Magnetic Resonance Imaging And Deep Learning Based Classification Of Alzheimer’s Disease

Alzheimer’s disease(AD)is a neurodegenerative disease characterized by chronic progressive development,which is irreversible.China is about to enter a moderately aging society and the number of elderly population is increasing rapidly.The prevalence of AD is increasing year by year,which has caused huge economic burden and mental pressure to the society and families.The accurate diagnosis of AD is of great significance for the treatment of the disease,which is expected to help reduce the cost of long-term care.With the development of medical imaging and deep learning technology,the combination of the two is one of the main research hotspots of modern medicine.This thesis studies the classification of Alzheimer’s disease based on magnetic resonance imaging(MRI)and deep learning,mainly including the following two aspects.(1)The t-test and Dense Net-SVM learning classification algorithm for Alzheimer’s disease diagnosis is proposed.The brain region of interest is easy to lose the distinguishing features and the global image has a lot of redundant information in the existing methods.Therefore,the thesis proposes to establish a t-test statistical model at the voxel level of GM to automatically calculate the significance of voxels between AD and NC.Then the significance of slices from the axial plane,sagittal plane and coronal plane is calculated in turn.Finally,the image representation is obtained by concatenating the significance slices.Meanwhile,combined with the advantages of the strong nonlinear expression ability of CNN and the feature that SVM classification is only related to support vector and solve the global optimal problem,a method combining Dense Net and SVM is proposed to construct the classification model.Dense Net reuses the multi-level features of the network through skip connection to alleviate the problem of gradient disappearance,and then extracts the multi-level fusion features of the samples and inputs them into the subsequent SVM classifier for diagnosis.In addition,the proposed image representation method is compared with other image representation methods(hippocampus based method and global GM image based method),and the classification models of Dense Net and Dense Net-SVM are compared.The experimental results show that the classification performance of the proposed method is improved.(2)Anatomical landmarks and DAG network feature learning algorithm for Alzheimer’s disease diagnosis is proposed.To solve the problem that the size of the image representation extracted by the method based on t-test is large,the thesis further proposes an image representation method based on anatomical landmarks to further remove redundant information and enhance the generalization of the algorithm.The locations with local minimum in the results of two sample t-test are identified as discriminative landmarks,and landmarks are screened for many times,so as to achieve better classification effect with as few biomarkers as possible.Considering the complex correlation between different regions of the brain,all the image patches of each sample are concatenated together to form a large patch as the image representation.Then a CNN network is applied to deeply fuse the features of patches.Meanwhile,the thesis build a DAG network with a layer of skip connection,which combines high-level strong semantic features and low-level high-resolution features.Finally,SVM is used for classification.The experimental results show that the identified landmarks can be used as important biomarkers for the diagnosis of AD and the classification performance is improved by the proposed CNN network.In addition,compared with the region of interest(ROI)-based method,voxel-based method and patch-based deep features concatenation method,the proposed method has better classification performance.