A Transfer Learning and Multi-Scale Feature Fusion Framework for Early-Stage Alzheimer's Disease Classification from MRI Images

Abstract

Early and accurate classification of Alzheimer's disease (AD) stages from magnetic resonance imaging (MRI) remains a critical challenge in clinical neuroimaging, particularly for distinguishing early-stage cognitive decline from normal aging. Existing deep learning approaches applied to publicly available augmented MRI datasets frequently report near-ceiling accuracy; however, many of these evaluations are vulnerable to data leakage introduced by improper train/test partitioning of augmented image collections, which can substantially inflate reported performance. In this study, we propose a transfer learning and multi-scale feature fusion framework based on an EfficientNetB0 backbone pretrained on ImageNet, in which feature maps extracted from three distinct network depths are pooled and fused prior to classification. We identify and correct a data leakage pattern present in commonly used baseline pipelines for this dataset, applying a single stratified, leakage-free train/validation/test split verified through explicit disjoint ness checks. Our model achieves 99.23% accuracy on the leakage-corrected augmented test set and 99.36% accuracy on a fully independent, non-augmented holdout set, indicating strong generalization beyond the augmented training distribution. We further show that this generalization performance is sensitive to training duration, with extended training substantially closing a generalization gap observed under shorter training schedules. Interpretability analysis using Gradient-weighted Class Activation Mapping (Grad-CAM) demonstrates that the model attends to anatomically plausible regions, including the ventricles and cortical areas associated with neurodegeneration, both on correctly and incorrectly classified examples. These findings contribute a rigorously validated classification framework and highlight methodological considerations relevant to evaluation practices on this dataset family.