Investigation On Macaque Nucleus Segmentation Based On Multi-atlas

The macaque brain is similar to the human brain in structure,and some advanced functions such as the visual system are also similar.Therefore,human can study some mechanisms of the human brain by studying the macaque brain.In addition,cross-species comparison by comparing the differences in the structure of the macaque brain and human brain can also reveal more evolutionary mysteries.The main method of studying the macaque brain is to analyze the collected MRI images.Among them,sub-cortical nucleus segmentation is the most difficult and the most important.However,at present,the segmentation of sub-cortical nucleus is mainly carried out by human brain MRI analysis software,and then manually modified and improved them.There are few automatic segmentation tools for macaque,so it is significant to study the automatic segmentation method of macaque sub-cortical nucleus.In order to improve the above situation,this paper studies the brain MRI image segmentation algorithm,selects the sparse representation multi-atlas segmentation algorithm for improvement,and applies it to the segmentation of the sub-cortical of macaque brain.The main research contents are as follows:(1)For the weighted fusion of sparse representation multi-atlas segmentation algorithm without considering the global similarity,an improved sparse representation multi-atlas label fusion algorithm is proposed.The algorithm introduces mutual information and improves the calculation method of information entropy in mutual information,considering the similarity between the target image block and the label image block from the global and local.The calculation method of information entropy in mutual information is improved by adding the probability ratio of gray matter to gray-white matter as a parameter to the calculation process of information entropy.At the same time,the information entropy of the label image is calculated,and the Dice coefficient is added to the label image as a parameter for mutual information calculation.The two parts of information entropy are added to obtain an improved information entropy calculation formula.The improved mutual information is used to measure the overall similarity between the target image and the atlas image.These two measures make the weight of each atlas more reasonable,and an improved sparse representation multi-atlas segmentation algorithm is proposed.(2)To improve the shortcoming that the sparse representation label fusion algorithm may lose the image block information,a new label fusion algorithm is proposed.The algorithm uses the distance measurement index D_Ri to fuse the segmentation result of the non-local block label fusion algorithm and the sparse representation fusion algorithm segmentation according to a certain rule,so as to achieve a better segmentation result.The index is the sum of the reciprocal of the L-Dice coefficient and the cosine distance.Among them,the L-Dice coefficient is an improvement of the Dice coefficient.Because the Dice coefficient cannot quantitatively reflect the number gap between the two sets,the algorithm adds the absolute value of the difference between the number of the target image label voxels value and the number of label voxels value to the Dice coefficient.In order to ensure that the Dice coefficient is not zero,a constant d is introduced,and finally the improved L-Dice coefficient is obtained.(3)Verify the two algorithms proposed in this paper.In this paper,the macaque data set released by Oxford University is selected as the test data set,and the D99-SL atlas,NMT atlas and Huanghao atlas are used as the participating atlas.The majority voting method(MV),local weight voting(WV)method and nonlocal patch-based weighting(PATCH)method and the proposed method are used to segment the test data set of sub-cortical nucleus,and then select the hippocampus,striatum and claustrum to evaluate the accuracy of segmentation.The results prove that the method proposed in this paper have improved the accuracy of segmentation of three nucleus compared with other algorithms,especially the results of hippocampus and striatum is better.In addition,two methods have better robustness.