Multiclass Alzheimer’s Disease Classification from MRI Images Using Full-Dataset Benchmarking and DenseNet121

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder and the most common cause of dementia worldwide. Early diagnosis remains challenging because structural brain changes may be subtle in the early stages and magnetic resonance imaging (MRI) interpretation requires significant clinical expertise. This paper presents a two-stage framework for multiclass Alzheimer’s disease classification using a large public MRI dataset containing approximately 44,000 images distributed across four categories: Mild Demented, Moderate Demented, Non-Demented, and Very Mild Demented. In the first stage, a full-dataset benchmark was established using all 44,000 images. A lightweight stochastic-gradient-based classifier was trained on down sampled grayscale inputs, achieving 40.47% accuracy, 0.7520 macro-AUC, and 0.3558 weighted F1-score. In the second stage, a transfer-learning-based DenseNet121 model was trained on a balanced subset of 4,000 images using 128×128 inputs over 10 epochs. The model achieved 89.17% best validation accuracy, 88.67% test accuracy, 0.8858 weighted F1-score, and 0.9725 macro-AUC. These findings show that transfer learning provides strong multiclass MRI discrimination; however, because the source dataset is augmented and up sampled, the results should be interpreted as image-level computational findings rather than patient-level clinical validation.