Multimodal And Longitudinal Classification Research Of Support Vector Machine In The Progression Of Alzheimer's Disease

Alzheimer's Disease(AD)is a latent neurodegenerative disease that mostly occurs in the elderly,while mild cognitive impairment(MCI)is a transitional stage between healthy elderly and AD.The healthy elderly is usually defined as normal control(NC).Magnetic resonance imaging(MRI)technique can display the structure and function of brain tissue noninvasively.It can be used to effectively classifly disease by combining with multivariate pattern analysis method of machine learning.This thesis use the support vector machine(SVM)algorithm to analyze the difference of brain structural MRI(sMRI)and classification performance among the NC,MCI and AD groups,aiming to explore a diagnostic pattern of AD based on big data and artificial intelligence.Meanwhile we classifly the functional MRI data of normal ageing population for different age groups,hoping to provide new clues for the functional research of AD from a methodological perspective.The main work is as follows:Chapter 1:First of all,we introduce research background,research significance and current research situation at home and abroad for AD.Then,the basic principle and appli-cation of multimodal(structural/diffusion/functional)MRI and the imaging database-the Alzheimer's Disease Neuroimaging Initiative(ADNI)and Cambridge Centre for Age-ing and Neuroscience(Cam-CAN)used in this thesis are introduced.The ADNI database includes the NC,the patients with sub-type MCI and AD,aiming to analyze the patho-logical mechanism in the process of evolution of AD from different perspectives.By contrast,The Cam-CAN database includes the normal ageing population from 18 to 88 years old,hoping to provide new clues for different and overlapping features of brain between normal ageing and AD.Finally,the main research contents of this thesis are summarized.Chapter 2:The theoretical foundation of support vector machine(SVM)is intro-duced.Firstly,the steps of machine learning in neuroimaging are described.Secondly,the basic principle and kernel function method of SVM are described.Finally,the application research of SVM and the solving strategies of high dimensional feature are introduced.Chapter 3:The cortical features extracted from sMRI data in AD patients and NCs are used for statistical analysis and classification research.Firstly,the sMRI data of all subjects are preprocessed to obtain five parameters including the cortical thickness,surface area,gray matter volume,curvature and sulcus depth.Significant differences can be found in the bilateral entorhinal cortex and left medial orbitofrontal cortex according to the results of statistical analysis between-group.Then the recursive feature elimination(RFE)method is used for feature selection and the SVM method combined with leave-one-out cross validation(LOOCV)is used to classifly the NCs and AD patients base on single parameter and multi-parameter combination.The classification performance before and after feature selection is compared.In addition,the receiver operating characteristic(ROC)curves and the areas under the curves(AUC)are drawn to verifly the stability of the classifier models.The results show that high accuracy(90.76%)and good robustness(AUC=0.94)can be obtained by multi-parameter combination.Finally,the SVM plane classifiers are constructed by using the first two features of weight value in feature selection of different parameters,and found that the classification performances are related to the effect of feature selection.Chapter 4:The SVM and logistic regression(LR)algorithms are used to classifly the diffusion MRI(dMRI)data for the patients with AD,early MCI,late MCI and NC-s.First of all,the dMRI data of all subjects are preprocessed to obtain six parameters including the fractional anisotropy(FA),mean diffusivity(MD),axial diffusivity(DA)and radial diffusivity(RD),local diffusion homogeneity used spearman's rank correla-tion coefficient(LDHs)and Kendall's coefficient concordance(LDHk).Then,the SVM and LR algorithms are used to classifly different group based on the parameters after feature selection and the permutation tests are used to determine whether the classifi-cation accuracy of multi-parameter is significantly higher than that of the random case.In addition,the performance of the classifier model is verified by the ROC curve.The results demonstrate that the classification performance of SVM is better than that of LR,and the classification effect of multi-parameter combination is better than that of single parameter classification.Finally,by listing the discriminative white matter(WM)features,it can be found that the uncinated fasciculus,cingulum,corpus callosum,coro na radiate,external capsule and internal capsule are the features that play a significant contribution in the classification of the four groups.These results suggest that multi-type and multi-region of WM features can effectively improve the diagnosis accuracy of AD and MCI.Chapter 5:The resting-state functional MRI(rs-fMRI)data with different age groups are classified.First,the data sets of 18?88 years old from Cambridge Centre for Ageing and Neuroscience(Cam-CAN)database are divided into young group(18?39 years old),middle-aged group(40?59 years old)and elderly group(60?88 years old),and the pre-processing of rs-fMRI data of all subjects(including global signal regression(GSR)and no global signal regression(NGSR))is conducted.We compute the functional connec-tivity strength of the whole brain(WHFCS),the left hemispheric functional connectivity strength(LHFCS)and the right hemispheric functional connectivity strength(RHFCS)as classification parameters based on an atlas of intrinsic connectivity of homotopic areas(AICHA).Then we use RFE method for feature selection and use SVM for classification.The results show that the classification performance improve through feature selection.In addition,the ROC curves and AUCs are drawn to validate the robustness of classifi-er.Next,we relate the discriminative features to the physiological mechanism of ageing.Finally,we extract the features based on the human brainnetome atlas(BNA)and verify the applicability of different atlas for classification.Chapter 6:The longitudinal sMRI data of AD,stable MCI,converted MCI and NC are used for statistical analysis and classification research.Firstly,the whole brain gray matter volume(GMV)of all subjects is extracted as a classification parameter during preprocessing.The differences of GMV patterns between-group at different time point and the differences of longitudinal changes of GMV within the same group are compared respectively.It is found that gray matter is a gradual change process when developing from NC to MCI and AD,which also proves that MCI is the transitional stage between NC and AD.Meanwhile,it indicates that the frontal and limbic lobes are the most severely damaged brain regions.Next,the longitudinal feature combination strategy is used to construct the classification model via using SVM combined with nested LOOCV method.The results demonstrate that the longitudinal feature combination strategy contributes to improve the classification performance.Finally,we associate the features which contribute significantly to classification with the pathological manifestations of AD.