Transformer-Based Brain MRI Classification for Early Alzheimer’s and Parkinson’s Disease Detection

Abstract

Alzheimer’s Disease (AD) and Parkinson’s Disease (PD) are progressive neurodegenerative disorders where early diagnosis is critical yet clinically challenging due to reliance on subjective observation. This paper proposes a transformer-based deep learning system for automated three-class classification of AD, PD, and cognitively normal subjects from structural brain MRI scans. Three architectures are evaluated: Swin Transformer, Vision Transformer (ViT), and BEiT (Bidirectional Encoder Representations from Image Transformers). A hybrid BEiT + Swin Transformer model is proposed, combining self-supervised masked image pre-training with hierarchical shifted window attention for robust neuroimaging feature extraction. Experiments on a balanced dataset of 450 brain MRI scans (150 per class) from ADNI and OASIS repositories show that the proposed model achieves 96.44% classification accuracy, outperforming standalone Swin Transformer (94.67%), BEiT (93.11%), and ViT (91.33%) baselines.