Alzheimer’s Disease Diagnosis via Specific-Shared Representation Learning in Multimodal Neuroimaging

Mild cognitive impairment (MCI) is a potential prodromal state of Alzheimer’s disease (AD), an irreversible progressive neurodegenerative disease. Early diagnosis and intervention of AD are crucial. Recent studies have shown that deep learning methods exhibit excellent performance in predicting AD from multimodal neuroimaging. However, previous studies tend to focus on one aspect of the analysis of intra-modal information or inter-modal shared information, while ignoring the integration of the two. To solve this problem, this paper proposes a fusion network that combines modality-specific representations with shared representations between modalities to assist AD diagnosis. The network first extracts modality-specific representations from each modality data, then introduces a modality shared representation extraction network to supplement modality shared information, and finally performs modality representation fusion and performs AD prediction. In addition, in order to reduce the interference of non-disease-related regions between modalities on shared information, the network combines the shallow representation extracted by the single-modal network to guide the extraction process of shared representation. Combining these representations can provide a comprehensive view of multimodal data, thereby more accurately assisting in AD prediction tasks. The experimental results of this article on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database prove the effectiveness and excellent performance of this model in assisting in the diagnosis of AD.