A Multi-view Multimodal Deep Learning Framework for Alzheimer's Disease Diagnosis

Jianxin Feng^1*Xinyu Zhao¹Zhiguo Liu¹Yuanming Ding¹Feng Wang^2*

• ¹Communication and Network Key Laboratory, Dalian University, Dalian, China
• ²Dalian University Affiliated Xinhua Hospital, Dalian University, Dalian, China

Early diagnosis of Alzheimer's disease (AD) is challenging, primarily due to the high similarity in characteristics between AD, mild cognitive impairment (MCI), and cognitively normal (CN) individuals, as well as the influence of confounding factors such as population heterogeneity, label noise, and variations in imaging acquisition. Although multimodal neuroimaging techniques such as MRI and PET can provide rich complementary information, existing methods remain inadequate in multimodal feature fusion and multi-scale feature information aggregation. To address this issue, we propose a novel multimodal diagnostic framework called Alzheimer's Disease Multi-View Multimodal Diagnostic Network (ADMV-Net), which aims to improve recognition accuracy across all stages of AD. We first design a dual-pathway structure using a Hybrid Convolution ResNet module that effectively fuses global semantic and local boundary information, providing robust support for extracting feature information from three-dimensional medical images. Additionally, we introduce a Multi-view Fusion Learning mechanism that encompasses a Global Perception Module, Multi-level Local Cross-modal Aggregation Network, and Bidirectional Cross-Attention Module. This mechanism captures and efficiently integrates multimodal features from multiple perspectives, fully exploiting complementary information between modalities. Finally, we incorporate a Regional Interest Perception Module to precisely locate important brain regions associated with AD pathology, enhancing the model's perception of disease characteristics. Comparative experimental results on public datasets demonstrate that the proposed ADMV-Net achieves 94.83% accuracy and 95.97% AUC in AD versus CN classification tasks, significantly outperforming current mainstream methods. The model also exhibits excellent discriminative ability and strong generalization performance in multi-classification tasks.The code is available at https://github.com/zhaoxinyu-1/ADMV-Net.