To address the classification problem of AD and cognitively normal (CN) patients, this study proposes a multi-atlas feature fusion model. First, the input brain imaging data undergo standardization and are then parcellated using different brain atlas templates. Fiber tractography is performed to compute adjacency matrices, constructing graph-structured data that incorporate topological information. To fully capture brain network characteristics, this study introduces a graph convolution module capable of extracting high-order topological features, along with a graph pooling module that reduces computational cost while preserving essential structural information. Finally, a multi-branch feature fusion framework is employed to effectively integrate features from different brain atlases, enhancing the model's generalization ability. Experimental results demonstrate that the proposed model achieves effective feature fusion across multiple brain atlas templates, efficiently extracting both local and global brain network features to improve classification accuracy. Additionally, the proposed graph convolution and graph pooling modules not only enhance classification performance but also significantly reduce computational costs, offering new insights into graph-based neuroimaging analysis.