Neural network architecture for the classification of Alzheimer's disease from brain MRI

Abstract

Alzheimer's Disease (AD) is a neurological condition in which the decline of brain cells causes memory loss and cognitive decline. Various Neuroimaging techniques have been developed to diagnose AD; among those, Magnetic Resonance Imaging (MRI) is one of the most prominent ones. Historically, expert radiologists were solely responsible for making decisions of a patient's AD situation by manually analyzing brain MR images. However, the recent progress in medical image analysis using deep learning especially has automated this task significantly. Although the state-of-The-Art architectures have achieved human-level performance in classifying AD images from Normal Control (NC), they often require predefined Regions of interest as a basis for feature extraction. This condition not only requires specialized domain knowledge of the human brain but also makes the overall design complicated. In this paper, we designed a 14 layer Neural network architecture that can facilitate AD diagnosis without being dependent on any neurological assumption. The network was tested over ADNI-1, a benchmark MRI dataset for AD research, and found an accuracy of 87.06 % (\mathbf{AUC}=\mathbf{0. 9 3}.