Abstract:Alzheimer’s disease (AD) is an irreversible neurodegenerative disease, the most common form of dementia, affecting millions worldwide. Neuroimaging-based early AD diagnosis has become an effective approach, especially by using structural magnetic resonance imaging (sMRI). The convolutional neural network (CNN)-based method is challenging to learn dependencies between spatially distant positions in the various brain regions due to its local convolution operation. In contrast, the graph convolutional network (GCN)-based work can connect the brain regions to capture global information but is not sensitive to the local information in a single brain region. Unlike a separate CNN or GCN-based method, we proposed a brain-inspired global-local information fusion network (BGL-Net) to diagnose AD. It essentially inherits the advantages of both CNN and GCN. The experiments on three public data sets demonstrate the effectiveness and robustness of our BGL-Net. Our method achieved the best performance on three popular public data sets compared with the existing CNN and GCN-based methods. In addition, our visualization results of the learned brain connection on AD and normal people agree with many current AD clinical research.

BGL-Net: A brain-inspired global-local information fusion network for Alzheimer’s disease based on sMRI