Structure-Aware Routing 3D Swin Transformer for Early-Stage Alzheimer’s Disease Detection and Classification Using Structural MRI

Abstract

Alzheimer’s disease is a progressive neuroimaging disorder that causes severe cognitive and cellular decline in elderly people. Magnetic Resonance Imaging (MRI) is widely used as a non-invasive neuroimaging modality for evaluating brain atrophy patterns. However, existing models struggles to detect early stage structural changes due to rigid downsampling layers that blur critical tissue boundaries. To address this, SAR-Swin3D (Structure-Aware Routing 3D Swin Transformer) framework is introduced to effectively detect and classify the diseases stages. The process begins by feeding raw 3D T1-weighted brain volumes into patch embedding layer. The network then routes features through successive Swin3D blocks utilizing local 3D windowing methods to reduce whole-brain modeling complexity. A block-wise 3D SAR module replaces blind downsampling to extract continuous maps using coordinate attention. The proposed model introduces a specialized block-wise 3D-SAR module to extract continuous boundary maps. This mechanism maps subtle gray-matter tissue thresholds and prevents global, age-related brain shrinkage from complicating localized early stage disease pathology. Furthermore, local 3D windowing methods reduce whole brain modeling complexity from cubic to linear. The SAR-Swin3D model is evaluated on ADNI and OASIS datasets by performing accuracy of 98.47% for Cognitive Normal (CN) against Alzheimer’s Disease (AD), 95.04% for stable Mild Cognitive Impairment (sMCI) against progressive (pMCI) and 99.81%  for multi-class dementia staging.