NeuroVision-XNet: A Hybrid Vision Transformer Framework for Early Alzheimer’s Disease Prediction using Structural MRI

Traditional Alzheimer’s Disease (AD) prediction models using imaging inputs are not likely to give a correct early diagnosis since they are formulated on hand-crafted features and do not use all the spatial dependence in the brain. It is progressively challenging to separate Alzheimer’s disease stages, Mild Cognitive Impairment (MCI), and normal aging based on these limitations. Thus, the study introduces NeuroVision-XNet (Neuroimage Vision Transformer and Cross-Spatial Feature Extractor Network), a deep neural model adopted for direct image-based Alzheimer’s prediction from structural MRI scans. NeuroVision-XNet’s hybrid vision transformer backbone through spatial attention fusion and localized patch-based convolutional refinement allows for comprehensive understanding of both localized abnormalities and global patterns of brain atrophy. The model was trained and evaluated on the ADNI dataset with T1-weighted MRI scans divided into Normal Control (NC), MCI, and AD classes. The results of the experiment with 99.6% prediction accuracy, 95.4% sensitivity, and 96.8% specificity demonstrate the stability of the model in clinical-level Alzheimer’s disease staging and diagnosis using medical images.