracy and interpretability of ASD diagnosis using advanced graph structure learning.

Main Methods:

• A multi-view, low-rank subspace graph structure learning framework (M³ASD) was proposed.
• Functional connectivity matrices were constructed using multiple brain atlases from multi-center neuroimaging data.
• Low-rank subspace projection and multi-view consistency regularization were employed to mitigate heterogeneity and extract discriminative features.

Main Results:

• The M³ASD method achieved an accuracy of 83.21% on the ABIDE-I dataset.
• The model demonstrated superior performance compared to existing methods for ASD diagnosis.
• The approach successfully identified common functional brain connections across different atlases.

Conclusions:

• The M³ASD method effectively improves diagnostic accuracy for ASD using multi-center rs-fMRI data.
• The framework enhances the interpretability of ASD diagnosis by revealing common functional brain network patterns.
• Validation on the ABIDE dataset confirms the robustness and effectiveness of the proposed approach.