A deep residual attention-Recurrent model for early and multi-stage Alzheimer's disease detection

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that requires accurate and early diagnosis. Deep learning methods have shown significant potential for automated MRI-based AD classification.

A hybrid deep learning framework integrating ResNet152V2, Convolutional Block Attention Module (CBAM), and Bidirectional Gated Recurrent Unit (Bi-GRU) was developed for three-class classification of Normal Cognition (NC), Mild Cognitive Impairment (MCI), and Alzheimer's Disease (AD). The model was trained using 2,100 ADNI subjects and externally validated using 900 OASIS subjects.

The proposed framework achieved 94.5% classification accuracy with an AUC of 95.0% on the ADNI dataset and 92.8% accuracy on the OASIS dataset. Comparative analyses demonstrated improvements of 5.2-7.0% over baseline models. Ablation studies confirmed the contribution of CBAM and Bi-GRU to overall performance.

The integration of deep residual feature extraction, attention-based refinement, and sequential modeling effectively captures disease-related anatomical patterns. The results suggest that the proposed framework may support automated multi-stage Alzheimer's disease classification and provide a foundation for future computer-aided diagnostic systems.