AtroFormer: Atrophy-Guided CNN-Transformer Network with Ordinal Severity Learning for Alzheimer's Classification from MRI

Alzheimer’s disease (AD) is a progressive brain disorder that leads to permanent cognitive decline. Detecting its severity at an early stage using structural MRI is essential for timely and effective clinical treatment. However, accurately classifying AD into different severity levels is difficult because the structural differences between neighboring stages are often very subtle, clinical datasets are imbalanced, and models struggle to generalize across different imaging centers. Traditional CNN-based methods are good at capturing local brain atrophy patterns but fail to understand long-range spatial relationships, while Transformer-based approaches lack the anatomical precision needed for detailed severity classification. To address these challenges, this study proposes AtroFormer, a hybrid CNN-Transformer model. It combines several key components, including Atrophy-Sensitive Region Attention (ASRA), Multi-Scale Feature Tokenizer (MSFT), Severity-Conditioned Cross-Attention (SCCA), and an Ordinal Severity-Aware Loss (OSA-Loss), to improve multiclass AD classification. The model was trained and tested on the ADNI dataset, which includes 1,763 subjects across four severity categories (CN, VMD, MiD, MoD) and 35,260 MRI slices. It was further validated on the OASIS-3 dataset with 1,412 subjects without any additional fine-tuning. AtroFormer achieved an accuracy of 95.42%, a macro F1-score of 93.87%, and a kappa score of 0.936 on ADNI, outperforming the best baseline models by 1.74% in accuracy and 2.63% in F1-score. On the OASIS-3 dataset, it reached an accuracy of 89.36% with only a 6.06% drop in performance, demonstrating strong generalization. These results show that AtroFormer is an effective and clinically practical approach for automatically assessing Alzheimer’s disease severity using structural MRI.