A SELF-SUPERVISED HYBRID 3D CNN–VISION TRANSFORMER FRAMEWORK FOR MULTI-STAGE ALZHEIMER’S DISEASE CLASSIFICATION FROM MRI

Abstract

Alzheimer's disease is an irreversible, progressive brain disorder that slowly destroys memory and thinking skills and, eventually, the ability to carry out the simplest tasks. Conventional deep learning methods usually find it difficult to simultaneously extract local anatomical features and long-range contextual relationships from brain MRI images. To this end, we introduce a hybrid3D CNN–Vision Transformer architecture for multi-stage Alzheimer’s disease classification in this work. The proposed model integrates volumetric feature extraction with a 3D CNN and global attention modeling via the Vision Transformer, together with a self-supervised pretraining approach. Experiments show that the proposed model can achieve an overall classification accuracy of 98.23%, which outperforms other ensemble CNN and multimodal fusion methods. The results suggest better feature representation, higher class-wise performance and more effective generalization among disease stages.