A DEEP LEARNING FRAMEWORK FOR ALZHEIMER'S DISEASE DIAGNOSIS USING A VISION TRANSFORMER APPROACH

Abstract

Alzheimer's Disease (AD) is a progressive neurological disorder which impacts memory, thinking capability and cognitive functions. It's critical that Alzheimer's be diagnosed early and correctly, so that treatment and disease management can be undertaken as soon as possible. One new imaging method that has emerged as a valuable tool in determining brain changes associated with Alzheimer's disease is Magnetic Resonance Imaging (MRI). However, manual analysis of MRI scans is time-consuming, and relies heavily on interpretation skills. In this project, an automated Alzheimer's disease detection system was proposed, employing a pretrained Vision Transformer (ViT) model for feature extraction, and a hybrid classification model that combines K-Nearest Neighbors (KNN), Radius Neighbors, and Voting Classifier. Firstly, MRI brain images are acquired and preprocessed for resizing, normalization and augmentation to enhance image quality and diversify the dataset. The pretrained ViT model is able to learn deep and meaningful feature representations by considering local and global relationships in brain images using a self-attention mechanism.The extracted feature vectors are classified by KNN and Radius Neighbors classifiers. Voting Classifier: It takes the output of both classifiers and gives the final disease prediction with an aim to increase the reliability of prediction and to reduce misclassification. The proposed system divides the MRI images into four categories: Non-Demented, Very Mild Demented, Mild Demented and Moderate Demented. The experimental results have shown the proposed method provides better classification performance, robustness, and generalization capabilities than the single-classifier approaches. The developed system can be used as a computer-aided diagnostic system to help health care professionals in early detection and classification of Alzheimer's disease.