Abstract: Aimed at the problem that the current early diagnosis methods of Alzheimer’s disease (AD) are difficult to effectively use key regions and their interrelationships for diagnosis, an early diagnosis method of AD based on a multi-level selection strategy is proposed. In this method, the structural magnetic resonance imaging (sMRI) was divided into several image patches, and the voxel-level key information of the image patches was enhanced using an attentional convolutional neural network. A genetic algorithm was introduced to select image patch-level features that were decisive for diagnosis, and multi-head self-attention was used to mine and use the relationship between patches to guide feature fusion, so as to achieve AD diagnosis. The experimental results show that the method is superior to the existing comparison methods in terms of Acc, AUC and F1-score.