A Hierarchical Global-Local Feature Fusion Network for Alzheimer’s Disease Classification

To improve the diagnostic accuracy for Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI) using structural Magnetic Resonance Imaging (sMRI) , this paper proposes a deep learning architecture by integrating Convolutional Neural Network (CNN) and Transformer frameworks, which is designed to efficiently capture both local and global features from sMRI data. Specifically, the model utilizes a Biformer block with its Bi-Level Routing Attention (BRA) to capture long-range dependencies across the whole-brain sMRI, an Inverted Residual Convolution block to enhance the extraction of local texture and edge features, and a Semantic Difference Module (SDM) to refine feature representation in boundary-blurred regions. Experiments on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate that the proposed model achieves classification accuracies of 92.56% for AD and 82.69% for MCI conversion prediction, outperforming several state-of-the-art classification methods.