Attention-Enhanced 2D CNN for Multi-Class Alzheimer’s Disease Classification with Grad-CAM Visualization

Abstract

Early and accurate diagnosis of Alzheimer’s disease remains a critical challenge in clinical neuroscience, particularly during the mild and moderate stages, when timely intervention can still have meaningful impact. In this study, we propose a deep learning framework based on a 2D convolutional neural network (ResNet-18) enhanced with a Convolutional Block Attention Module (CBAM) for multi-class classification of Alzheimer’s disease stages from brain MRI slices. To address the pronounced class imbalance inherent in the dataset, a class-weighted loss function was combined with aggressive online data augmentation during training. The model was evaluated on a large public dataset comprising over 86,000 MRI slices spanning four diagnostic categories: Non-Demented, Very Mild Dementia, Mild Dementia, and Moderate Dementia. Quantitative evaluation demonstrated a slice-level classification accuracy of 93.7%, with robust performance across all classes, including the underrepresented stages. Interpretability was provided through Grad-CAM visualizations, which highlighted anatomically meaningful regions consistent with established Alzheimer’s disease biomarkers. While these results underscore the potential of attention-augmented 2D CNNs for automated staging of Alzheimer’s disease, the framework has only been validated on a single public dataset and has not yet been tested on independent cohorts. Consequently, this study should be considered a proof-of-concept, and further multi-center, out-of-distribution validation is needed before any clinical deployment can be contemplated.