Research On Classification Methods Of Alzheimer’s Disease Progression With Brain Mri Based On Deep Learning

Alzheimer’s disease(AD)is one of the prevalent neurodegenerative diseases that leads to cognitive impairment in middle-aged and elderly people.Following the accelerated speed of population aging,the incidence of Alzheimer’s disease is on the rise,which brings a spiritual and economic burden to the patient’s family and society.Mild cognitive impairment(MCI)is a preclinical stage of AD.Early intervention in the condition of MCI patients is helpful to prevent and delay the occurrence of AD.Therefore,using effective methods to timely and accurately distinguish AD,MCI,and normal aging(NC)will play a positive role in the optimal management,prevention and treatment of Alzheimer’s disease.In this paper,to improve the accuracy of early diagnosis of Alzheimer’s disease and the interpretability of the model,the author studies the segmentation and classification method of MRI medical images based on deep learning,taking brain magnetic resonance imaging(MRI)data as the research object.Focusing on the feature extraction and classification of brain MRI images,the network models suitable for the classification of AD courses are constructed by combining deep learning technology.In order to improve the reliability of the deep learning model,various methods are used for interpretable analysis,which provides technical support for the early clinically accurate diagnosis of AD.The main work of this paper is as follows:(1)A Multi Res UNet brain tissue segmentation method combining dilated convolution and non-local mean attention is proposed.This method expands the convolutional receptive field through hole convolution to realize global information coordination.Multi Res Block and Res Path structures are used to reduce the difference between encoder and decoder features and reduce memory requirements.Considering the effect of high noise on brain sulcus detail segmentation,a non-local mean attention model is constructed to reduce noise interference in the mapped organization category.The experiment of brain tissue segmentation has achieved a better segmentation effect.(2)A dual-path fusion network based on ga SCE and gm SCE attention mechanisms is proposed and applied to feature extraction and classification of two-dimensional brain MRI images.A dual-path fusion network model is constructed to provide a basis for capturing feature pattern dependence,and multi-scale and multi-channel features are integrated into more expressive fusion features;the attention mechanism modules of ga SCE and gm SCE are designed to capture the feature dependence of MRI brain images in the channel dimension;a feature map reduction module is constructed increasing the field of view of the feature map and the proportion of feature area information;aiming at the problem that attention model will lose the dependency relationship between features,a weight function module is proposed to meet the dependency relationship of feature fusion,thus improving the classification ability of network model.Experiments are carried out on MRI images in the ADNI public data set,and the results show that the proposed methods improve the classification effect and obtain effective parameter compression.Ablation and contrast experiments verify the feasibility and superiority of the proposed method.(3)A lightweight-depth capsule network model based on Deep Caps combined with convolutional full connection capsule structure is proposed to classify and diagnose the course of AD.The Con-FC-Caps layer is constructed in the model,which converts the extracted features into vectors,reduces the input space of the convolution layer,and realizes the lightweight of the model;In order to restrain the non-information capsule and highlight the identification capsule,the compression function is designed by limiting the activation value of the capsule in the primary capsule layer;A regularization method combining similarity measurement between capsules and Dropout technology is proposed to improve the ability of model feature extraction and nonlinear representation.The generalization experiments on ADNI and OASIS public data sets show that this classification method achieves excellent classification results in the case of small samples.(4)A 3D CNN network model based on the MS-CAM-Inception module is proposed to classify the course of AD.The MS-CAM module is embedded in the model to improve the feature fusion mode and solve the problem of difficult fusion of medical image features;The 3D Inception structure is constructed to solve the problems of over-fitting and gradient disappearance;In order to improve the reliability of the network models,five interpretable techniques including Occlusion Maps,SA-3DUCM,3D-Grad-CAM,SHAP and LIME,are explored respectively.Through the visualization of model diagnosis process,the interpretability of the models is enhanced.