APLG-Net: an anatomy-guided local-global hybrid network with progression-aware supervision for structural MRI-based NC/MCI/AD classification.

INTRODUCTION

Structural MRI-based Alzheimer's disease classification remains challenging due to subtle anatomical variations and the intermediate nature of mild cognitive impairment (MCI).

METHODS

We propose APLG-Net, an anatomy-guided local-global hybrid network with progression-aware supervision for NC/MCI/AD classification. The model integrates a global whole-brain encoder and a local ROI-based encoder, followed by cross-attention fusion and vector-gated integration. An ordinal supervision strategy is introduced to model disease progression.

RESULTS

On the ADNI dataset, APLG-Net achieves 87.1% accuracy, 86.4% balanced accuracy, 86.8% Macro-F1, and 85.6% MCI F1, outperforming CNN-based, Transformer-based, and hybrid baselines.

DISCUSSION: The results demonstrate that incorporating anatomical priors, local-global feature interaction, and ordinal supervision significantly improves MCI discrimination and overall classification robustness.