Progression-Aware and Explainable CNN–Transformer Framework for Multiclass Alzheimer’s Disease Staging Using MRI

<b>Background:</b> Alzheimer disease (AD) is a neurodegenerative condition that progressively develops structural changes in the brain, resulting in different stages of severity, which makes accurate multiclass classification from magnetic resonance imaging (MRI) challenging. Despite promising outcomes of deep learning models, a great number of current methods disregard disease progression, suffer from evaluation leakage, or lack interpretability. <b>Objectives:</b> This paper introduces DeepAttentionADNet, a lightweight hybrid CNN-Transformer framework designed for multiclass staging of Alzheimer's disease using MRI images. <b>Methods:</b> The proposed model integrates convolutional feature extraction with transformer-based global context modeling. To capture the ordered nature of disease severity, a progression-aware ordinal learning objective is proposed. Moreover, consistency regularization is utilized to enhance robustness by imposing consistent prediction with spatial perturbation. A leakage-free k-fold cross-validation protocol is adopted, in which data splitting is performed prior to augmentation. Also, to promote interpretability, token-level importance maps based on transformer embeddings are utilized to visualize spatial regions that were used to make classification decisions. <b>Results:</b> The experimental findings on a multiclass MRI dataset of Alzheimer disease demonstrate consistent and high performance across cross-validation folds (mean F1-score (0.991 ± 0.003) and AUROC (0.9998 ± 0.0002)), without losing transparency and progress awareness. <b>Conclusions:</b> The proposed framework provided a robust and interpretable method of Alzheimer disease severity classification using MRI.