Early Alzheimer's disease classification via structure and feature-based graph attention network from multi-center...

With the increasing Alzheimer's disease (AD) prevalence, morbidity, and mortality, its early diagnosis is particularly important. The multi-modal data can be used to analyze intermodal feature relationships and enhance the feature expression ability related to AD. The stable and discriminative features are obtained by fusing multiple regions of interest among different modalities. In this paper, a novel graph attention network is proposed that integrates image as well as non-image information for each subject, leveraging a structure-based, and feature-based approach. Initially, we devise a group sparse representation approach that takes into account the strength of the connections among the blood-oxygen-level-dependent (BOLD) signals in the corresponding brain regions of subjects, which constrains the correlation between brain functional networks. We utilize structural networks as a means of supervising the construction of functional networks to obtain brain networks, considering their biological significance. Then, the construction of the sparse graph involves incorporating gender and age information of the subjects with multi-center neuroimaging features. Eventually, the information on individual and common embedding is extracted from existing node features for classification. The diagnostic performance is improved by utilizing fused weight information obtained through the use of topological structure and its combination, parameter sharing, and attention mechanism. We evaluate the proposed method using both the public Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and the in-house dataset. The comprehensive experimental results indicate that our approach surpasses the conventional methods.