Multimodal Dense 3D Convolutional Networks for Alzheimer's Disease Classification

The multi-modal data fusion has gained prominence as a critical strategy for enhancing diagnostic accuracy by leveraging the complementary features across diverse imaging modalities. Nevertheless, existing approaches are often hindered by challenges such as inadequate feature alignment, scale inconsistencies, and suboptimal fusion strategies. To overcome these limitations, this study introduces a novel Multi-modal Dense 3D Convolutional Network (MD3DNet) framework that integrates a 3D Spatial Pyramid Decomposition module (SPD3D) for lossless downsampling and multi-scale feature encoding. Specifically, the modality alignment module (MSDI3D) that employs adaptive pooling and trilinear interpolation is complemented to ensure cross-modal feature space consistency and address resolution disparities between modalities. Furthermore, the model incorporates a densely connected 3D convolutional architecture to enhance deepen representational capabilities. Experiments on ADNI database demonstrate that the proposed method enables high-precision automated classification of Alzheimer's Disease (AD) and its early stages, including Early Mild Cognitive Impairment (EMCI) and Late Mild Cognitive Impairment (LMCI), outperforming existing methods.