Determining Alzheimer’s Disease Progression: Stage Identification Using Ordinal and Regular Deep Learning Classification on OASIS MRI Scans

Abstract

Deep learning offers transformative potential for automated Alzheimer’s dis-ease (AD) staging from brain MRI scans, supporting earlier clinical intervention and more personalised care. This paper proposes and evaluates a systematic deep learning pipeline for four-class ordinal AD stage classification using the OASIS MRI image dataset (86,437 images; 4 classes: Non Demented, Very Mild Dementia, Mild Dementia, Moderate Dementia). We compare four CNN architectures—a Basic CNN, EfficientNetV2B1, Basic+EfficientNetB0 hybrid, and Basic+EfficientNetV2B1 hybrid—under two classification paradigms: Ordi-nal Classification (cumulative binary encoding) and Regular Classification (cat-egorical cross-entropy). All models were trained on an NVIDIA Tesla T4 GPU (Google Colab) with batch size 32, Adam optimiser (lr = 0.001), ReduceLROn-Plateau scheduling, and early stopping (patience = 5). Augmentation included random flips, rotations, zoom, brightness adjustment, and Gaussian noise. The best performing configuration—Basic+EfficientNetV2B1 hybrid with Ordinal Classification—achieves 97.82% test accuracy, micro-average F1 = 0.978, macro AUC-ROC = 0.999, Scott’s π = 0.97, and Cohen’s κ = 0.97. McNemar’s test confirms that ordinal classification provides a statistically significant improve-ment over regular classification (χ2 = 61.47, p < 0.001). Ninety-five percent confidence intervals are reported for all key metrics. These results demonstrate that exploiting the inherent ordinal structure of Alzheimer’s severity labels materially and significantly improves classification performance.