Abstract: A self-attention mechanism based model is proposed for the prediction of Alzheimer's disease (AD). Magnetic resonance imaging (MRI) images were pre-processed to extract primary features for brain anatomical structures. A self-attention mechanism based feature processing unit (SAFPU) was designed, and by the theory of residual blocks, multiple SAFPUs were stacked to build a reliable network for predicting AD, which could automatically analyze the dependencies of different brain anatomical structures to generate high-level features for MRI images. The empirical results demonstrate the proposed model outperforms existing AD classification methods, which achieves 99.36% (98.90%) of the maximum accuracy for the AD (early stage of AD, i.e., Mild Cognitive Impairment) classification task.