Research Of Brain Structure Segmentation Algorithms Of Infant Based On Multi-atlas Fusion And Dictionary Learning

Infantile period is a critical period of brain development.The risk of brain disease is much higher than that of other periods.Therefore,it is of great importance to explore infant brain diseases.Segmentation of infant brain structure by using nuclear magnetic resonance images is the basis for morphometric analysis of infant brain structure,studying infant brain development and diagnosing diseases.The structure of the deep brain consists of the corpus callosum,the hippocampus,and the ventricles of the brain.The morphological changes are closely related to many diseases.The deep brain structure of children is in development,and the information of gray and shape is more complex and irregular than that of adults,and its segmentation is more difficult than that of adults.In this issue,we take the hippocampus and corpus callosum as an example.Under the framework of multi-atlas,the three-dimensional automatic segmentation method of the brain structure of infants is studied in detail based on dictionary learning.The main work and research results of this issue are as follows:(1)In view of the characteristics of fuzzy edge structure in infant brain MR images.A multi graph fusion algorithm based on discriminative dictionary learning is studied and implemented,and an improved algorithm of discriminative dictionary learning combined with structural information is proposed.The improved algorithm discriminative dictionary learning based on the structural information to the target classification of voxels around the most select some adjacent elements.Each of these voxels and target voxel dictionary learning and sparse representation,multiple tags are weighted analysis comprehensive discrimination to the target label.This method takes the local information of brain images more comprehensively,avoids the drawback of single discrimination of target voxels,increases the dependency between labels,and improves the discriminating ability of brain structure edge tags.(2)In view of the poor quality of infant brain MR images,based on the multi dictionary fusion algorithm based on kernel dictionary learning,a brain structure segmentation algorithm based on fusion of image texture features and kernel dictionary learning is proposed.In the process of multi graph fusion dictionary learning,the training data and test data are nucleated to form high dimensional virtual samples.The method maps data to high-dimensional space,and then integrates LBP texture features of image information,improving the discriminant ability of image blocks.The algorithm has better segmentation results for low quality images and improves the segmentation precision of the algorithm.(3)In view of the low contrast of infant brain MR images,a multi graph fusion algorithm based on label consistency dictionary learning is studied,and the kernel label conformance dictionary learning algorithm is first applied to multi-atlas fusion to achieve infant brain structure segmentation.The label conformance dictionary learning uses the markup relation between the original dictionary and the training dictionary to construct the mark error term,which enhances the dictionaries' discriminability.The nuclear label consistency dictionary learning based on the fingerprint image from the original feature space to a high dimensional space.The dictionary learning in high dimensional space,solves the problem of nonlinear data,further improve the discriminative dictionary,overcomes the disadvantages of low contrast image.