An Early Diagnostic Method For Alzheimer's Disease Based On Superv I Sed Locally Linear Embedding Algorithm

Early diagnosis of Alzheimer's disease(AD)is one of the unsolved problems.High-dimensional data collected by magnetic resonance imaging(MRI)contain a large amount of redundant feature inevitably and noisy.As a result,it is difficult to distinguish two early stages of sMCI and aMCI between cognitive normal(CN)and AD patients.Therefore,in order to improve the classification accuracy,many dimensionality reduction methods of machine learning have been proposed by researchers,which are used to reduce redundant information and extract more discriminative features.However,the nonlinear of the early stages of AD often be neglected by the traditional linear dimensionality reduction methods,which lead to the intrinsic relationship between data is difficult to find in high-dimensional space.So the early classification of AD's improvement is not significant with these methods.In this thesis,the supervised locally linear embedding algorithm(SLLE)in manifold learning is applied to the early diagnosis of AD,and the proposed improvement algorithm is also used to early diagnosis of AD based on ADNI dataset.These algorithms not only extract effectively the nonlinear structural features from the data;but also make full use of the label information to improve the early classification effect of AD.The main works of this thesis can be summarized as follows:1.Because the traditional linear feature extraction method neglects the nonlinear distribution of data in the AD early stage,and the locally linear embedding(LLE)method in extraction of nonlinear features dose not make full use of the label,this thesis proposes a SLLE method for early diagnosis of AD.SLLE not only mine the potential topological structure of data,but also make full use of the sample label information to optimize the loss function.And it achieves the discrete dimensionality reduction effect between classes to improve the accuracy of AD classification.2.Aiming at declining the limitations of LLE methods for the data of dense homogeneous distribution and the influence of neighboring points,we propose an improved distance locally linear embedding algorithm(MLLE).The nearest neighbors are obtained by calculating the geodesic distance between sample points.The sample distance is adjusted by using the reciprocal of the average distance between the sample's neighbors,so that the whole distribution tends to be uniform.Compared with the LLE algorithm,the MLLE algorithm reduces the influence of the number of nearest neighbors.Combined with the idea of supervised learning,an improved supervised locally linear embedding method(SMLLE)is proposed in this thesis for the early diagnosis of AD.3.In order to solve the problem of the limited improvement of recognition rate in early diagnosis of AD based on single modal data of MRI,this thesis proposes to improve the recognition rate of dichotomous classification of AD by combining the features of volume of interest(VOI)and cerebral cortical thickness(CTH)based on MRI with the biomarker characteristics of cerebrospinal fluid(CSF).At the same time,it is a multi-classification problem to diagnose the different stages of disease in clinical practice.Therefore,this thesis combines LLE,SLLE,MLLE and SMLLE with multi-classifier support vector machine(SVM)to verify the effectiveness of the proposed method.