Research On Alzheimer’s Disease Classification Method Based On Attention Mechanisms And Fully Convolutional Network

Alzheimer’s Disease(AD)is a degenerative disease of brain nervous system whose onset is concentrated in the elderly.Its early symptoms are mainly memory decline,cognitive impairment,language expression ability decline and visual space ability damage.With the intensification of population aging,AD has brought a huge burden to patient’s family and social development.Mild Cognitive Impairment(MCI)is the early stage of AD and the important period of diagnosis and prevention for patients.If highrisk AD patients can be accurately identified at the early stage,the progression of disease can be delayed as much as possible.This paper uses clinical conventional Magnetic Resonance Imaging(MRI)to achieve AD classification by combining attention mechanisms,Fully Convolutional Network(FCN)and Multi-Layer Perception(MLP).The specific research contents are as follows:(1)A classification method(EA-FCN)based on External-Attention(EA)mechanism and Fully Convolutional Network is proposed for AD.The feature information of disease probability map of the sample is generated by training the Fully Convolutional Network,and obtain the disease state of local brain region;At the same time,External-Attention mechanism is introduced into the network to establish the longdistance dependence between image regions and potential correlation between samples,so as to obtain the hierarchical feature representation of image.In addition,the double normalization method of Softmax and L1 norm is used to reduce the weight difference of each pixel in the attention map,and avoid the influence of maximum or minimum eigenvalues on the attention map.The experimental results show that combine ExternalAttention mechanism with the double normalization method can learn the most discriminative high-order features of the dataset,which can improve the classification performance of the model for Alzheimer’s disease and the accuracy of early diagnosis for AD patients.(2)A classification method(ADF-PSA)combining Anisotropic Diffusion Filtering(ADF)and Pyramid Squeeze Attention(PSA)mechanism is proposed for AD.Firstly,different image filtering methods(median filtering,Gaussian blur filtering and anisotropic diffusion filtering)are used in image preprocessing to denoise and smooth MRI images,and preserve edge features;Secondly,the Pyramid Squeeze Attention mechanism is introduced into the network,which can integrate the spatial information of feature maps at different scales and establish the long-term dependence between multiscale channel attention.The experimental results show that anisotropic diffusion filtering is more advantageous than median filtering and Gaussian blur filtering for the model to distinguish images of AD patients and Normal Cognitive(NC).And the Pyramid Squeeze Attention mechanism can enable the network to learn abundant multi-scale feature representations,so as to enhance the recognition ability of the model for AD images.Therefore,combining anisotropic diffusion filtering with Pyramid Squeeze Attention mechanism is great significance for exploring AD classification.This paper has studied the impact of External-Attention mechanism and Pyramid Squeeze Attention mechanism on model performance in AD classification tasks,which extract the feature representation of images based on spatial attention and channel attention respectively.The results of this paper show that combining attention mechanisms,Fully Convolutional Network and Multi-layer Perception can obtain more accurate feature information of disease probability map of AD patients,thereby can effectively improve the AD classification performance of model.In addition,anisotropic diffusion filtering method can suppress the interference noise of the MRI image,which is conducive for model to distinguish AD and NC images.