Title: Using CNNs for AD classification based on spatial correlation of BOLD signals during the observation

Abstract: Resting state functional magnetic resonance images (fMRI) are commonly used for classification of patients as having Alzheimer's disease (AD), mild cognitive impairment (MCI), or being cognitive normal (CN). Most methods use time-series correlation of voxels signals during the observation period as a basis for the classification. In this paper we show that Convolutional Neural Network (CNN) classification based on spatial correlation of time-averaged signals yield a classification accuracy of up to 82% (sensitivity 86%, specificity 80%)for a data set with 429 subjects (246 cognitive normal and 183 Alzheimer patients). For the spatial correlation of time-averaged signal values we use voxel subdomains around center points of the 90 regions AAL atlas. We form the subdomains as sets of voxels along a Hilbert curve of a bounding box in which the brain is embedded with the AAL regions center points serving as subdomain seeds. The matrix resulting from the spatial correlation of the 90 arrays formed by the subdomain segments of the Hilbert curve yields a symmetric 90x90 matrix that is used for the classification based on two different CNN networks, a 4-layer CNN network with 3x3 filters and with 4, 8, 16, and 32 output channels respectively, and a 2-layer CNN network with 3x3 filters and with 4 and 8 output channels respectively. The results of the two networks are reported and compared.