A Multiscale MixFormer Network for Computer-Aided Alzheimer's Disease Diagnosis

Deep learning has been widely utilized in medical-assisted detection and diagnosis. In the classification and diagnosis of Alzheimer's disease (AD), deep learning can combine low-level features into more abstract high-level representations to aid disease detection. In this paper, we propose a model architecture based on a Multiscale MixFormer. This architecture incorporates local window self-attention mechanisms with a multi-scale convolutional structure, enabling bidirectional interconnections across channels and spatial dimensions. This addresses the issues of limited receptive fields and inadequate modeling capabilities of standard local window self-attention models. In classification experiments between AD and normal controls (NC), AD and mild cognitive impairment (MCI), and MCI and NC, the model achieved accuracies of 95.09%, 93.58%, and 94.47% respectively. In a three-category experiment for AD, MCI, and NC, the 89.63% classification accuracy achieved outperforms related studies for these diagnostic groups. Our research provides an effective deep learning solution for the computer-aided diagnosis of Alzheimer's disease.