AD-ViTNet: Modeling Global Structural Dependencies in Brain MRI for Alzheimer’s Disease Detection Using Vision Transformer

Although convolutional neural networks (CNNs) have shown promising results for automated detection of Alzheimer’s disease (AD) from structural MRI, their reliance on local receptive fields limits their ability to capture the diffuse and spatially distributed neurodegenerative changes characteristic of the disease. To address this limitation, this study proposes the AD-ViTNet, a Vision Transformer-based framework to model global structural dependencies in brain MRI through self-attention mechanisms. The proposed AD-ViTNet framework trained and evaluated on structural MRI data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) processes MRI scans using patch-wise representations and multi-head self-attention to learn global contextual features across the entire brain volume. Comprehensive comparative experiments are conducted against widely used CNN architectures, including ResNet50, EfficientNetB0, InceptionNet, and VGG16, for binary classification between cognitively normal control subjects and individuals diagnosed with AD. Experimental results show that the AD-ViTNet consistently outperforms all benchmark models across multiple evaluation metrics. In the primary evaluation setting, AD-ViTNet achieves an accuracy of 96.8%, F1-score of 96.8%, and AUC of 0.983, indicating strong discriminative capability and robust feature representation. Statistical significance is verified using McNemar’s test, with all comparisons yielding p-values <0.05, confirming that the observed performance improvements are not attributable to chance. These findings demonstrate the effectiveness of global attention mechanisms in capturing heterogeneous atrophy patterns associated with AD and highlight the potential of transformer-based architectures for reliable, computer-aided neuroimaging diagnosis.