AlzheiNN: a convolutional neural network-based model for Alzheimer’s disease classification

Abstract

Alzheimer’s disease (AD) is a neurological degenerative brain disorder which leads to a continuous diminishing of medullary functions over time, usually affecting elderly people. Neurofibrillary tangles and the accumulation of ß-amyloid plaques throughout the brain are clinically diagnosed through pathogenesis. However, lack of knowledge for underlying reason leads to no cure for this disease. This becomes the main reason for the identification of AD at prodromal stages, and traditional methods often face limitations in terms of accuracy, reliability, and early detection. These traditional methods such as neuro-psychological tests and clinical assessments, are costly and time-consuming, which has less accessibility for all individuals. On the other hand, ML algorithms provide better and efficient results. Automatic and early detection techniques implemented using DL approaches has the potential for more efficiency and scalability and might be useful to predict a larger population for the disease. Although, there have been existing studies that provided CNN models with enough accuracy but optimized with minimal hyper-parameters, including batch size, learning rate and optimizer in transfer learning models. This study proposed a customized CNN model which proved to be 3.37% more efficient and accurate than the baseline model along with impressive precision and F1-score. In addition to that, a 5-fold stratified cross-validation strategy was implemented which gives an average model accuracy to be 98.47% ± 0.26%. It includes convolutional layers with fine-tuned hyper-parameters and filters which classifies probability of Alzheimer’s disease. This approach holds the potential to revolutionize early screening for AD and facilitate timely interventions.

Subjects