Research On Alzheimer's Disease Classification Algorithm Based On Convolutional Neural Network

Alzheimer’s disease is a degenerative disease of the nervous system that starts insidiously and develops progressively in the body.The disease is predominantly found in the group of people over 60 years old.With the arrival of aging society,an increasing number of elderly people will suffer from this disease.Studies have found that abnormal brain atrophy in the early stages of dementia is very alike to healthy brain atrophy,making the diagnosis of the MCI stage particularly important.s MRI can detect structural changes caused by brain atrophy,so it is widely used in computer-aided diagnosis of neurodegenerative diseases such as dementia.In recent years,machine learning methods that can consider the relationship between regions have become a popular computer-aided diagnosis technique,and have been widely used in the automatic diagnosis and analysis of neuropsychiatric diseases.In this thesis,AD classification algorithms based on convolutional neural network are studied.The specific work is described as follows:Firstly,an Alzheimer’s disease classification algorithm focused on feature reconstruction and convolutional neural networks is proposed.The algorithm reconstructs the artificially extracted features into 2D matrix and feeds them into a CNN model pre-trained on the source domain for learning.Specifically,it uses the extracted gray matter volume data as the source data to construct the feature reconstruction matrix.At the same time,a shallow CNN architecture is built and pretrained on the SVHN dataset.According to the size of the pre-training image,the matrix is flipped to expand the data set,and cropped to fit the CNN input.Finally,the proposed method is evaluated on the ADNI data set,and the accuracy rate is 90.2% in the AD/NC classification task and 79.1%in the p MCI/s MCI classification task.Secondly,an Alzheimer’s disease classification algorithm based on transfer learning and densely connected convolutional network is proposed.First of all,for s MRI,Gaussian filter is used for image enhancement before skull stripping to minimize noise and correct geometric distortion.Excess tissue is been removed according to threshold and morphological operations,gray matter image segmentation was performed using HEICA.Using image entropy to select the top 32 s MRI slices of each subject to reduce the input of redundant information.Using Dense Net-BC pretrained on Image Net as the network architecture to practice s MRI.Finally the proposed method was validated on the ADNI data set.In the four classification tasks of AD/NC,AD/MCI,NC/MCI and AD/NC/MCI,the accuracy rates of 96.43%,90.57%,91.21%,and 85.46% were obtained respectively.This thesis proposes two CNN-based AD classification algorithms to address the joint application of artificial extraction of s MRI features and convolutional neural networks,as well as the compactness of brain tissue segmentation and a limited sampling size of medical photographs that increases the precision of the classification of the various stages of Alzheimer’s disease.