A Transformer–Rule Fusion Framework for Early Multistage Alzheimer’s Detection from MRI Scans

Abstract

Alzheimer’s disease is the most common form of dementia among older individuals and poses a major global health challenge, with nearly 10 million new cases reported annually. This progressive neurological disorder leads to gradual neurodegenerative changes, beginning with mild memory impairment and advancing to loss of social interaction and environmental awareness. According to Alzheimer Disease International (ADI), approximately 75% of dementia cases remain undetected worldwide, making early diagnosis particularly difficult. Currently, effective solutions to halt disease progression remain limited due to the lack of reliable diagnostic and treatment methods. In this work, high-quality Magnetic Resonance Imaging (MRI) scans are analyzed using a pretrained Swin Transformer model to extract high-dimensional visual embeddings that capture subtle neuro-structural variations across four stages: Normal, Very Mild, Mild, and Moderate Alzheimer’s disease. These extracted features are evaluated using multiple baseline classifiers, including Generalized Linear Model (GLM), Generalized Learning Vector Quantization (GLVQ), and Perceptron, to establish comparative performance. The proposed Optimal Rule Fit (ORF) classifier achieves superior results by effectively modeling complex, non-linear relationships within the MRI feature space. A Tkinter-based Graphical User Interface (GUI) is developed to facilitate dataset upload, preprocessing, model training, evaluation, and visualization through confusion matrices and Receiver Operating Characteristic (ROC) curves. Additionally, the system integrates a remote prediction server enabling real-time, cross-device MRIbased Alzheimer detection. Explainable Artificial Intelligence (XAI) techniques further enhance interpretability by identifying brain-region atrophy, ventricular enlargement, and disease severity. With an accuracy of 99.43%, the proposed framework demonstrates strong reliability for early diagnosis, remote clinical support, and neuroimaging research.