STDCformer: Spatial-temporal dual-path cross-attention model for fMRI-based autism spectrum disorder identification

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for understanding Autism Spectrum Disorder (ASD) mechanisms.
• Deep learning shows promise in analyzing rs-fMRI data for ASD research.
• Accurate differentiation between ASD and healthy controls using rs-fMRI remains a challenge.

Purpose of the Study:

• To propose STDCformer, a dual-path cross-attention framework designed to enhance ASD identification accuracy from rs-fMRI data.
• To effectively preserve and integrate both temporal and spatial patterns within rs-fMRI signals.
• To develop a more precise neuroimaging biomarker for ASD.

Main Methods:

• Developed STDCformer, a model with dual-path embeddings for temporal and spatial patterns.
• Incorporated perturbation positional encoding for temporal data and Gramian angular field similarity for spatial networks.
• Utilized interleaved cross-attention and 2D convolution for spatiotemporal feature extraction and analysis.

Main Results:

• STDCformer demonstrated competitive performance against state-of-the-art methods on the ABIDE dataset.
• The model successfully preserved spatiotemporal correlations at multiple scales, reducing information distortion.
• Interpretative analyses provided preliminary insights into potential physiological mechanisms of ASD.

Conclusions:

• STDCformer shows significant potential for improving the accuracy of ASD identification using rs-fMRI.
• The proposed framework offers a promising approach for developing advanced neuroimaging biomarkers for ASD.
• Deep learning techniques, exemplified by STDCformer, are valuable tools for advancing ASD research.