Multivariate Data Fusion Method And Application Research For High-dimensional Medical Image Processing

Alzheimer’s disease(AD)is a progressive neurodegenerative disease.Computer-aided imaging techniques,such as magnetic resonance imaging(MRI)and positron emission tomography(PET),are widely used to help identify AD.This thesis starts with structural magnetic resonance imaging(SMRI),studies the feature extraction of high-dimensional medical data,and multivariate data fusion methods to improve the accuracy of AD classification and diagnosis.Specifically,it involves how to extract effective classification features while maintaining the three-dimensional structure information of the data under the condition of small samples;how to use the information of multiple brain regions to jointly learn and combine features to improve feature classification performance.Considering the high-dimensional data expression ability of tensor representation,this thesis analyzes the tensor-based feature dictionary learning method,the method of determining the effective rank in tensor dictionary learning,and the feature fusion method based on multiple brain regions.The main research content is as follows:(1)High Dimensional Feature Extraction Based on Tensor Dictionary LearningAiming at the problem of loss of spatial information when applying the traditional matrix dictionary learning method to high-dimensional data,this thesis uses tensor to represent high-dimensional data,adopts tensor Tucker decomposition,extracts the common decomposition factors of training samples as the basis,and uses This set of basis inverse operations yields the corresponding kernel tensors,i.e.features.Using the dimensionality-reduced features,this thesis introduces an attention mechanism into the3 D convolutional neural network to further improve the classification performance.(2)Optimal Rank Optimization AlgorithmIn the above tensor dictionary learning,the rank of tensor decomposition is a priori condition,and determining an appropriate rank can preserve the useful information of the data to the greatest extent.Therefore,this thesis conducts further research on the selection of tensor low-rank approximation ranks for three-dimensional brain data,and proposes a model for automatically finding the optimal rank of tensor decomposition,which can avoid repeated trials of manual rank selection.In addition,the data is processed in blocks,and different brain regions use different ranks to achieve better feature extraction performance while meeting the data compression rate.(3)Multi-brain region feature fusion methodSince different brain regions have different performances in different patients,in order to fuse and analyze the influence of the characteristics of each brain region on the lesion,this thesis adopts an early feature fusion strategy to improve the ability of features to distinguish diseases.Specifically,strategies such as maximum pooling,Transformer modular fusion,tensor splicing,and hierarchical decomposition are proposed.Based on the above feature extraction,tensor decomposition rank optimization,and multi-brain region feature fusion methods,a series of experiments were done to verify the performance of the algorithm in this thesis.Experiments show that compared with the existing AD classification methods,the method proposed in this thesis has a greater improvement in classification performance.And made a visual interactive interface to integrate the algorithm.