PAM-MoE-AD: A plane-aware multi-stage mixture-of-experts framework for Alzheimer's disease classification from sMRI

Alzheimer's disease (AD) is a progressive neurodegenerative disorder for which accurate and early diagnosis remains difficult. Structural magnetic resonance imaging is widely used to assess disease-related brain changes, but automated classification methods are frequently affected by class imbalance, inter-subject anatomical variability, and orientation-dependent structural patterns. This study addresses three recognised limitations of existing automated systems, namely orientation-blind fusion, fixed expert weighting, and the absence of subject-adaptive plane relevance estimation, by proposing a plane-aware multi-stage mixture-of-experts framework, designated PAM-MoE-AD, for T1-weighted structural MRI. Axial, coronal, and sagittal 2.5D slice stacks were constructed through a content-aware slice-selection procedure, and each plane was modelled with an independent Swin Transformer V2 encoder. Training proceeded through three sequential stages comprising plane-specific representation learning, supervised refinement of plane-specific experts, and mixture-of-experts fusion governed by a learned gating network. Experiments were conducted on subject-level data splits derived from the AD Neuroimaging Initiative. Under five-fold subject-level cross-validation the framework achieved a mean accuracy of 97.21% and a mean balanced accuracy (BACC) of 93.14%, with 95% confidence intervals from 96.71% to 97.71% and from 92.47% to 93.81% respectively. External evaluation on a 124-scan AIBL subset, performed without fine-tuning, retained an accuracy of 93.52% and a BACC of 91.31%. The framework outperformed three same-split baselines comprising a three-dimensional (3D) ResNet-18, a 3D Swin Transformer, and a multi-plane CNN. Quantitative Grad-CAM analysis grounded model decisions in established atrophic regions including the hippocampus and entorhinal cortex.