An Explainable Connectome Convolutional Transformer for Multimodal Autism Spectrum Disorder Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Autism spectrum disorder (ASD) diagnosis is challenging due to heterogeneity and lengthy behavioral assessments.
• Automated neuroimaging tools show promise but struggle with multi-site data variability.
• Integrating functional magnetic resonance imaging (fMRI) and structural magnetic resonance imaging (sMRI) data is complex.

Purpose of the Study:

• To develop a multimodal deep learning framework, the Connectome Convolutional Transformer (CCTF), for accurate ASD diagnosis.
• To integrate functional and structural brain connectivity from fMRI and sMRI data.
• To enhance feature representation by incorporating diverse connectivity metrics and morphological properties.

Main Methods:

• Developed the Connectome Convolutional Transformer (CCTF), a multimodal deep learning framework.
• Integrated functional connectivity metrics from fMRI and structural covariance networks from sMRI.
• Utilized a connectome convolutional embedding module, transformer encoder, and node-to-graph pooling for biomarker identification.

Main Results:

• CCTF outperformed state-of-the-art methods on the multi-site ABIDE dataset.
• Achieved high accuracies: [Formula: see text] (fMRI), [Formula: see text] (sMRI), and [Formula: see text] (fMRI+sMRI ensemble) in intra-site validation.
• Demonstrated robustness and generalizability with [Formula: see text] accuracy in inter-site cross-validation.

Conclusions:

• The CCTF framework offers a robust and generalizable approach for automated ASD diagnosis using multimodal neuroimaging.
• Identified brain regions align with known ASD neurobiology, suggesting potential for advancing understanding of the disorder.
• CCTF shows promise in overcoming multi-site data integration challenges for neuroimaging-based diagnostics.