SOGAF-Net: Spectral–Ordinal Gated Adaptive Fusion for Multi-Class Alzheimer’s Disease Severity Classification from Brain MRI

Abstract

Accurate classification of Alzheimer’s disease severity from magnetic resonance imaging (MRI) remains challenging because adjacent disease stages often exhibit overlapping structural characteristics and publicly available datasets are frequently imbalanced. This study proposes SOGAF-Net, an Ordinal–Spectral Adaptive Fusion Network for four-class Alzheimer’s disease severity classification. The framework integrates spatial representations extracted using a pretrained EfficientNet-B0 backbone with frequency-domain information obtained through a dedicated fast Fourier transform-based spectral branch. A sample-adaptive gating mechanism dynamically combines spatial and spectral features, while ordinal supervision explicitly models the ordered progression from non-demented to very mild, mild, and moderate dementia. Class-balanced focal loss and a categorical–ordinal consistency constraint are incorporated to improve discrimination across disease stages and mitigate class imbalance. Experiments were conducted using a public brain MRI dataset containing 6,400 images, with separate training, validation, and test partitions. SOGAF-Net achieved an accuracy of 96.64%, a macro-F1 score of 97.10%, and a macro one-vs-rest ROC-AUC of 99.83%. Bootstrap evaluation produced a 95% confidence interval of 95.55–97.58% for accuracy, indicating stable predictive performance. Grad-CAM-based visual analysis further provided interpretable evidence of image regions contributing to model predictions. The results demonstrate that jointly exploiting spatial, spectral, and ordinal information provides an effective framework for multi-stage Alzheimer’s disease classification from brain MRI.