Multi-Modal Fusion Network Integrating Imaging and Clinical Tabular Data for Alzheimer's Disease Classification

Accurate diagnosis of Alzheimer's disease relies on the integration of multimodal data, with imaging providing structural and functional insights, and clinical tabular data capturing patient-specific variables. Together, these sources offer a comprehensive view of the disease. However, traditional approaches often fail to take advantage of the complex relationships between clinical data and neuroimaging, limiting the effective use of complementary information. Common methods for spatial and channel feature extraction struggle to capture deeper interactions within structural magnetic resonance imaging and positron emission tomography data. Additionally, these methods often overlook critical intermodality relationships. To address these challenges, we propose the image-table cross-merging network, a multimodal framework that integrates spatial and channel features from sMRI and PET via cross-attention. The framework employs an innovative image-table fusion module with multilayer cross-attention and self-attention, enabling robust interactions between clinical and imaging data to enhance multimodallearning. Experimental results demonstrate significant improvements in AD classification accuracy, highlighting its diagnostic effectiveness.