Deep neural network CSES-NET and multi-channel feature fusion for Alzheimer's disease diagnosis

Abstract

Alzheimer's disease (AD) is an irreversible brain disease. The structural Magnetic Resonance Imaging (sMRI) has been widely used in the diagnosis of AD. However, the characteristic information from a single-mode is not comprehensive. In this paper, we proposed a Convolutional- Squeeze-Excitation-Softmax-NET (CSES-NET) deep neural network combined with multi-channel feature fusion for the diagnosis of AD. First, three kinds of features were extracted including patches based on voxel morphology, cortical features based on surface morphology, and radiomics features. Next, the residual network CSES-NET was proposed to extract the deep features from the patch images in which the features were re-scaled in the residual structure in order to fit the correlation between channels. Then, the fused features of the three channels were applied to classify AD/EMCI/LMCI/NC with the fully connected neural network. Finally, radiomics and cortical features were combined with genetic data for genome-wide association study to assess genetic variants. We performed experiments with 1539 subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The experimental results verified that the proposed method improved the effectiveness of the model by extracting nonlinear deep features and fusing the multi-channel features. In addition, the genome-wide association study identified multiple risk SNPs loci which were associated with the pathological of AD and contributed to the early prevention and control of AD.