| Alzheimer’s disease(AD)is an irreversible,progressive neurodegenerative disease,which has brought serious threats to the health and safety of the elderly.Mild cognitive impairment(MCI),commonly accompanied by a measurable memory impairment but largely intact cognitive function,is referred to as the prodromal stage of AD.Identifying MCI patients who will convert into AD from elderly people and providing well-targeted treatment can help prevent or delay the disease.Structural magnetic resonance imaging(s MRI)has been widely used in the recognition of MCI due to its non-invasive nature,moderate costs,and high spatial resolution.Compared with AD,the alterations of brain structure and function caused by MCI are not significant.How to extract discriminative features from high-dimensional s MRI data is a key problem to be solved in the early recognition of MCI.In this dissertation,the approaches in fuzzy mathematics and machine learning are used to investigate the feature extraction methods based on s MRI.And then,the proposed features are verified in MCI diagnosis,prediction of MCI-to-AD conversion and the disease’s mechanism research.Results indicate that these works provide reliable biomarkers for the investigation of computer-aided diagnosis and pathological mechanism of MCI.The main contributions are summarized as follows:1.A novel method is developed to extract structural connectivity of individual brain based on the indirect relation of brain regions.Previous studies extract structural connectivity only using the information of two brain regions,ignoring the relationship between brain regions and their common neighbors.To solve this problem,this work constructs individual structural network based on the indirect relation of brain regions.Specifically,two fuzzy sets are composed of the relations between two brain regions and their common neighbors,and the connectivity are extracted by describing the correlation of the two fuzzy sets.Experimental results show that the proposed structural connectivity can more accurately reflect the abnormal alterations of brain structure,and effectively improve the efficiency of MCI diagnosis and conversion prediction.In addition,the Mini-Mental State Examination(MMSE)scores and direct relation features are included to further improve the classification performance.The final accuracies of MCI diagnosis and conversion prediction within 3 years are 90.16% and 77.38%,respectively.In conclusion,this work proposes an effective idea for the extraction of individual structural connectivity,and provides reliable biomarkers for the MCI diagnosis and conversion prediction.2.A novel method is proposed to extract self-weighted grading features based on graph regularization and information propagation.In the traditional methods of predicting MCI-to-AD conversion with auxiliary information of AD and normal controls(NC),the local structure information of brain has not been explored.To overcome this defect,this work preserves the neighborhood relationship between brain regions of MCI subjects by introducing a graph Laplacian regularization term in linear regression model,which guarantees the accuracy of information propagation from AD and NC subjects to MCI subjects.And then the biomarkers are extracted for MCI subjects by a self-weighted method.In addition,the integration of biomarkers based on two morphological features in the classification framework provides complementary information,which enhances the classification performance.Experimental results show that the self-weighted grading features propagate the information from AD and NC subjects to MCI subjects more precisely.The features achieve high prediction accuracy of MCI-to-AD conversion,and further reveal the transformation mechanism of MCI.3.A feature extraction method is proposed based on correlation analysis and local structure preservation.How to obtain low-dimensional discriminative features from high-dimensional s MRI data through spatial transformation is a key problem to be solved in the prediction of MCI conversion.This work first learns two projection matrices using subspace learning based on AD and NC data.In this process,the one-to-one correspondence between sample and its label,correlation between two morphological features of the same image and the nearest neighborhood of samples are kept as much as possible.And then,the data of MCI subjects are mapped into low dimensional subspace by the two projection matrices.Finally,the biomarkers are extracted for MCI subjects by the information integration of two morphological features.In the experiments of predicting the conversion from MCI to AD,the proposed features achieve good performance with different classifiers.This proves that the features have strong discrimination ability for MCI subtypes and universal adaptability for different classifiers.4.A novel method is developed to extract structural connectivity of individual brain based on the fusion of direct and indirect relation of brain regions.Previous studies quantify structural connectivity only using direct or indirect information of brain regions.However,in the investigation of structural covariation pattern among the cortical regions for MCI patients,it is important to fuse the information of direct and indirect relation to achieve more comprehensive and scientific representation of the relationship between brain regions.This work constructs a fused network,which fuses the direct and indirect correlation for brain regions with common neighbors,and utilizes the direct correlation for brain regions without common neighbors.By conducting an effective way of information fusion,the structural connectivity can be quantified more scientifically and accurately.Experimental results show that the accuracy of MCI conversion prediction within 3 years is 77.83%.The inclusion of MMSE scores further enhances the prediction performance,and the final accuracy is 79.64%.This work provides an effective method to predict the development of brain diseases from the perspective of constructing network connectivity. |
PDF Full Text Request