Alzheimer’s Disease Detection From MRI Using Deep And Transfer Learning Approaches

Abstract

Alzheimer's disease is a common neurological disorder that affects millions of people around the world. It is important to find the disease early so that it can be treated effectively and in a fast manner. Not long ago, deep learning methods became useful for looking at medical images and helping to find diseases. This work shows a new way to use MRI images to find people with Alzheimer's disease by combining deep learning and transfer learning methods. This work focuses on using deep learning approaches more especially, transfer learning to identify Alzheimer's disease with MRI data. It compares with popular deep learning models using ResNet50, VGG19, Xception, DenseNet169 for image classification. Our custom CNN architecture performs better than these mature models, with the single layer perceptron net giving a very high 97% accuracy. The work contributions discuss the pro and cons of each model and detail of our design choice for the custom CNN and its capability to learn complex patterns. The custom CNN design is very detailed, showing why it made the choices it did and how its hyper parameters are set up. The thesis goes into detail about the reasoning behind each design choice, which helps show how the model best finds complicated patterns in the data. This in-depth study not only helps us understand how deep learning works, but it also gives us a good starting point for building unique architectures in the future. In addition, the study looks into whether information gained from the custom CNN can be used in other areas. It focuses on how flexible and useful the proposed architecture is for generalization. The findings of the study are shown to be important beyond the current dataset by talking about their practical implications and possible uses in the real world. In the end, this thesis gives a full review of well-known deep learning architectures and offers a new custom CNN architecture that does better at image classification tasks than existing models. The study adds to the current conversation in artificial intelligence by showing how useful customized model design can be and by giving useful suggestions for future progress in the field