Work

An enhanced graph convolutional network with bidirectional cooperative cross-attention for early diagnosis of Alzheimer’s disease using multi-scale functional connectivity and structural connectivity

Year: 2026
Type: article
Abstract: Abnormal functional connectivity (FC) and structural connectivity (SC) during the progression of Alzheimer’s disease (AD) in specific brain regions may suggest potential biomarkers. In this paper, we first established approaches to computing multi-scale FC using resting-state functional magnetic resonance imaging (rs-fMRI) and SC using diffusion tensor imaging (DTI), while investigating FC and SC patterns at whole-brain and network levels across normal controls (NC), mild cognitive impairment (MCI), and AD groups. Then, we proposed an enhanced graph convolutional network model with bidirectional collaborative cross-attention (BCCA-EGCN) for AD recognition, which is primarily composed of the EGCN, BCCA, and multi-layer perceptron (MLP) modules. The EGCN module extracts deep features from functional graph and structural graph constructed utilizing multi-scale FC and SC, the BCCA module fuses these features, and the MLP module performs final classification. Experiments revealed that at the whole-brain level, AD showed significantly decreased FC and SC strength between the lateral occipital cortex and precentral gyrus versus NC (potential AD biomarker), and reduced FC across cognition-related regions versus MCI. At the network level, AD exhibited decreased FC within Somato-Motor B (SOM-B), Ventral Attention (VATTN), and Default Mode Network B (DMN-B) networks versus NC, and diminished FC specifically in SOM-B versus MCI. The BCCA-EGCN model achieved accuracies of 0.8974 (AD vs. NC), 0.8347 (AD vs. MCI), and 0.8021 (MCI vs. NC) via 5-fold cross-validation on ADNI. This work established a computational method that reveals multi-scale FC and SC evolution in cognition-related brain regions and networks throughout AD progression, and demonstrated the BCCA-EGCN model’s effective AD classification performance with significant clinical potential.
Language: English
Open Access status: green
Award: G683114012
OpenAlex ID: W7128781237
FWCI: 5.569
Cites: 53
Cited by: 3
Related to: 10