With the acceleration of global population aging, Alzheimer’s disease (AD) has turned into a progressively formidable public health problem. As deep learning technology has achieved substantial headway, a growing quantity of deep learning approaches have been utilized in the domain of medical image analysis. Against this backdrop, we herein present a novel 2.5D deep learning model, specifically designed for the diagnosis of AD. The 2.5D diagnostic model comprises two modules. The first is a 3D spatial feature compression module, which allows us to compress the spatial features of 3D MRI image data into a 2D feature map. The second is a 2D feature extraction module, which utilizes a 2D Swin Transformer to extract global features from the 2D feature map, thereby obtaining classification results. The 2.5D model alleviates the computational resource demands of 3D models and significantly shortens the training cycle, while overcoming the issue of 2D models being unable to capture spatial information, ensuring excellent diagnostic accuracy. We validated the performance of this 2.5D model on the ADNI dataset and focused on analyzing the model’s advantages in terms of computational volume. Our 2.5D model has a significant advantage in computational volume, reducing it by 8 times compared to traditional 3D deep learning models while maintaining a high level of accuracy.