Abstract: Aiming at the problems of difficult detection and low recognition accuracy of brain cognitive diseases, a multi-head attention-guided convolutional neural network ( MAGINet: Multi-Head Attention-Guided Convolutional Network) is proposed. This integrates the local dependent modeling ability of the convolutional neural network with the global dependent modeling ability of the attention mechanism. It is used to identify NC (Normal), EMCI(Early Mild Cognitive Impairment), LMCI(Late Mild Cognitive Impairment), and AD (Alzheimer’s Disease), and to explore the complete evolution from NC through MCI(EMCI and LMCI) to AD. First, the complementary information of three FCN(Functional Connectivity Network) variants is integrated to form a multi-view learning framework. Secondly, a new multi-head attention module is designed in the convolutional neural network module in each view. By completing self-attention, channel attention, and spatial attention successively, it helps to model the global dependence relationship, compensates for the local mechanism of the convolutional neural network, optimizes the performance of the model, and proves its effective information extraction ability. Finally, the model is applied to several encephalopathy classification experiments to prove the strong universality and repeatability of the model. 

Key words: Alzheimer’s disease, convolutional neural network, attention mechanism, functional connection network