FOR EARLY ALZHEIMER’S DISEASE DETECTION FROM MRI: A COMPARATIVE STUDY AND VISION TRANSFORMER-BASED ENHANCEMENT

Abstract

Alzheimer’s disease is a brain problem that gets worse over time, causing people to lose memory, have trouble thinking, and experience changes in the brain's structure, so finding it early is very important for giving help and taking care of it. New improvements in brain scanning and computer methods have made it possible to create automatic systems that can find Alzheimer’s by looking at brain pictures from MRI scans. In this research, we study and compare different ways of using advanced computer learning to automatically sort MRI images to identify Alzheimer’s, using images from the OASIS public collection, which has detailed brain images and information about the patients' dementia levels. We used and tested three well-known image-processing designs: a special-made Convolutional Neural Network, EfficientNetB0, and ResNet50. Our systems were able to correctly classify images with 94.6%, 92.73%, and 98.24% accuracy, respectively, showing that ResNet50 was the best at telling the difference between people with normal brain function and those with dementia. While regular image-processing designs are good at finding small details, they often have trouble understanding how different parts of the brain connect over long distances, which is important for noticing small brain changes that happen early in Alzheimer’s. To fix this problem and improve on the usual CNNs, we suggest using Vision Transformers together with image-processing feature finders, making a combined design that finds both small local details and overall context in brain MRI scans. This method uses self-attention to understand how things relate over long distances, which could make the system better at finding the varied and spread-out signs of Alzheimer’s. Early tests with improved transformer systems show good increases in how well the system classifies images and create attention maps that are easier to understand, pointing out brain areas affected by the disease, which is helpful for doctors making decisions and for making the system easier to understand. Our results suggest that advanced computer learning systems, especially those that use transformer designs, are very promising for correctly and automatically finding Alzheimer’s disease. Future work will focus on growing this combined system to sort Alzheimer’s into multiple levels of severity, studying brain changes over time to predict how the disease will progress, and adding tools that explain how the system works to build trust and help doctors use it more in their work.