Research And Application Of Brain Image Analysis Algorithm Based On Multi-modal Data

In recent years,the development of brain imaging technology has provided enormous help in understanding specific brain regions and their functions.A multi-modal brain image analysis method can help doctors analyze the pathological mechanism of brain diseases by using a variety of information.At present,multi-modal brain image analysis is still facing certain difficulties,such as the small sample size of multi-modal brain image data,the high feature dimensionality of brain image,heterogeneous data and other problems.Finding the biomarkers associated with disease and improving the accuracy of diagnosis are the current focuses in the multi-modal brain imaging data analysis.Existing brain image analysis methods usually only used single modality of brain image data,or combined multiple kinds of brain image data together through simple concatenating,which is not fully utilized between modal information.Fully considering the structural information and correlation relations in brain image data,this thesis proposes two kinds of multi-modal brain image analysis methods,the main work and innovation are as follows:In order to make full use of the correlation information in multi-modal brain image data,this thesis presents an association analysis model between genes and multi-modal brain image with the feature sparsity constraints.Considering the high dimensionality of multimodal brain image data and the presence of noise,this thesis adds sparsity constraints to the objective function to discover the biomarkers of brain regions associated with risk genes.Furthermore,by taking advantage of the characteristics of different degree of a correlation between each modal and gene,and the existence of a correlation between modalities,the modal correlation constraint term is added,so that the model can obtain the correlation coefficient between different modalities and gene.The experimental results confirm that the proposed method can identify the brain regions highly associated with pathogenic genes from the functional and structural modalities.In order to make the most of the information provided by brain image data and effectively fuse different data in feature selection,this thesis also presents a novel feature selection method with a multi-modal feature selection method based on consistent measurement constraints.Considering that there is a prior distribution between samples and the distance relation between samples in the positron emission tomography and magnetic resonance imaging modals is different,they provide brain function and structure information respectively.In this thesis,the random forest is used to provide a consistent pair similarity measurement method for multi-modal samples,which can describe the distance of samples under consistent metrics.In order to combine data of multi-modal at the same time,we induce group sparsity constraints into the model.So that the model not only preserves the similarity information between samples on each modality,but also can jointly select the shared features from different modalities.The proposed method can classify the samples and identify the brain regions associated with the disease.The proposed method is able to classify samples and identify brain regions associated with the disease.The proposed algorithm is compared with the state-of-the-art multi-modal brain image analysis methods on the ADNI database.The experimental results show that the proposed multi-modal brain image analysis model can improve the experimental accuracy and find more discriminative brain region features by adding sparse constraints.