Research On Alzheimer’s Disease Image Classification Algorithm Based On Deep Learning

Alzheimer’s Disease(AD)is an irreversible neurodegenerative disease of the brain that has no cure once it develops.With the deepening of the aging degree in our country,the number of Alzheimer’s disease patients is increasing,which has brought a heavy burden to the family and society.The increasing number of patients has greatly increased the workload of doctors in related fields,and the rapid development of computer technology has made it widely used in the medical field.In recent years,the rapid development of deep learning technology has made deep learning technology widely used in the fields of medical image classification and segmentation.Therefore,the use of deep learning technology to classify the magnetic resonance imaging(MRI)of the brain,identify patients with Alzheimer’s disease and mild cognitive impairment(MCI)and normal control(NC),assist doctors,and improve the efficiency of diagnosis and treatment is of great significance to doctors and patients..In this paper,based on the convolutional neural network(CNN)technology,the MRI images obtained from the ADNI website are classified,in order to help doctors improve the efficiency of identifying AD and MCI patients through MRI images.Aiming at the situation that MCI patients were divided into pMCI that developed AD at 36 months and sMCI that did not develop AD,the classification efficiency was improved by further improving the model.The main research contents of this paper are as follows:1.The lightweight DenseNet model is combined with 3D convolution technology and a simple,parameter-free convolutional neural network attention module(SimAM)to complete two binary classification tasks of MCI/NC and AD/NC.In view of the situation that ordinary two-dimensional convolution cannot learn the spatial information of three-dimensional MRI images,this paper uses three-dimensional convolution,and at the same time,by adding an attention mechanism module(SimAM)that can pay attention to the full neuron weight of the feature map to enhance the model’s ability to extract key features.When building the model,by reducing the number of network layers,the computational efficiency is improved,and the over-fitting phenomenon when training a small number of MRI image data sets is slowed down,and the 3D-SimAMDenseNet model is successfully built.Finally,in the two classification tasks of MCI/NC and AD/NC,the classification scores of 86.66% and 93.33% were obtained,respectively.2.On the basis of the 3D-SimAMDenseNet model,a multi-scale feature extraction module based on atrous convolution is added,and three models with different parameters are integrated with the integrated learning idea,and the integration of the models is completed by probabilistic fusion of the results of the three models,which is the ensemble model of 3DMFSimAMDenseNet,which is used to complete the three-classification tasks of AD,MCI,and NC.The multi-scale feature extraction module is based on the hole convolution technology,which effectively improves the ability of the model to extract features.The model integrating three different parameters can alleviate the over-fitting situation and also obtain a good threeclassification accuracy of 72.22%.3.Based on the ensemble model of 3D-MFSimAMDenseNet,the model is extended to a structure that can train multi-modal data to classify and identify pMCI and sMCI.The features of pMCI and sMCI are too close.By training on multi-modal data,we can learn the distinguishing features between the two as much as possible,and obtain a better classification accuracy of 79.63% compared with the comparison algorithm.