Research On Brain Image Analysis Based On Sparse Structural Feature Learning And Their Applications

Brain diseases not only endanger the health and life of patients,but also bring heavy economic burden and great mental stress to patients,their families and the society.The early diagnosis and early treatment is the only way to eliminate the harm of brain diseases and neuroimaging is an important tool for early diagnosis and treatment of brain diseases.Medical image analysis and processing technology based on computer science has been widely used in brain image analysis.Among them,sparse learning technology has been paied much attention.Because it can discover the inherent structure of data.One of the challenges and difficult problems in brain image analysis is how to learn and utilize effective features,especially the structural features of brain information such as multimodal structures and brain network structures.Brain image analysis generally includes data acquisition,brain segmentation,feature extraction,feature learning and classification steps.Based on sparse learning,this dissertation focuses on the brain segmentation and feature learning in brain image analysis and studies the structural feature representation.The specific research work and innovation are as follows:(1)In order to make full use of the priori knowledge of atlas,this paper proposes a Hierarchical Sparse Representation(HSR)method for multi-atlas segmentation.In this method,the tree structure is used to construct the dictionary(main dictionary and residual dictionary)for different atlas images,so that the constructed dictionary has compact and efficient expression ability.Furthermore,in order to eliminate the noise and reduce the impact of the background image patches on the final segmentation results,the main dictionary is combined with the residual dictionary for the sparse reconstruction of the target image patch,to achieve more detailed organization and characterization.Finally,after the optimal sparse reconstruction coefficient is obtained,the residuals of the target patch are used to determine the label to which the target image patch belongs.In the ADNI and PPMI datasets,segmentation experiments were performed on the brain regions associated with Alzheimer s disease,brain stem and basal ganglia related to Parkinson's disease,respectively,which validates the effectiveness of the proposed segmentation method.At the same time,it also proves that enhancing the ability of dictionaries plays an important role in sparse reconstruction.(2)In order to maximize the retention of the classification structure and label information of different modality data,and to fully utilize the discriminating ability of data features,a Label-aligned Multi-modality Feature Selection(LMFS)is proposed.The method uses group sparse constraint to ensure discriminative features can be jointly selected from multi-modality data.Secondly,in order to embed the category structure information of data,label alignment regularization term is introduced into the objective function of the traditional multi-modality feature selection method.Based on the proposed algorithm,an Accelerated Proximal Gradient(APG)method is used to optimize the problem.Finally,a multi-kernel support vector machine is used to fuse the selected multimodal data features for final disease diagnosis.The proposed method can find disease-sensitive biomarkers on the ADNI dataset and achieve better classification performance in multiple groups of diseases.(3)In order to overcome the problem of inaccuracy of the similarity measurement between high-dimensional features,an adaptive neighborhood learning strategy is introduced.And a multi-modality feature selection method based on adaptive neighborhood learning is proposed(ASMFS).The traditional method of feature selection which introduces data structure often needs to know the structure relationship of a given paired-sample before learning the feature,such as the similarity between two pairs of samples.If the given similarity measure is not accurate,it will affect the performance of the subsequent task.To solve this problem,the proposed method is to model the similarity of samples and the sparse weight coefficients simultaneously in multi-modality feature selection.This in turn ensures that the correct modal structure is embedded and induces features with more discriminability.In addition,in order to better characterize the inherent correlation between modalities,it is assumed that all modality data share the same sample similarity matrix.Finally,the proposed iterative optimization algorithm is used to solve the proposed problem,and the selected features are classified by using the multi-kernel support vector machine.A comparative experiment with the state-of-the-art methods on the ADNI dataset verifies the effectiveness of the proposed method.(4)In order to depict the correlation of high-order structures among multiple brain regions,a Transductive Hypergraph Learning based Subnetwork Selection(THLSS)method is proposed.At present,many studies based on brain networks focus on the correlation between pairs of regions of interest in the brain,ie,second-order relationships,while ignoring the association features of more brain regions(eg,three brain regions).More brain regions can form subnetworks.Such subnetworks with high-order relationships may provide guidance in the early diagnosis of the disease.In response to this problem,by introducing the hypergraph laplacian regularization term and the group sparse constraint in transductive learning,the discriminative sub-network is excavated to help find and discover valuable biomarkers.Experiments were performed on ABIDE and ADHD200 brain network datasets using Support Vector Machine and Support Tensor Machine.The results show that using the proposed method enables the discovery of potential subnetwork links related to the disease,which can be utilized in classification to improve diagnostic capabilities.