2D-MAPNet: A Multi-Attention Probabilistic Network for Automated Alzheimer’s Disease Classification from MRI with Uncertainty-Guided Attention

Abstract

Alzheimer's disease (AD) is the leading cause of dementia, affecting approximately 57 million individuals globally and projected to reach 152 million by 2050 [49]. Early three-class differentiation of Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and AD stages from structural T1-weighted MRI is clinically critical, yet existing deep learning approaches are challenged by severe class imbalance (AD constitutes only 15% of samples), morphological overlap between adjacent disease stages, and the absence of calibrated predictive uncertainty in deterministic systems. We propose 2D-MAPNet, a Multi-Attention Probabilistic Network that jointly addresses all three limitations through four synergistic components: (1) a Gibbs-MRF Probabilistic Layer employing Monte Carlo dropout [16,27] with T=20 forward passes to produce per-sample predictive uncertainty estimates; (2) the novel Uncertainty-Guided Attention (UGA) gate, g^UGA = q_att·exp(−γσ̂²), which is the first coupling of MC-dropout variance inversely into a multi-scale attention mechanism in any AD classification network; (3) a Multi-Scale Attention Module (MAM) providing hierarchical feature attention at three spatial resolutions; and (4) a 2D Masking Branch (2DMB) for seed-map guided spatial suppression of false-positive activations. Evaluated on the Falah/Alzheimer_MRI benchmark (n=6,400) under rigorous 5-fold stratified cross-validation, 2D-MAPNet achieves accuracy of 92.63±0.75% (95% CI: 91.91–93.20%), AUC=98.58±0.27%, MCC=0.878, and a directly measured Expected Calibration Error (ECE)=0.021, representing a 4× improvement over the deterministic baseline and the lowest ECE [29] among all compared architectures including ResNet50, DenseNet121, EfficientNet-B0, ConvNeXt-Tiny, Swin-T, and ViT-B/16. With only 0.21 M parameters, 2D-MAPNet is 122× more compact than ViT-B/16 while uniquely providing calibrated uncertainty quantification suitable for clinical triage.