Deep Learning-based Classification of MRI Images for Early Detection and Staging of Alzheimer's Disease

Abstract

Alzheimer's disease (AD) poses a significant challenge to modern healthcare, as effective treatment remains elusive. Drugs may slow down the progress of the disease, but there is currently no cure for it. Early AD identification is crucial for providing the required medications before brain damage occurs. In this course of research, we studied various deep learning techniques to address the challenge of early AD detection by utilizing structural MRI (sMRI) images as biomarkers. Deep learning techniques are pivotal in accurately analyzing vast amounts of MRI data to identify Alzheimer's and anticipate its progression. A balanced MRI image dataset of 12,936 images was used in this study to extract sufficient features for accurately distinguishing Alzheimer's disease stages, due to the similarities in the characteristics of its early stages, necessitating more images than previous studies. The GoogLeNet model was utilized in our investigation to derive features from each MRI scan image. These features were then inputted into a feed-forward neural network (FFNN) for AD stage prediction. The FFNN model, utilizing GoogLeNet features, underwent rigorous training over multiple epochs using a small batch size to ensure robust performance on unseen data and achieved 98.37% accuracy, 98.39% sensitivity, 98.50% precision, and 99.45% specificity. Most remarkably, our results show that the model detected AD with an amazing average accuracy rate of 99.01%.