Multi-channel Deep Model for Classification of Alzheimer’s Disease Using Transfer Learning

Abstract

AbstractIn this paper, we perform the classification of Alzheimer’s disease (AD) using 3D structural magnetic resonance imaging (sMRI) images through 2D convolutional neural networks (CNNs). Most existing methods that use 2D convolutional neural networks for AD classification, extract 2D image slices from each 3D MRI scan along the three anatomical planes of view of the brain. A CNN is trained separately on images from each plane of view. However, these methods only consider images from one plane of view at a time which leads to loss of 3D information. We address this issue by proposing a novel way of using an ensemble of multi-channel convolutional neural networks wherein, given a location in the brain, a multi-channel model looks at images from all three planes of view at a time around that location to obtain 3D information from 2D images. Multiple such locations from the brain are considered, and ensemble learning is used to give predictions at a subject level. Transfer learning is adopted wherein each channel in the multi-channel network utilizes state-of-the-art pre-trained CNNs, customized to the classification task. The proposed model obtains 98.33% accuracy for the Alzheimer’s disease (AD) vs Cognitively Normal (CN) classification task, outperforming current state-of-the-art methods.KeywordsAlzheimer’s diseaseDeep learningMulti-channel modelTransfer learning2D CNN